From Official Election Data to AI-Generated Answers: Responsibility Across the Voter-Information Chain

Who should put things right when an AI assistant gives voters the wrong answer? A companion analysis on Malaysian election integrity.

By Muhammad Syawalfiza.

Part of a four-essay series: This article is the second companion analysis to "Protecting Voter Autonomy in the Age of Artificial Intelligence (AI)" , following "The Invisible Architect: How AI Is Quietly Shaping What We See".

The original essay examines how AI could influence political judgement and voter autonomy. The first companion explores how information selection and presentation can quietly shape the choices people see. This second companion addresses a practical question: when an AI-generated election answer is wrong, who is responsible for correcting it?

1. Follow the information, not just the answer

Imagine a voter asking an AI assistant where to vote, what identification is required or whether a registration deadline has changed. The response is quick, clear and confident. It may even include a source link.

But confidence is not proof of accuracy.

A Malay proverb captures this neatly: "Sepandai-pandai tupai melompat, sekali gawal juga". Even the most skilful can make a mistake. An AI system should therefore be judged by the evidence behind its answer, not by how confidently or fluently it speaks.

The information might have originated with the Election Commission of Malaysia (SPR), passed through a data intermediary, been retrieved and summarised by an AI system, and finally appeared in a search interface or messaging application.

Think of it as a relay race:

SPR's official information → data intermediary → AI-generated answer → digital platform → voter

If something goes wrong, the mistake may have occurred at any point along the route.

The appropriate response is not to blame SPR automatically because the information began with an official record, or to assume that the AI provider is responsible for everything that appears on a platform. Responsibility should follow each actor's function, control over the process and ability to prevent or correct the error.

SPR should answer for official information within its control. Intermediaries should preserve its source and currency. AI providers should address failures in retrieval, synthesis and presentation. Platforms should address the distribution and reporting functions they control.

This is distributed but differentiated responsibility: each organisation should be accountable for the part of the process it operates.

2. The problem is already documented internationally

Research has established that AI assistants can produce inaccurate, incomplete or outdated election information. In a study of 300 responses to 100 questions about the 2024 UK general election, Simon, Fletcher and Nielsen (2024) found that ChatGPT-4o and Perplexity.ai provided correct answers in 78% and 83% of cases, respectively.

These figures apply to the particular systems, questions and testing period. They should not be treated as universal error rates or evidence that AI-generated election information has changed voting behaviour.

The wider governance challenge is recognised in International IDEA's "Artificial Intelligence for Electoral Management" (2024) and the United Nations Development Programme's "Global Landscape Analysis: Artificial Intelligence in Electoral Administration", published on 19 October 2026. The latter draws on a survey of 53 electoral management bodies across six regions (United Nations Development Programme [UNDP], 2026).

These precedents matter. The research opportunity is not to claim that nobody has considered AI accuracy or accountability in elections. It is to determine whether responsibilities are adequately allocated when official information passes through several independent systems—and whether a voter can obtain an effective correction when an answer fails.

A practical example already exists in the United States. Democracy Works' Elections API supplies data drawn from official government sources, offers links through which users can examine sources and report problems, and supports integration with generative-AI services.

On 20 October 2026, OpenAI reported that its Democracy Works integration had gone live in ChatGPT in the United States. It provides information about registration, deadlines, voting locations and other requirements, with links to official election sources (OpenAI, 2026).

These initiatives demonstrate a feasible approach to source-verified voter information. They are not, by themselves, proof that every generated answer is correct or evidence of a universal legal duty applying to all AI providers. The problem is already documented internationally

Research has established that AI assistants can produce inaccurate, incomplete or outdated election information. In a study of 300 responses to 100 questions about the 2024 UK general election, Simon, Fletcher and Nielsen (2024) found that ChatGPT-4o and Perplexity.ai provided correct answers in 78% and 83% of cases, respectively.

These figures apply to the particular systems, questions and testing period. They should not be treated as universal error rates or evidence that AI-generated election information has changed voting behaviour.

The wider governance challenge is recognised in International IDEA's "Artificial Intelligence for Electoral Management" (2024) and the United Nations Development Programme's "Global Landscape Analysis: Artificial Intelligence in Electoral Administration", published on 19 October 2026. The latter draws on a survey of 53 electoral management bodies across six regions (United Nations Development Programme [UNDP], 2026).

These precedents matter. The research opportunity is not to claim that nobody has considered AI accuracy or accountability in elections. It is to determine whether responsibilities are adequately allocated when official information passes through several independent systems—and whether a voter can obtain an effective correction when an answer fails.

A practical example already exists in the United States. Democracy Works' Elections API supplies data drawn from official government sources, offers links through which users can examine sources and report problems, and supports integration with generative-AI services.

These initiatives demonstrate a feasible approach to source-verified voter information. They are not, by themselves, proof that every generated answer is correct or evidence of a universal legal duty applying to all AI providers.

3. Does Malaysia already have relevant laws?

Yes. Malaysia has electoral legislation, communications regulation, an online-safety framework, data-governance rules and national AI governance initiatives. The issue is whether these arrangements work together effectively for third-party AI answers.

MCMC and the Communications and Multimedia Act

On 15 November 2025, the Malaysian Communications and Multimedia Commission (MCMC) reminded social-media users to act responsibly during Sabah's state election. Its statement referred to section 233 of the Communications and Multimedia Act 1998 (Act 588) and the enforcement of applicable restrictions on online communications (Malay Mail, 2025).

This illustrates the regulator's public approach to online conduct during an election. However, a news report summarising a regulator's warning is not a substitute for the statutory text. Liability depends on the provision in force, its elements and the circumstances of the case.

An AI system giving an outdated polling location is not automatically equivalent to a person committing a criminal offence. Nor does section 233, by itself, establish a universal duty requiring every AI provider to cite official sources, update election information or correct every factual error within a prescribed time.

The Communications and Multimedia Act 1998 (Act 588) must be distinguished from the Malaysian Communications and Multimedia Commission Act 1998 (Act 589), which establishes MCMC and defines its institutional functions. Both Acts were amended in 2026, with the relevant amendment Acts (A1798 and A1799) coming into force on 14 October 2026. Their implications must be assessed provision by provision, taking account of the statutory wording, applicable service definitions and regulatory powers. The commencement of these amendments does not, by itself, establish a blanket duty to correct every AI-generated election answer.

The Online Safety Act

The Online Safety Act 2025 (Act 866) came into force on 1 January 2026. It establishes duties for specified service providers within its scope, including requirements relating to harmful-content risks, user assistance and reporting mechanisms.

This is an important regulatory development. Nevertheless, its harmful-content framework should not be treated as a general guarantee that every factual statement generated by an AI system will be accurate. An error about a polling location may require correction even where it does not fall within a particular statutory category of harmful content.

SPR and existing complaint channels

SPR already publishes official electoral information and maintains public complaint and feedback mechanisms. Its customer charter sets out targets for handling public complaints and forwarding relevant election-period feedback to the responsible parties.

This is useful existing infrastructure, not something the research should overlook. The unresolved question is whether these channels can handle errors originating outside SPR's own systems.

Suppose SPR publishes the correct polling information, an intermediary retains an old version and an AI assistant subsequently delivers it to a voter. SPR may be able to confirm the error, but who corrects the intermediary's data? Who fixes the AI system's retrieval process? How does the voter know that the correction has reached the service used?

Those questions require investigation of actual procedures. The existence of a complaint channel does not, by itself, establish that the whole correction chain is connected.

Malaysia's proposed AI Governance Bill

Malaysia's National AI Office has been developing a proposed AI Governance Bill. Its 2026 public consultation has closed, but the proposal should not be treated as enacted law.

The emerging framework offers an opportunity to consider sector-specific safeguards, including the treatment of errors passed between developers, data intermediaries and deployers. The research must establish what powers and obligations the eventual legislation actually creates rather than assume that a horizontal AI law will automatically resolve election-information errors.

The legal conclusion at this stage is measured: Malaysia has relevant laws and institutions, but the material reviewed does not establish a complete, universal process for correcting ordinary factual errors in every third-party AI voter-information service.

4. What should a workable accountability system look like?

For election information, that means checking source links, update arrangements and correction procedures before polling begins, rather than waiting until a voter receives the wrong answer.

A proportionate system should connect five responsibilities.

First, authoritative information. SPR should make important voter information identifiable, dated and accessible. Where practical, official records should have version identifiers and formats that digital services can reliably process.

Second, intermediary integrity. An aggregator should preserve source links, record update times, distinguish official information from unverified material and propagate corrections when its upstream source changes.

Third, AI-provider responsibility. Providers should test important voting-related questions, use authoritative sources where available, show sources that genuinely support material claims and acknowledge uncertainty when information cannot be verified. They should provide an accessible route for reporting errors and address defects within their control.

Fourth, voter redress. Voters should not have to diagnose the technical failure before reporting it. A coordinated process should receive the complaint, identify the likely point of failure, refer it to the responsible organisation and communicate the outcome.

Fifth, oversight. Regulators should act within their legal powers, require proportionate remediation where authorised and provide transparent reasons and opportunities for review.

The key distinction is between finding a mistake and fixing it. A warning that an AI answer may be wrong does not restore an outdated polling location. The underlying record, retrieval process or generated answer must be corrected where necessary.

5. Five Additional Safeguards for AI-Mediated Voter Information

As the Malay saying Sediakan payung sebelum hujan reminds us, preparation should come before trouble arrives (DBP, n.d.-b).

A workable system needs clear official sources, traceable intermediaries, tested AI answers, accessible reporting and proportionate oversight. Five additional safeguards could help Malaysia develop such a system.

5.1. Establish a local compliance route

Major AI providers may operate from outside Malaysia. Contacting an overseas provider during an election does not necessarily ensure a timely response.

Malaysia could examine whether designated high-reach or high-impact AI services should appoint a local legal representative or registered compliance contact, where legislation provides a valid basis. The EU's Digital Services Act and AI Act provide precedents for local representatives for specified non-EU providers.

A named contact could receive lawful notices, preserve relevant records and coordinate responses. But local representation would not automatically expand MCMC's or SPR's powers, guarantee overseas compliance or justify blanket data-localisation requirements. ASEAN cooperation and shared incident-reporting procedures could provide additional support (ASEAN Secretariat, 2025).

5.2. Make correction speed proportional to risk

A wrong polling location is more urgent than an error in general electoral history. Correction targets should reflect the severity and timing of the potential harm.

Malaysia could distinguish routine incidents, election-period errors and critical failures close to polling. A proposed 12–24-hour correction window could be tested, but should not be presented as an established international standard; it may be too slow for an error affecting voters on polling day.

For critical procedural questions, systems should use a verified fallback. If the current answer cannot be confirmed, the AI should stop improvising and direct the user to the appropriate official SPR service.

Australia's government AI technical standard supports risk mitigation through interventions such as safe-state recovery, manual override and system logging.

The practical rule is simple: when an AI system cannot verify where or when someone should vote, it should say so and direct the person to an authoritative source—not guess.

5.3. Connect reporting to a coordinated correction process

AI providers already have reporting mechanisms. Google's Gemini guidance lets users flag inaccurate or unsafe responses. OpenAI's reporting guidance explains how to report content and notes that reports about unreliable sources may lead to mitigations.

The gap is not an absolute absence of feedback tools. It is whether an election-related report can be traced to the responsible organisation and corrected throughout the information chain.

A proposed Malaysian protocol could record the disputed answer, time, cited source and error type, then route the report to SPR, MCMC or the provider according to their respective responsibilities. Case identifiers, response logs and repeat-error checks would make performance auditable.

A coordinated procedure could work as follows:

Voter reports error → responsible organisation checks it → source or system is corrected → provider updates or limits the faulty answer → outcome is recorded.

SPR should correct official information within its control; providers should remedy errors in their systems. A coordinated procedure may be sufficient initially; creating a new task force should depend on demonstrated need and clear legal authority.

5.4. Develop a privacy-preserving SPR data interface

AI systems often retrieve election information from web pages, search results and commentary that may be inconsistent or outdated. A structured application programming interface (API)—a controlled connection between computer systems—could allow participating services to obtain published official information directly.

International precedents include Democracy Works' Elections API and Google's Civic Information API guidelines. Google specifies that voting-location and contest information should not ordinarily be cached for more than 24 hours when no different cache-control instruction applies.

SPR could examine a versioned, read-only interface for non-personal information, including election dates, general voting procedures, published polling arrangements, source links and update timestamps. Its reliability would depend on accurate underlying records and effective update processes: an API is not automatically a source of truth.

Personal voter information must remain protected. Services such as MySPR Semak involve individualised lookups and should not become unrestricted public data feeds. Secure integration or direct links to SPR would often be safer than transferring identifying information to a general-purpose AI service.

5.5. Use cryptographic verification as one layer of defence

An attacker might attempt to place misleading information in websites or retrieval sources later used by an AI system. Research on knowledge poisoning in retrieval-augmented generation, including knowledge-graph-based systems, demonstrates that maliciously introduced information can undermine system performance (Zhao et al., 2026). This is a technical risk, not evidence that such an attack has occurred against SPR or a Malaysian election.

Digital signatures can help confirm that a record came from a recognised source and has not been altered since signing. The C2PA provenance standard illustrates how cryptographic methods can support the verification of origin and history.

SPR could consider signing suitable official feeds, using trusted keys and version identifiers. Providers could verify signatures, check freshness and prioritise approved sources for high-risk procedural answers.

However, a valid signature does not prove that information is correct or current. A signed record can still be outdated, an official source can contain an error, and an AI system can misinterpret genuine information. Cryptographic verification should therefore complement—not replace—source validation, monitoring and correction.

Together, the five (5) safeguards form a proposed Malaysian election-information assurance protocol. Their value should be assessed through measurable indicators such as source accuracy, correction time, repeat errors, verification failures and privacy impacts. They are adaptations of established practices, not five newly invented concepts.

6. Accuracy must not become a pretext for censorship

Any Malaysian framework should distinguish verifiable procedural information from contested political expression.

Polling locations, registration deadlines, ballot instructions and certified results can generally be checked against an authoritative record. Political opinions, policy judgements, satire and predictions are different. They should not be treated as though SPR were the sole authority on political truth.

A factual error may call for a correction, source link or update notice. Removal or sanctions require a separate legal basis and a proportionate assessment. Decisions should be transparent, reviewable and protected against partisan application.

A pantun recorded in DBP's Khazanah Pantun Melayu Riau offers a pointed reminder about political conduct:

Kalau hendak membeli kain,
Jangan dibeli kain luntur;
Kalau awak jadi pemimpin,
Jangan sekali bermain kotor.

The message is relevant beyond politicians. Those who administer election information, develop AI systems or enforce the law should apply consistent standards and resist the temptation to manipulate information for partisan advantage.

That does not mean every mistake is deliberate misconduct. It means that the response to an error should be fair, evidence-based and proportionate, regardless of which political interests may benefit from it.

Source labels should help voters understand the status of an answer. A useful label might identify whether information comes from an official SPR source, a secondary report, an AI-generated synthesis or an unresolved claim. Such labels cannot guarantee accuracy, but they can make verification easier and reduce unwarranted confidence.

The objective is not to suppress political discussion. It is to help voters distinguish reliable procedural information from material that needs further checking.

7. What would make this a meaningful Malaysian contribution?

The original essay, Protecting Voter Autonomy in the Age of Artificial Intelligence, considers how AI may influence political judgement. The Invisible Architect examines how the selection and presentation of information can shape what people see. This second companion addresses the next step: who is accountable when that information is wrong, and can the error be corrected?

A defensible research contribution would combine a provision-by-provision legal analysis with an empirical audit of publicly accessible AI systems used by Malaysian voters. It could also test whether simple interface features make errors easier to recognise and report—without creating a false impression that an answer is guaranteed to be correct.

Make reporting easy, but distinguish feedback from correction

Two major AI services already provide ways for users to flag problems.

Google's Gemini Apps Help guidance allows users to mark a response as good or bad, select a reason, add comments and, where available, attach a screenshot. Google also explains that the associated conversation and its prompts and responses may be included in the feedback, alongside other submitted material. Users should therefore consider privacy before attaching sensitive information.

OpenAI's content-reporting guidance describes ways to report conversations, responses, shared links, GPTs and other content that may violate its Terms of Use or applicable laws. It also states that reported domains or content may be reviewed by its Model Quality team, which may apply filters or other mitigations to reduce future reliance on unreliable sources.

These systems provide useful starting points: feedback is available close to the answer, users can give additional context, and reports can inform quality or safety processes. However, neither help page promises that every factual error will be corrected, that a correction will propagate to every subsequent answer, or that the person reporting it will receive a confirmed resolution.

For election information, a more targeted process could separate three types of report:

• Incorrect factual answer: for example, a polling location that conflicts with the current official record.

• Source or freshness problem: a cited page is unavailable, outdated or does not support the claim.

• Safety or legal concern: content that may breach applicable rules and requires a separate assessment.

A report should capture the answer, date and time, source link, type of problem and, where available, the product or model version. A screenshot may help, but should be optional. The interface should explain what information will be submitted, allow users to remove unnecessary personal details and provide a reference number or status where operationally feasible.

This would make a report more actionable than a generic thumbs-down rating. It would also help the service distinguish dissatisfaction with an answer from a verifiable factual error.

Use epistemic markers to show what is known—and what is not

An epistemic marker is a label or cue that helps a reader understand the evidence and status behind a claim. In this context, useful markers would say more about the source and verification process than about the model's own confidence.

A simple answer card might display:

Source: Official SPR page
Source updated: Date shown by SPR, if available
Checked by this service: Date and time of retrieval
Status: Linked source located; claim not independently verified
Correction: Report a factual error

Other appropriate labels could include “Sources conflict—check SPR”, “Official result”, “AI-generated summary” or “Correction issued”. The exact wording should match what the system can actually establish. A service should not display “verified” merely because it has inserted a link; verification requires checking that the source is authoritative, current and supports the material claim.

There is a recognised technical basis for preserving source lineage. The World Wide Web Consortium's PROV model describes how to record the entities, people and processes involved in producing or changing information. It can support traceability between the original record, an intermediary's transformation and the delivered answer (World Wide Web Consortium [W3C], 2013). It is a data-provenance model, not a guarantee that the underlying information is true.

Research on explanations in AI-assisted decision-making also points to three useful design goals: helping people understand the system, recognise uncertainty and calibrate their trust appropriately (Wang & Yin, 2022). These are objectives to evaluate, not outcomes that can be assumed merely because a label or explanation has been added.

In a preregistered experiment involving 495 US-based adults with postgraduate degrees, Ojewale et al. (2026) found that visual confidence indicators improved participants' subjective ability to distinguish correct from incorrect AI outputs, but also increased behavioural agreement with incorrect answers in the experimental setting. Combining visual indicators with verbal uncertainty cues produced still greater over-reliance and less independent source-checking. These findings do not mean that every visual label is harmful. They show why uncertainty cues must be evaluated by how people actually use them, rather than assumed to improve trust calibration simply because they are displayed.

The study does not show that every visual label is harmful. It does demonstrate that adding more cues is not automatically better. For electoral information, source identity, dates, actual verification status and a clear route to check the official record may be more useful than a decorative confidence gauge.

Use simple visual and productivity tools where they solve a real problem

The study does not need an elaborate dashboard or a new application to be useful. A few carefully chosen tools could make the accountability chain easier to understand and operate.

A simple information-flow diagram can show the voter how an answer travels from SPR through an intermediary and an AI service. A source card can display the actual source, its date and the answer's verification status. A correction tracker can record when a report was received, which organisation is handling it, what was corrected and whether the issue recurred. These tools make different responsibilities visible without requiring the voter to understand the underlying technology.

For researchers and regulators, a lightweight dashboard could summarise the number of errors found, source failures, average correction time, repeat errors and differences across tested languages. A secure spreadsheet or case-management queue may be sufficient for an initial pilot; a complex visual interface should be justified by an actual operational need.

The design should remain readable on a mobile phone, accessible to people with disabilities and usable in Bahasa Melayu and English. Colour should not be the only way to distinguish statuses. Labels should use ordinary language, and detailed provenance should be available without overwhelming the main answer.

These choices are consistent with work on data provenance and human–AI interaction, but their effectiveness in Malaysian election services still needs to be tested. The aim is not to make the answer look more trustworthy. It is to make verification and correction easier.

Test whether the tools actually help

The audit could compare three versions of the same service:

1. A plain AI answer with its existing citations.

2. The same answer with source, date and verification-status markers.

3. The marked answer with an accessible error-reporting process and a clear confirmation or status route.

Researchers could assess whether participants identify incorrect answers more reliably, open authoritative sources more often, submit complete error reports, understand the status labels and calibrate their trust to the actual accuracy of the output. The test should include Bahasa Melayu and English, and should report differences among users rather than assume that one design works equally well for everyone.

The evaluation should also measure possible drawbacks: false reassurance from a label, excessive dependence on a visual indicator, confusion over technical terms and unwillingness to consult the official source. Results should determine which features are retained.

This design would combine established principles with a testable intervention. It would not presume that epistemic markers, reporting tools or visualisations have already been proven to improve Malaysian voters' decisions.

The original essay, Protecting Voter Autonomy, concerns the protection of political judgement; The Invisible Architect considers how information presentation can influence what people see. This study adds a practical accountability question: can voters identify the source of an AI answer, understand its status and get an error corrected?

The answer should come from legal analysis and empirical evaluation, not assumption.

Conclusion: make responsibility traceable.

When a voter receives a wrong answer, the practical questions are straightforward. Where did the information come from? When was it last updated? Who controlled the process that introduced the error? Can the voter report it? Can the responsible organisation correct it? And can the correction be verified?

Malaysia already has relevant electoral, communications and online-safety institutions. The task is to determine whether their powers and procedures connect effectively across the voter-information chain.

The strongest approach is not to treat every AI error as a crime or to create sweeping powers over political discussion. It is to establish clear sources, traceable updates, tested systems, accessible complaints and proportionate remedies, with independent oversight and protection for lawful expression.

Responsibility should follow function, control and the capacity to correct. Voters should not have to understand the technology behind an answer before they can expect a reliable correction.

Research note: This is a policy analysis, not a formal legal opinion. 

------------------

Designing AI Governance for Election Integrity in Malaysia: Lessons from Regional Approaches

By Muhammad Syawalfiza

Introduction

Imagine asking an AI assistant a simple question: “Where and when do I vote?”

The answer looks clear and confident. Unfortunately, the polling date is wrong.

The mistake might arise because the system uses outdated information, retrieves the wrong record, misunderstands a source or generates an answer that the source does not support. The problem may become more serious if the answer is repeated by news outlets, social-media accounts or other AI systems.

Who should be responsible? The Election Commission, which maintains the official information? The company that developed the AI model? The organisation that incorporated the model into its service? Or the platform through which the answer reaches voters?

The answer depends on what went wrong, what each organisation could reasonably control and what the applicable law requires.

My earlier article, From Official Election Data to AI-Generated Answers: Responsibility Across the Voter-Information Chain, explored how incorrect AI-generated election information might be traced and corrected. This companion examines a related question: how should Malaysia's legal and regulatory institutions support that process?

The issue has become more pertinent following Japan's latest legislative and policy developments concerning election misinformation. Japan is proposing measures to improve the visibility of trustworthy information, label certain AI-generated material, address financial incentives for misinformation and strengthen platform responses. Some of these objectives overlap with my earlier proposal, although the mechanisms differ.

This article examines the similarities and differences, drawing on selected experience from Japan, Singapore, South Korea and ASEAN. Its purpose is not to claim that Malaysia needs an entirely new regulatory model or that the proposed approach is unprecedented. Instead, it considers how established accountability principles might be applied more precisely to election information in Malaysia.

1. Malaysia's regulatory position: an evolving framework

Malaysia is not without rules relevant to AI-generated election information. Its existing legal framework includes the Communications and Multimedia Act 1998 (CMA), the Online Safety Act 2025 (ONSA), the Personal Data Protection Act 2010 (PDPA), the Cyber Security Act 2024 and election-related legislation. These instruments have different legal statuses, scopes and functions. Some impose binding statutory duties or create offences; others provide institutional powers or regulatory mechanisms. They should not be treated as interchangeable, and none should be assumed to create a general law of AI-output accuracy merely because it may be relevant to an AI service [4–8].

1.1 Online safety and communications law

ONSA took effect on 1 January 2026. It is binding legislation that establishes duties for specified licensed service providers and provides mechanisms for addressing legally defined harmful content, reporting concerns and regulatory enforcement [6].

An important distinction must be maintained: an incorrect AI-generated polling date does not automatically constitute harmful content under ONSA or a criminal offence under the CMA. The relevant statutory definitions, legal elements and facts must be examined. Draft guidance, policy proposals or comparative approaches cannot expand those statutory provisions unless adopted through a legally valid instrument.

Consider three situations.

An AI assistant accidentally gives the wrong polling date. A political operator deliberately uses an AI system to distribute false voting instructions. A platform receives credible reports of incorrect information but fails to investigate or correct it.

All three situations could cause concern, but they involve different conduct and potentially different legal consequences. A proportionate approach should distinguish the original error from any subsequent failure to address it. The existence of a policy concern does not, without more, establish a breach of binding law.

The same caution applies to section 233 of the CMA. It should not be described as making every false AI-generated answer an offence. Responsibility depends on the applicable statutory requirements and the evidence concerning the relevant person's conduct. Any proposed guidance or policy reform should be presented as supplementary to that legal analysis, not as a substitute for it.

1.2 The 2026 communications-regulatory amendments

The Malaysian Communications and Multimedia Commission (Amendment) Act 2026 (Act A1799) came into operation on 14 October 2026. As enacted legislation, it forms part of Malaysia's binding legal framework. Among its changes are functions relating to standards for the safety, security, integrity and reliability of information assets and communications services, as well as the review or audit of specified information and activities [7].

These provisions are relevant to the institutional context in which additional safeguards might be developed. They do not, by themselves, establish a universal legal duty requiring every AI model to provide factually correct election answers. The applicability of a proposed requirement must be assessed against the amended legislation, any valid subsidiary instruments or regulatory directions, and the service concerned.

The separate Communications and Multimedia (Amendment) Act 2026 (Act A1798) principally amends section 202 of the CMA in relation to universal-service initiatives involving national security. It should not be presented as a general AI-liability statute [8].

The distinction matters. A proposal can be sensible as a matter of policy without necessarily falling within an existing regulator's legal authority. Similarly, a regulator's statutory function may support further rule-making or oversight without itself resolving every question of liability for an inaccurate AI output.

1.3 Malaysia's proposed AI Governance Bill

Malaysia's proposed AI Governance Bill adds another layer to this discussion, but it is not, at the date of this analysis, enacted law. Its public consultation ran from 10 July to 1 August 2026. The official consultation record identifies the process as closed and at the pre-drafting stage [4, 5].

The proposed framework is intended to establish central oversight, risk-based governance requirements and clearer responsibilities for AI actors. The Ministry of Digital has also indicated that the proposed Bill is not intended to regulate AI outputs directly; illegal-content issues would continue to be addressed through existing laws and regulatory frameworks [4].

These statements describe a policy direction and consultation position, not binding obligations presently enforceable against all AI actors. The final scope, legal effect and institutional powers would depend on the text enacted by Parliament and any subsequent regulations, codes or guidance.

This creates a useful policy question: can system-level safeguards improve the reliability of AI-generated election information without turning AI governance into a general mechanism for controlling political speech?

A possible approach would be to require defined services, where legally authorised and proportionate, to demonstrate how they verify critical factual information, investigate errors and correct answers. At present, this should be understood as a policy proposal rather than a general existing statutory duty. Whether an individual communication warrants further legal action would remain a separate question under the applicable law.

The objective should be to improve the reliability of the process, not to assume that every inaccurate output deserves the same response.

2. Japan's latest measures: an important regional development

Japan's latest initiative is particularly relevant because it addresses how election-related information is presented and distributed on large social-media platforms. It should, however, be separated into its binding legislative component and its non-binding implementation material.

The legal status needs to be stated carefully. Japan's Act Partially Amending the Public Offices Election Act and the Act on Measures against Infringement of Rights, etc. Arising from the Distribution of Information through Specified Telecommunications was promulgated as Act No. 58 of 2026 on 17 July 2026. The relevant amended provisions are scheduled to take effect on 1 March 2027 [22, 23]. The Act is enacted legislation, but the relevant provisions are not yet in force as at 23 October 2026.

On 23 October 2026, Japan's Ministry of Internal Affairs and Communications (MIC) announced a public consultation on draft guidelines intended to support the effective implementation of measures to mitigate adverse effects on electoral fairness. The consultation is scheduled to run from 24 October to 24 November 2026. The detailed guidance is therefore still a draft at the date of this analysis and should not be described as binding law or as evidence of settled implementation practice [22].

2.1 What is Japan proposing?

The draft guidelines and associated legislative materials address several aspects of the online information environment [22, 23]. The amended legislation supplies the binding legal framework, while the draft guidelines indicate how the Japanese authorities may expect covered operators to implement or explain relevant measures. The latter may be revised following consultation and should therefore be treated as proposed guidance rather than final regulatory requirements.

Prioritising reliable information. The proposed implementation approach would allow platforms to give greater visibility to information from official websites, candidates, political parties and other sources considered reliable. This is a proposed platform practice within the legislative and guidance framework, not a legal determination that every official statement is correct.

Labelling AI-generated material. The amended election law provides for labelling specified AI-generated or modified images and videos distributed online in covered election contexts, subject to the statutory conditions and exclusions. This is the clearest example in this section of a binding legislative requirement, although its practical operation will depend on the statutory text, commencement and any applicable implementation arrangements.

Addressing financial incentives. Proposed platform measures include suspending monetisation for qualifying election misinformation that could harm electoral fairness, as well as addressing accounts and content that generate revenue. These measures should be distinguished from the statutory labelling provisions: their precise legal status and implementation depend on the enacted provisions and the final guidance.

Responding to coordinated abuse. The proposed measures also address bot-driven activity, account transparency, warnings about synthetic media and cooperation with election-related stakeholders. These are elements of the legislative scheme and draft implementation approach, but their exact obligations should not be inferred beyond the enacted text or treated as final until the consultation process is complete.

Public reporting. The amended platform law requires covered large-scale platform operators to take necessary measures appropriate to their service characteristics and to publish information about the measures they implement annually [23]. This is an enacted statutory obligation, subject to its scope, commencement and the conditions specified in the law.

The approach is significant because it addresses more than whether an individual post is true or false. It also examines how content is promoted, whether misleading material generates revenue, and whether platforms can demonstrate the measures they take.

In plain language, Japan is trying to make it harder for misleading election information to dominate what people see online, while making it easier to find material from sources considered reliable. That description captures the policy objective, but it should not be mistaken for a single universal legal duty applying to every service or every item of election information.

That does not mean the government has proved that every official statement is correct or that every alternative source is unreliable. Nor does the proposed approach require every disputed political statement to be removed. The legislation and draft guidance must be read in the context of their scope, implementation requirements and protections for legitimate expression. The enacted provisions are binding once in force; the draft guidance remains subject to consultation and possible revision.

2.2 How does this compare with my earlier proposal?

There is a genuine similarity between Japan's proposal and the mechanism discussed in my earlier article: both recognise the importance of reliable sources when people seek election information.

The difference is where each approach intervenes and what legal status each component has.

Japan's enacted and proposed measures are principally concerned with the wider information environment: what platforms display, how they handle synthetic media, what incentives they provide and how they respond to misinformation. The statutory provisions scheduled to take effect in 2027 should be analysed as future binding law, while the MIC consultation document should be treated as draft guidance that may inform implementation but is not yet final.

My earlier proposal focuses more narrowly on the information used to generate a particular AI answer. It suggests a privacy-conscious, read-only SPR data interface for defined election facts, supported by source links, update timestamps, version information, reporting and correction procedures [15]. This remains a research and policy proposal, not an existing Malaysian legal requirement or government system.

In plain language, Japan's proposed approach is like helping people find the most reliable signposts in a crowded public square. My proposal is like checking the original timetable before giving someone a travel time. The first helps people find dependable information; the second checks whether a particular answer agrees with the relevant record. Japan's approach is a comparative example of legislative and regulatory design; the Malaysian mechanism is a proposed application of accountability principles.

These are complementary functions. A platform might promote an official election website without guaranteeing that an AI assistant has retrieved the correct information or accurately reflected it in an answer. Conversely, a well-designed election-data interface would not, by itself, prevent bot campaigns, manipulated candidate videos or monetised misinformation from spreading across social media.

The comparison therefore supports a limited conclusion: my mechanism is more specific to the verification and correction of AI-generated factual answers, while Japan's proposed measures address a broader set of platform-level election risks. This does not establish that one approach is more effective overall, and the proposed Malaysian mechanism has not yet been empirically evaluated. Nor does Japan's example, whether enacted or still in draft form, establish that Malaysia has adopted or is legally required to adopt an equivalent model.

2.3 What Malaysia can learn from Japan

Japan's approach, considered as a combination of enacted legislation, forthcoming implementation and draft guidance, suggests that election integrity may require several distinct responses rather than one universal rule. It is a comparative example from which Malaysia may draw policy lessons, not a source of binding obligations in Malaysia.

Where the problem is an incorrect polling date generated by an AI assistant, authoritative retrieval and output validation may be appropriate. In Malaysia, these would presently be safeguards to be developed through policy, procurement, service design or future legislation, unless an existing legal instrument clearly requires them in the particular context.

Where the problem is a realistic synthetic video falsely depicting a candidate, labelling requirements and rules governing election campaigning may be more relevant. Japan's enacted provisions provide a regional example of how such risks may be addressed, but their scope and operation should not be assumed to apply in Malaysia.

Where the problem is coordinated distribution or financial incentives for misinformation, platform-level measures may be needed. Japan's legislative framework and draft guidance illustrate possible approaches, while Malaysia would need to identify the relevant statutory authority and decide whether any comparable measures should be proposed.

Where the problem is deliberate deception, the applicable electoral, communications or criminal law must be assessed separately. A comparative policy proposal or draft guideline cannot itself establish criminal liability in Malaysia.

The important lesson is to identify the type of problem and the point at which intervention can reasonably prevent or remedy it. A rule designed for platform amplification should not automatically be treated as a solution to AI factual errors, and a data-verification mechanism should not be expected to replace measures addressing deceptive campaign media.

Figure 1. Two complementary approaches to election-information reliability.

Japan's proposed approach focuses on the visibility, labelling and distribution of election-related content. The Malaysian proposal focuses on checking critical facts before an AI system answers and correcting errors when they occur. The Malaysian diagram represents a research proposal, not an existing government system.

3. Other regional lessons

Japan is not the only relevant comparator. Singapore, South Korea and ASEAN provide different ways of thinking about correction, transparency and shared responsibility.

3.1 Singapore: correcting false information

Singapore's Protection from Online Falsehoods and Manipulation Act 2019 (POFMA) permits correction directions in qualifying circumstances involving false statements of fact. Its operation and criminal provisions have distinct legal requirements [10].

Singapore also introduced election-specific restrictions concerning certain realistic but false representations of candidates in online election advertising [11].

The lesson for Malaysia is not to copy Singapore's legislation wholesale. It is to recognise that correction, removal and criminal prosecution are different interventions.

For example, if a service accidentally provides an incorrect polling date, a prompt correction may be the most appropriate first response. If an actor deliberately uses AI to deceive voters, a different legal assessment may be required.

3.2 South Korea: transparency and defined AI duties

South Korea's AI Basic Act entered into force on 22 January 2026. It establishes binding transparency requirements for covered generative-AI products and services, alongside other role- and risk-related duties [12].

South Korea also has election-specific restrictions concerning covered deepfake material used in election campaigning [13].

These mechanisms show that general AI governance and election law can address different aspects of the same problem. A transparency label tells a user that material is AI-generated; it does not establish that the information is correct. Factual verification and content labelling should therefore not be treated as substitutes for one another.

3.3 Japan: platform responsibilities and practical guidance

Japan's proposed measures complement its broader AI-governance framework and other existing laws [14]. Their significance for Malaysia lies in the way they link platform responsibilities to the characteristics of the service while addressing election-specific risks.

However, the relevant election guidelines remain under consultation as at 23 October 2026. Their final wording and implementation should be examined before drawing conclusions about how they operate in practice.

3.4 ASEAN: shared responsibility

ASEAN's AI Governance and Ethics Guide and expanded generative-AI guide recommend shared responsibility, testing, assurance, incident reporting, security and content provenance [2, 3].

These principles are important prior policy work. The present proposal should not claim to originate the idea that multiple actors should contribute to responsible AI deployment.

Its potential contribution is narrower: to examine how these principles can be applied to critical election facts in Malaysia through clearer source-verification, output-validation and correction arrangements.

4. A practical proposal for Malaysia

SPR already operates an official open-data portal containing election-related information, including results, candidates, party information, voter statistics and constituency-related datasets [15].

This provides a starting point. Before mandating a new API, policymakers should assess the existing datasets, their update frequency, their metadata and the availability of documented interfaces for the particular information needed by voters.

A possible assurance mechanism would have four stages.

Stage 1: Use authoritative information

For defined facts such as polling dates, locations and official procedural requirements, the system should identify an appropriate authoritative source and the relevant version of the information.

Stage 2: Check the answer

The system should verify that the retrieved record supports the final answer. Where practical, critical structured facts should be rendered directly from verified data rather than generated solely from model memory.

Stage 3: Show the source

The answer should provide a source link and meaningful information about the currency of the record. A user should be able to check where the answer came from.

Stage 4: Report and correct errors

A service should provide a usable reporting mechanism. Where an error is confirmed, the relevant organisation should investigate, correct or withdraw the answer as appropriate, and record the corrective action.

In everyday terms, this is similar to a newsroom checking a date against an official announcement before publishing it. The difference is that an AI service may generate thousands of answers through automated processes, so checks need to be built into the system and tested rather than left entirely to individual users.

These proposed measures should be proportionate to the function and risk of the service. A general-purpose chatbot, a specialist voter-information service and a social-media platform do not necessarily have the same duties or degree of control.

A mandatory SPR interface is one possible implementation, but it should not be presumed to be the only acceptable approach. Equivalent methods could be considered if they demonstrate comparable reliability and traceability.

5. Responsibility, privacy and freedom of expression

The proposed process must not assume that all failures have one cause or one responsible actor.

If an official record is outdated, the investigation should examine the source and whether the service had appropriate freshness controls. If the data are correct but the retrieval system obtains the wrong record, the integration process becomes relevant. If the service receives a credible error report and does nothing, that subsequent conduct should be assessed separately.

More than one actor may have relevant responsibilities. Any legal consequences must depend on the applicable duties, statutory authority and evidence—not merely on which service was last seen by the voter.

Privacy also matters. Election-information systems may process user queries, account information and interaction logs even if the answer itself contains only public facts. Malaysia's PDPA and its applicable exceptions must be examined in relation to each organisation and processing activity. The proposed AI and Personal Data Protection Framework is a separate policy development that also warrants consideration [17].

Cybersecurity presents another distinct question. A compromised source, altered record or disrupted service requires different controls from a model that misinterprets correct information. Applicable cybersecurity duties should be considered where relevant, without assuming that every election dataset or AI service falls within a particular statutory designation.

Finally, reliability requirements must not become a general mechanism for restricting political debate. An official polling date is a verifiable administrative fact. Political opinions, criticism of candidates and contested interpretations of policy raise different questions.

UNESCO's Guidelines for the Governance of Digital Platforms emphasise that addressing online harms must be balanced with freedom of expression and access to information [16]. Any Malaysian proposal should therefore define the relevant factual problem, the evidence needed to establish non-compliance, the authority to act and the available review mechanisms.

Official sources should be accessible and verifiable, but not immune from scrutiny or correction.

6. How should the proposal be evaluated?

It would be premature to claim that an SPR data interface or source-verification protocol will necessarily eliminate hallucinations or prevent election misinformation. These are proposals whose effectiveness needs testing.

A practical evaluation could examine whether the system:

gives accurate answers to questions about polling dates and locations; 

1. Retrieves the correct, current source;

2. Produces answers supported by the cited record;

3. Abstains when information is unavailable or conflicting;

4. Allows users to report errors successfully; and

5. Corrects verified mistakes within a time appropriate to their seriousness.

Testing should include supported Malaysian languages and different question styles. Performance should be compared with a baseline system that does not use the proposed controls.

The results would help determine whether the additional safeguards improve reliability, how much they cost to implement and which requirements are justified for different services.

Japan's proposed measures should also be evaluated separately. Source prioritisation, synthetic-media labelling, monetisation controls and platform transparency target different risks; their effectiveness should not be assumed from their adoption alone.

Conclusion

Japan's latest election-related measures are a relevant development for Malaysia. They reinforce the importance of making reliable information easier to find, improving transparency around certain AI-generated content and requiring covered platforms to take appropriate measures against risks to electoral fairness.

There is a clear similarity with my earlier article, but the mechanisms are not identical. Japan's proposed approach addresses the broader distribution and presentation of election-related information. My proposal focuses on verifying critical facts before an AI assistant answers, preserving source information and providing a route to correction.

The two approaches can complement one another. Better visibility of official information does not guarantee that every AI-generated answer is accurate, while a reliable data interface cannot by itself prevent coordinated misinformation or manipulated campaign media.

For Malaysia, the practical task is to assess existing SPR data infrastructure, identify the legal basis for any new obligations and determine which safeguards are justified by evidence. The CMA, ONSA, election law, data-protection framework and proposed AI Governance Bill have different functions; their relationship should be examined carefully rather than assuming that one instrument can solve every problem.

The proposed relay-race approach is best understood as an application of established accountability principles to a specific election-information problem. Its value will depend on whether it makes errors easier to detect, attribute and correct while preserving legitimate political expression.

The aim is not to promise error-free AI. It is to make critical election information more reliable, make failures easier to remedy and ensure that accountability follows the evidence.

------------------

References:

1. Cobbe, J., Veale, M., & Singh, J. (2023). Understanding accountability in algorithmic supply chains. In 

Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (pp. 1186–1197). 

Association for Computing Machinery. https://doi.org/10.1145/3593013.3594073

2. ASEAN Secretariat. (2024). ASEAN guide on AI governance and ethics. https://asean.org/book/asean-guide-

on-ai-governance-and-ethics/

3. ASEAN Secretariat. (2025). Expanded ASEAN guide on AI governance and ethics: Generative AI. 

https://asean.org/book/expanded-asean-guide-on-ai-governance-and-ethics-generative-ai/

4. Ministry of Digital, Malaysia. (2026, July 10). Ministry of Digital initiates engagement on proposed Artificial 

Intelligence (AI) Governance Bill. https://www.digital.gov.my/en-GB/siaran/Kementerian-Digital-Mulakan-

Libat-Urus-Cadangan-Rang-Undang-Undang-Tadbir-Urus-Kecerdasan-Buatan-%28AI%29

5. Unified Public Consultation Portal, Malaysia. (2026). Public consultation feedback: Proposed Artificial 

Intelligence (AI) Governance Bill. https://upc.mpc.gov.my/view-consultation/264

6. Malaysia. (2025). Online Safety Act 2025 (Act 866). Laws of Malaysia. https://lom.agc.gov.my/act-detail.php?

act=866&lang=BI

7. Malaysia. (2026). Malaysian Communications and Multimedia Commission (Amendment) Act 2026 (Act 

A1799). Laws of Malaysia.

8. Malaysia. (2026). Communications and Multimedia (Amendment) Act 2026 (Act A1798). Laws of Malaysia.

9. Infocomm Media Development Authority & AI Verify Foundation. (2024). Model AI governance framework

for generative AI. https://www.imda.gov.sg/resources/press-releases-factsheets-and-speeches/factsheets/

2024/gen-ai-and-digital-foss-ai-governance-playbook

10. Singapore. (2019). Protection from Online Falsehoods and Manipulation Act 2019. Singapore Statutes 

Online. https://sso.agc.gov.sg/Act/POFMA2019

11. Singapore. (2024). Elections (Integrity of Online Advertising) (Amendment) Act 2024 (No. 34 of 2024). 

Singapore Statutes Online. https://sso.agc.gov.sg/Acts-Supp/34-2024/Published/20241121

12. Republic of Korea. (2025). Framework Act on the Development of Artificial Intelligence and the Creation of 

a Foundation for Trust. National Law Information Center. https://www.law.go.kr/lsInfoP.do

13. Republic of Korea. (1994, as amended). Public Official Election Act, Article 82-8. National Law 

Information Center. https://www.law.go.kr/LSW/lsInfoP.do?lsiSeq=285629

14. Cabinet Office, Government of Japan. (2025). Act on Promotion of Research and Development, and 

Utilization of AI-Related Technologies. https://www8.cao.go.jp/cstp/ai/ai_act/ai_act.html

15. Suruhanjaya Pilihan Raya Malaysia. (n.d.). Portal Data Terbuka SPR. https://opendata.spr.gov.my/

16. United Nations Educational, Scientific and Cultural Organization. (2023). Guidelines for the governance of

digital platforms: Safeguarding freedom of expression and access to information through a multistakeholder 

approach. https://www.unesco.org/en/articles/guidelines-governance-digital-platforms

17. Department of Personal Data Protection, Malaysia. (2026, October 5). Public consultation paper 

No. 1/2026: Artificial intelligence (AI) and personal data protection framework. 

https://www.pdp.gov.my/ppdpv1/en/public-consultation-paper-no-1-2026-personal-data-protection-

framework/

18. European Commission. (2024). Guidelines for providers of very large online platforms and very large 

online search engines on the mitigation of systemic risks for electoral processes. https://digital-

strategy.ec.europa.eu/en/policies/dsa-guidelines

19. Juneja, P. (2024). Artificial intelligence for electoral management. International IDEA. 

https://doi.org/10.31752/idea.2024.31

20. Xu, J., Lee, T., & Goggin, G. (2024). AI governance in Asia: Policies, praxis and approaches. 

Communication Research and Practice, 10(3), 275–287. https://doi.org/10.1080/22041451.2024.2391204

21. National Institute of Standards and Technology. (2024). Artificial intelligence risk management 

framework: Generative artificial intelligence profile (NIST AI 600-1). https://doi.org/10.6028/NIST.AI.600-1

22. Ministry of Internal Affairs and Communications, Japan. (2026, October 23). [Public consultation on draft 

guidelines for the appropriate and effective implementation of measures to mitigate adverse effects on electoral

fairness] [In Japanese]. https://200.180.31.150.static.iijgio.jp/menu_news/s-news/01ryutsu02_02000497.html

23. House of Representatives of Japan. (2026). [Outline of the bill partially amending the Public Offices 

Election Act and the Act on Measures against Infringement of Rights, etc. Arising from the Distribution of 

Information through Specified Telecommunications] [In Japanese]. 

https://www.shugiin.go.jp/internet/itdb_gian.nsf/html/gian/honbun/youkou/g22105026.htm

Beyond Deepfakes: How AI Can Narrow the Choices Voters See

From the impending United States midterms to Israel's October 2026 elections and Malaysia's preparations for its 16th General Election (GE16), we like to believe our votes are the product of independent thought. We gather facts, weigh policies and make rational choices. But what happens when the digital systems we rely on to gather those facts present information in ways that shape our judgments—sometimes unintentionally, and sometimes through deliberate misuse?

This companion analysis builds on Protecting Voter Autonomy in the Age of AI by examining how generative AI intersects with behavioural economics. The concern extends beyond false content: AI systems may also shape which political information voters encounter, how it is presented and which options they consider.

​As we navigate the 2026 global election cycle, here are four of the most critical, often-overlooked dynamics at play, and what experts say we can do to build resilience against them.

​1. The Demand-Side Challenge: Why Misleading Political Content Can Gain Traction

Political misinformation can spread not only because people mistake false information for truth, but also because prior beliefs, group identities and emotional reactions can influence how they evaluate and share political content.

In a polarised environment, some voters may be more inclined to engage with material that reinforces their existing views or criticises a political opponent. This does not mean that voters knowingly want falsehoods, or that every person who encounters misleading content will believe or share it.

Disclosure labels and watermarks may help, but they should not be treated as a complete solution. Their effectiveness depends on how people interpret them, whether they trust the source, the context in which the content appears and what other information is available. Experimental research on AI-generated political images suggests that well-designed labels can reduce perceived credibility and, in some settings, reduce belief in misleading claims or willingness to share them. However, effects vary with the label design and research setting. Labelling can be a useful safeguard, but it should not be treated as a guarantee that voters will reject misleading content.

The practical question is therefore not whether voters secretly want lies. It is which combinations of disclosure, source verification, timely correction and media literacy best reduce the belief and spread of misleading political content.

These risks are becoming increasingly relevant to election campaigns, although deliberate psychological manipulation must be established case by case.

In the lead-up to Israel's 2026 elections, the potential use of AI to shape political perceptions warrants close scrutiny. In his October 14, 2026 essay, Dean Livne, acting director-general of Israel's Central Elections Committee, warns that AI systems may distort voter choices through misleading information, incomplete candidate coverage and opaque recommendations.

These concerns identify credible risks, but they should not be treated as proof that specific political campaigns are already using AI to manipulate voters' emotions. Establishing such activity requires separate evidence, such as documented campaign materials, independently verified incidents or credible investigations.

Whether confirmation bias or motivated reasoning explains why a person shared a particular deepfake should be established through evidence about that incident or audience.

​In Layman's Terms (Visualize This): Imagine a political deepfake is a box of junk food. Regulators think that putting a giant "Unhealthy Ingredients" warning label on the box will stop people from eating it. But if that junk food perfectly satisfies a person's craving for outrage against a politician they already dislike, they may still be tempted to consume it. The problem isn't just that AI makes the junk food; outrage can make it tempting to consume.

​2. The Illusion of Algorithmic Parity: Confident but Incomplete

​We often focus on whether AI can write a better speech than a human. But the true danger lies in its scale and manufactured confidence. The States United Democracy Center's study found that AI platforms directed voters to official state election websites less than half the time overall. ChatGPT mentioned a state website in 39.4% of responses, while Google AI did so in 55.6%. Both systems also showed gaps in the completeness and sourcing of election information.

The Brennan Center’s 2026 tests found that several chatbots struggled to recognise synthetic images, including images generated by their own tools. After OpenAI released a tool to verify images generated by its systems, ChatGPT identified some of its own tested images but not all. The broader tests found that most chatbots still struggled to identify AI-generated images.

​In Layman's Terms (Visualize This): Imagine a highly confident tour guide who speaks perfectly but only shows you half the map. They assure you that you've seen the whole city, and because they sound so professional, you don't bother to double-check. The AI isn't necessarily lying to you; it is just confidently leaving out crucial pieces of reality.

​3. The Erasure of the Local: How Algorithms Narrow Our Choices

​The Dutch findings illustrate how AI chatbots can act as “choice architects”: the parties they recommend may influence which political options voters notice. An audit by the Dutch Data Protection Authority, published in March 2026, found that specific local political parties were recommended as the first choice in fewer than 1% of the chatbot voting recommendations examined, despite local parties having received more than 30% of the votes in the preceding municipal elections.

The regulator identified limited and less up-to-date online information about local parties as one possible explanation. The findings demonstrate a problem with the representation of political choices in the systems tested, but they do not establish a single cause for that behaviour.

For Malaysia, this raises a legitimate question rather than a proven outcome: would widely used AI systems provide complete and balanced information about parliamentary, state and local political choices? The answer should be established through systematic testing across Malaysian parties, languages and constituencies, rather than assumed from overseas findings.

Malaysia's electoral map raises a separate issue: malapportionment, where constituencies have different numbers of voters and therefore unequal relative voting weight. Gerrymandering—the strategic drawing of boundaries for political advantage—is a distinct problem. Neither is the same as an AI system omitting candidates from its answers. Whether AI systems provide complete and balanced information about Malaysian candidates should be tested across constituencies and languages, measuring candidate coverage, factual accuracy and source quality. The Dutch findings justify conducting those tests; they do not establish that Malaysian systems reproduce the same pattern. Any numerical comparison of Malaysian constituencies should identify the election year, cite official Election Commission electorate figures and show how the disparity is calculated.

​In Layman's Terms (Visualize This): You walk into a massive supermarket (the AI) to buy groceries (political information). You want local, farm-fresh produce. But the store's systems have left local items out of sight in the back warehouse and only stocked the main aisles with two giant national brands. You end up buying a national brand, thinking it was your own free choice.

​4. The Decentralisation Dilemma: The Open-Weight Models and Enforcement”

Regulatory approaches vary, but safeguards aimed at major AI providers cannot address every form of misuse. Some open-weight models can be run locally, potentially allowing users to generate synthetic political material without relying on a major commercial chatbot. The practical barriers vary with hardware, model capability and technical expertise. This complicates enforcement when content is generated outside a provider's hosted service, but it does not mean every local model is unrestricted or that oversight is impossible.

In a July 2026 opinion piece, Chithra Latha Ramalingam argued that Malaysia's proposed AI Governance Bill may not adequately address election-related deception. She questioned whether a risk framework that prioritises physical harm would fully recognise societal harms, including misleading voters. With GE16 ahead, her argument reinforces the case for examining these gaps before the next campaign.

A Malaysian case illustrates the importance of distinguishing allegations of AI-enabled election deception from independently verified findings. During the July 2026 campaign for the Chuah by-election in Negeri Sembilan, MCA alleged that AI-generated images and doctored material were being used to link its candidate to a viral cash-handout video. The Star reported that the candidate lodged a police report on July 29 and noted that the underlying cash-handout footage remained unverified at the time of publication.

This incident highlights the potential for disputed or manipulated media to become part of an election controversy. However, the report alone does not establish that all the disputed material was AI-generated, who created it, or whether an investigation subsequently confirmed the allegations. Those questions require separate verification.

​In Layman's Terms (Visualize This): Regulators are trying to stop bad information by controlling major broadcasting towers. But some technically capable users can run their own pirate radio stations from home. The government may be guarding the front door while new side doors keep opening.

​Building Civic Resilience: Moving Beyond Doomerism

​It is easy to look at these behavioural vulnerabilities and adopt a fatalistic "doomer" mindset. But human cognition is not a lost cause; it is a system that requires updating. We cannot rely solely on legal bans or tech companies to save democratic autonomy.

As the Malay proverb puts it, “bagai menghela rambut dalam tepung: rambut jangan putus, tepung jangan terserak.” DBP records this proverb as a reminder that difficult matters must be handled with great care. That is precisely the balance election safeguards must strike: addressing AI-enabled deception without unnecessarily damaging legitimate political speech, fair competition or due process.

Expert advice for 2026 points toward immediate, proactive resilience:

1. ​Immediate Electoral Intervention: Dr. Chithra Latha Ramalingam argues that Malaysia cannot wait for overarching AI bills to pass. Her recommendations include requiring labels on AI-generated political campaign material, assigning an accountable person to each political avatar, and asking the Election Commission to act before GE16 rather than waiting for the proposed bill to pass.
2. ​A Multipronged, Defensive Approach: At an August 2026 roundtable convened by Singapore's S. Rajaratnam School of International Studies, panellists concluded that no single technical solution could adequately address disinformation. They advocated a multipronged approach combining AI-based and non-AI-based defence.
3. ​Build Algorithmic Literacy: Discussing the 2026 Israeli elections, expert Haleva Amir warns that we cannot regulate our way out of this: "Laws will not be enough or sufficient because technology always exceeds the law... The real answer to all these things should come from the people…by raising awareness and educating people and giving them tools to criticise and reflect."

Voters should learn to recognise cognitive ease—the feeling that information is familiar or easy to process—and remember that fluency is not the same as accuracy. When an AI answer sounds convincing, the next step is to check its sources and verify important claims independently. As the Malay proverb reminds us, “sediakan payung sebelum hujan”: safeguards and public education must be prepared before misleading content spreads, not improvised after a crisis. Protecting democracy in 2026 no longer just means securing the ballot box; it means safeguarding the human mind's capacity to choose.

Protecting Voter Autonomy in the Age of Artificial intelligence (AI).

An evidence-informed revision of Dean Livne’s proposals, with implications for Malaysia

By Muhammad Syawalfiza.

Central finding. The scientific record supports precautionary, testable safeguards for AI used in election information. It does not establish that chatbots have already changed an election result in Malaysia. The strongest evidence is more specific: controlled studies show that AI-generated messages and dialogues can shift political attitudes; audits show that chatbots can produce factual errors or systematically omit candidates; and Malaysian research documents a pre-existing misinformation environment that requires locally validated, multilingual protection.

1. The democratic question: inform, do not decide

In his 14 October 2026 By Invitation essay, “Israel’s election overseer on how to stop AI distorting votes,” Dean Livne, acting director-general of Israel’s Central Elections Committee, draws a useful line between helping citizens understand politics and a chatbot making a voting choice on their behalf. AI can explain an electoral system, summarise manifestos, compare verifiable policy positions and help a voter interrogate claims. The risk emerges when an answer that appears impartial and personalised silently embeds stale information, uneven candidate coverage, unreliable sourcing or value judgments that belong to the voter (Livne, 2026).

Livne proposes five core commitments for AI companies: do not recommend or rank candidates for voters; make electoral information verifiable; defend information sources against manipulation; prevent political profiling based on personal data; and make safeguards testable through evaluation and independent scrutiny. These are a sound starting point. The evidence suggests they should be retained but operationalised more precisely, with measurable performance standards, safeguards against multi-step abuse, multilingual testing and public incident reporting.

The objective should not be to suppress political discussion. A blanket refusal to discuss candidates, parties or policy would reduce civic utility, particularly for first-time voters or people facing language and accessibility barriers. The more defensible boundary is between neutral, source-grounded civic assistance and the system taking the final evaluative step for the voter. An assistant may compare parties against criteria selected by a user; it should not present a single candidate as the objectively “best” choice, convert sensitive personal information into a political profile, or obscure the assumptions behind a ranking.

2. What the scientific evidence actually establishes

2.1 AI can persuade, but the size and setting of the effect matter

The clearest direct evidence comes from preregistered experiments. Lin et al. (2025), published in Nature, randomly assigned participants in the United States, Canada and Poland to converse with an AI model instructed to advocate for one of the leading candidates. Across the three election settings, the dialogues produced statistically significant shifts in candidate preference and voting intentions. The US experiment involved 2,457 participants starting the study; the Canadian and Polish experiments also reported effects on candidate preference. In a separate Massachusetts ballot-measure experiment, the authors found sizeable shifts in support for psychedelic legalisation. Some effects were larger than those commonly reported for conventional political advertisements.

This is strong evidence of a capability: under experimental conditions, a conversational model can influence stated preferences. It is not proof that ordinary chatbot use has altered a national election outcome, changed actual ballots, or will produce the same effects in Malaysia. The experiments measured pre-to-post responses in controlled interactions; campaign context, prior knowledge, election system, language, model design and persistence of effects all matter. The defensible inference is that personalised dialogue should be included in election-safety testing rather than treated as equivalent to a static web page.

The same study makes factual accuracy a governance issue. Researchers rated 8,134 statements produced by the AI models; overall accuracy was generally high by their rating method, but they found notable differences between advocacy conditions. In all three countries, the models advocating for the more right-leaning candidates produced more inaccurate claims than those advocating for the more left-leaning candidates in those specific experiments (Lin et al., 2025). This result must not be generalised into a claim that every model is inherently biased against or in favour of a political side. It does show why matched, candidate-by-candidate accuracy testing is needed: a superficially balanced tone does not guarantee equal evidentiary quality.

A complementary set of three preregistered experiments involving 4,829 participants found that LLM-generated messages shifted policy attitudes relative to neutral control messages and were about as effective as messages written by ordinary human participants (Bai et al., 2025). The results included contested policy topics such as an assault-weapons ban, a carbon tax and paid parental leave. The effects were generally small, but statistically meaningful and comparable to ordinary human persuasion. These experiments test policy messaging more directly than candidate choice; they show persuasive capability, not election-scale impact.

A separate online experiment with 739 participants examined a candidate-operated AI chatbot against a conventional campaign video. Lee and Yuan (2025) found that interactive, personalized chatbot exchanges increased candidate likability and social presence; a sense of behavioural interdependence increased reported voting intention, while perceptions of artificiality and lower authenticity worked in the opposite direction. This is direct evidence that political chatbots can shape candidate impressions and stated intention, but it studies a campaign-controlled chatbot rather than a neutral public chatbot giving voting advice, and stated intention is not an actual vote.

Importantly, the broader synthesis is more nuanced than a claim that AI is inherently more persuasive than humans. Hölbling et al. (2025) meta-analysed seven studies comprising 17,422 participants and 12 effect-size estimates. They found no statistically significant overall difference between LLMs and human communicators in persuasive effectiveness (Hedges’ g = 0.02, p = .530), alongside substantial heterogeneity between studies (I² = 75.97%). The evidence therefore suggests that LLMs can persuade and may match human persuasion, but that the result depends heavily on context and implementation. This strengthens the case for evaluation rather than sensational claims about inevitable mass persuasion.

At the wider misinformation level, Park and Nan (2026) scoping review synthesized 24 empirical studies and concluded that generative AI can both create convincing misinformation and assist detection or mitigation; results for audience impact and safeguards were mixed. This review is useful as a map of the field, but it combines heterogeneous tasks and does not establish the size of an election-specific effect in Malaysia.

2.2 Political deepfakes: experimental effects are possible, but are uneven

Murphy, Ching and Meehan (2025) experimentally compared text, synthetic audio and low-cost political deepfake videos using a fabricated scandal (N = 443). The content could plant false memories, and some conditions reduced reported voting intention for one target; deepfakes were not consistently more persuasive than text. A follow-up with 300 participants found no lasting measurable changes to beliefs or memories after debriefing. The study supports a credible risk and the value of rapid correction, not the claim that deepfakes inevitably alter elections.

Field evidence further cautions against equating technical capability with a flood of election content. Kruschinski and Votta (2026) analyzed nearly 69,000 social posts in Germany’s 2025 federal election and about 65,000 in the Netherlands’ 2025 election. Visual generative-AI content appeared in about 1.4% and 1.3% of posts, respectively, but was disproportionately used by smaller actors and far-right parties; 87% of detected AI visuals in Germany and 75% in the Netherlands were unlabeled. Their synthesis concludes that who deploys AI, what the content depicts and how it is amplified may matter more than assuming synthetic visuals automatically persuade more than ordinary images.

Conrad et al. (2025) tested labels on AI-generated political images in an EEG study (N = 26) and an online experiment (N = 276). Labels affected perceived trustworthiness and intentions to follow political recommendations, indicating that disclosure design matters. These modest, constructed experiments do not show that labels reliably neutralize misinformation in real campaigns. Provenance labelling should therefore complement—not replace—source verification, contextual correction and platform-level controls.

2.3 Chatbots can distort the political choice set without explicit persuasion

A different risk is omission or skewed coverage. In October 2025, the Dutch Data Protection Authority reported that four widely used chatbots repeatedly placed the same two parties at the top of voting advice: in more than 56% of cases, the first recommendation was either the PVV or GroenLinks–PvdA, regardless of how the prompt was framed; one chatbot did so in more than 80% of cases. In a follow-up study published in March 2026 on municipal elections, the regulator reported that specific local parties were recommended first in less than 1% of chatbot recommendations, despite local parties having received more than 30% of votes in the previous municipal elections (Autoriteit Persoonsgegevens, 2025, 2026). These are regulator-led audits, not randomized voter-outcome experiments, but they demonstrate a concrete representational failure: voters can receive an incomplete or distorted map of their choices.

The UK think-tank Demos tested five chatbots and AI search services during the 2026 Scottish pre-election window and found factual errors in 34.1% of responses. Errors included incorrect election dates, incorrect voter-identification information, a fabricated candidate and invented allegations about candidates. A survey conducted in the run-up to the local elections found that 20% of adults had used an AI chatbot or AI search service for election information, 47% were concerned about inaccuracies, and 49% did not trust chatbots for election-related information (Hancock & Moores, 2026). This is a time-bounded audit by an independent think-tank rather than a peer-reviewed experiment, but it is highly relevant to operational accuracy and trust.

A peer-reviewed 2026 audit offers a complementary, narrower snapshot. Jang, McGregor and Neill examined 107 election-related prompts to ChatGPT-4o mini from 25 October 2024 and assessed 592 verifiable claims; 57 claims (9.63%) were coded as inaccurate. They also documented outdated candidate information, significant contextual omissions and plausible-sounding but unsupported responses. Because this was one model version and one date, the percentage is not an estimate of current chatbot error rates; it does establish why audits should score omissions and timeliness as well as literal falsehoods.

These findings support Livne’s proposal to avoid vote recommendations, but the rule must be implemented carefully. It should prohibit direct candidate endorsements, rankings and compatibility scores that purport to decide how someone should vote. It should not prevent neutral, auditable comparison of parties’ published positions. Where a voter asks for a recommendation, the system should explain its boundary, ask which policy criteria matter to the voter if appropriate, and offer a structured comparison with source links and explicit uncertainty.

2.4 Source manipulation and multi-step operations create a different attack surface

A chatbot does not need to be directly compromised to reproduce a political manipulation. If retrieval systems draw on an information environment that has been flooded with fabricated interviews, duplicate articles, fake social accounts or deceptive “news” pages, a model may present contaminated material in a calm, authoritative voice. Livne describes this risk as attackers manipulating the sources that AI systems consult (Livne, 2026). International IDEA likewise warns that generative AI can help actors create deepfakes, distort facts, microtarget audiences and overwhelm information systems, while stressing that empirical research on AI’s electoral impact remains limited (Juneja, 2024).

Anthropic’s 14 October 2026 account offers a useful but company-reported operational lesson. It analysed a random, privacy-preserving sample of 250,000 US Claude conversations from 21–27 September 2026. Broadly political, public-policy or social-issue conversations represented about 5.75% of the sample; election or voting conversations represented about 0.51%; requests asking whom or what to vote for represented about 0.01%; and requests classified as violating election-integrity policy represented about 0.002%. These figures describe one product, one country, one week and the company’s own classification method; they should not be treated as population-wide use estimates or independent confirmation of safety.

More consequentially, Anthropic reports that joint testing with the Institute for Strategic Dialogue and the McCain Institute found that the model generally refused direct harmful requests, but could miss a malicious campaign distributed across smaller, individually plausible tasks—such as audience research, building a site, drafting content and planning distribution. Anthropic says it added classifiers to better detect novel and multi-step misuse (Anthropic, 2026). The lesson for all providers is that request-by-request refusal tests are insufficient. Red teams must test the whole campaign workflow and evaluate whether benign-looking fragments combine into coordinated deception.

An independent Brennan Center expert audit in August 2026 found a similar need to test both information provision and content generation. In limited exploratory tests, chatbots often challenged overt election-conspiracy prompts, yet factual or citation failures remained, and image, audio or video tools did not consistently block deceptive political assets. The report’s methods appendix describes an API-based red-team exercise and a limited set of scored outputs; it is an independent operational audit, not a peer-reviewed representative study. Its strongest contribution is to show that refusing a blatant conspiracy query is not the same as reliably sourcing election facts or preventing the generation of deceptive campaign material.

2.5 What Malaysian research establishes—and what it does not

Malaysia already has credible research on election misinformation and digital campaigning, but a smaller body of directly relevant evidence on generative-AI chatbots influencing Malaysian vote choice. This distinction is important: evidence about social-media misinformation establishes the local information-risk context, not the causal effect of AI chatbots.

Neyazi, Tan and Kuru (2026) used a two-wave panel survey during Malaysia’s 15th General Election, with 1,800 respondents in wave one and 901 in wave two. They found that conservative respondents were more likely to believe and share misinformation on Facebook and WhatsApp. Religiosity had no direct association with misinformation susceptibility; however, stronger religious outlooks moderated sharing behaviour among conservative respondents, who were less likely to share misinformation. The result argues against simplistic demographic stereotypes. Risk assessments should be based on tested behaviours and exposure pathways, not assumed attributes of ethnic, religious or political groups.

A 2026 mixed-methods study of political communication in Malaysia, Indonesia and Thailand adds recent regional evidence (Suherman et al., 2026). It combined a survey of 611 adults, analysis of 15,000 public social-media posts from January to June 2025, 120 political-content artifacts and 30 interviews; 205 survey respondents were in Malaysia, of whom 78% reported seeing or interacting with content they perceived as AI-related and 60% reported emotional effects. These are respondent reports, not forensic proof that each item was AI-generated, a nationally representative prevalence estimate or a causal finding about vote choice. The cross-sectional association between exposure and emotional engagement is a reason to study local pathways, not a basis for claiming that AI changed Malaysian election outcomes.

Tapsell (2023) analysed the growth of video campaigning during GE15, showing the importance of platform-native content, livestreaming and campaign innovation in an already competitive digital environment. Balakrishnan et al. (2021) analysed a corpus of 190,224 tweets from the 2018 General Election using machine-learning sentiment classification and human content analysis. Their best reported English-language sentiment model reached accuracy in the approximate 64–67% range; their results also differed between Malay- and English-language tweets. This is not evidence that sentiment models can identify misinformation. Rather, it is a warning that simple automated classifiers can be imperfect and that performance can vary by language and corpus.

Sabri (2024) provides a Malaysia-focused narrative review of AI-generated deepfakes and misinformation, drawing on a limited set of literature from 2019–2023. It is useful as a threat map, but a narrative review cannot establish the prevalence of deepfakes in Malaysian elections or quantify their causal impact on voters. Zailani, Pirus and Adnan (2026) provide a recent literature-oriented discussion of Facebook, AI and political support in Malaysia and internationally; this should likewise be treated as contextual synthesis, not as an experimental estimate of chatbot persuasion. 

Daud (2020), using qualitative legal analysis of GE14, highlights an enduring governance risk: rules against false information can be turned into instruments of political censorship if definitions, due process and independent oversight are weak.

Taken together, Malaysian evidence supports three priorities: election-specific misinformation monitoring; robust language and platform coverage; and rights-respecting response mechanisms. It does not yet justify a claim that AI chatbots have changed Malaysian election results, nor does it justify treating any political, ethnic or religious community as intrinsically more vulnerable. The missing evidence should become a defined research programme rather than be filled with speculation.

3. Evaluation of Livne’s five commitments

1. Preserve voter autonomy. Adopt a clear prohibition on unsolicited or requested direct endorsements, candidate rankings and “best match” scores framed as voting advice. Continue to allow policy explanation, comparison of verifiable promises, and debate preparation. Any comparison should state the criteria used, distinguish manifesto commitments from independently verified performance, include material trade-offs and let the voter reach the conclusion.

2. Make election information verifiable. For registration, eligibility, polling locations, deadlines, identification requirements, voting procedures and results, prioritise current information from the responsible election authority. In Malaysia, this means a governed, continuously updated link to official Election Commission material, rather than general web popularity. Display source links, dates, jurisdiction and uncertainty. If no authoritative information is available, say so instead of improvising.

3. Defend the information supply chain. Use provenance checks, source-quality assessment and corroboration from genuinely independent sources. Repetition across websites that copy one another must not be treated as independent confirmation. Test retrieval-augmented systems against poisoned sources, fake interviews, synthetic candidate statements, prompt injection and coordinated publication campaigns. Preserve a rapid correction path with electoral authorities and trusted fact-checkers.

4. Prohibit political profiling from personal data. Do not use conversational memory, sensitive disclosures, inferred anxieties, health or financial context, location traces or predicted turnout to steer a person’s electoral choice. Apply data minimisation, purpose limitation, short retention periods and strict access controls. Ban political persuasion based on inferred vulnerabilities, including when targeting is performed by an intermediary rather than the AI provider itself.

5. Make safeguards independently testable. Publish an election-specific evaluation protocol, aggregate results, known limitations and remediation status. Permit independent researchers to conduct privacy-preserving audits and repeat tests across model versions. Self-reported evaluations by providers are useful disclosures, but should not substitute for external replication, adversarial testing or public accountability.

These controls can be organised through a continuous risk-management lifecycle: govern, map, measure and manage risks before deployment and throughout the election period. That approach is consistent with the NIST AI Risk Management Framework’s Generative AI Profile, which emphasises context-specific risk identification, measurement, mitigation and monitoring rather than a one-time compliance check (National Institute of Standards and Technology [NIST], 2024). International IDEA similarly recommends that electoral management bodies prepare for both the benefits and the risks of AI and notes that research on AI’s electoral impact remains comparatively limited (Juneja, 2024).

The supplied policy documents reinforce the need for layered controls but should be cited according to their type. RAND’s 2024 perspective is a literature-based map of threats and response options; Gehringer, Nehring and Łabuz (2024) explicitly caution that deepfake effects are difficult to attribute and may be cumulative; and the European Parliament’s 29 April 2025 text is a committee working document proposing priorities for preparedness, red-teaming, media literacy and electoral resilience, not a binding law or an evaluation of intervention effectiveness. Together they support a risk-management rationale, not a causal estimate of voter impact.

The five commitments are substantively sound, but each needs a measurable control and an accountable owner. Without that operational layer, “neutrality,” “verifiability” and “privacy” risk remaining aspirations rather than properties that can be tested before polling day.

4. Strengthened framework: from principles to controls

A. Test the voter experience, not just the model prompt. Run matched-prompt audits across candidates and parties, using identical questions about integrity, competence, policy and record. Repeat prompts with different wording, languages, levels of detail and multi-turn follow-ups. Measure whether outputs vary unfairly, whether important candidates disappear, and whether the model adds uneven caveats or selectively demands stronger evidence from one side.

B. Establish election-specific factual benchmarks. Build a version-controlled, publicly documented test set containing official electoral procedures, candidate identity, manifesto claims, dated public statements, documented corrections and common local misinformation narratives. Score every factual claim for accuracy, source support, date sensitivity and correct jurisdiction. Separate “claim is false” from “claim is unverified,” “opinion” and “campaign promise.”

C. Expand multilingual and inclusion testing. For Malaysia, test Bahasa Melayu and English as a minimum, and include Mandarin, Tamil and locally important languages or dialects where practical. Evaluate code-switching, transliteration, colloquial language, minority and local parties, candidates with limited web presence, and accessibility formats. A model that performs well in English but misstates procedures or omits candidates in Malay or another language is not election-ready.

D. Evaluate multi-step and coordinated misuse. Test staged workflows that split a deceptive campaign into research, asset creation, web publishing, account creation, targeting and distribution tasks. Combine technical red-teaming with human review of suspicious patterns, clear policy rules and escalation channels. Use privacy-preserving signals and proportionate enforcement to avoid turning ordinary civic work, journalism or political organising into presumptive abuse.

E. Treat synthetic-media labels as one layer, not a cure. Watermarks and provenance standards can help identify some AI-generated material, but do not establish that a claim is true, identify all manipulated media or survive every transformation. Combine them with origin metadata where possible, source verification, rapid rebuttal, platform enforcement and media literacy. Disclosure must be legible, placed where content is actually encountered and backed by a workable complaint and correction process.

F. Protect rights and procedural fairness. Define prohibited conduct narrowly around deception, impersonation, suppression, covert coordinated manipulation and misuse of personal data. Provide notice, an explanation and an appeal route for significant enforcement actions. Establish independent oversight and publish transparency reports. Disinformation controls should not become a proxy for suppressing criticism, opposition speech, satire or legitimate uncertainty.

G. Measure operational readiness before election day. Set minimum release criteria and a freeze-and-review process for election-sensitive changes. Track factual error rates, citation validity, candidate coverage, party and language parity, privacy failures, policy-evasion rates, incident acknowledgement time and correction latency. Publish a public incident channel and identify the teams accountable for decisions.

5. How current AI-provider policies fit the evidence

OpenAI’s 27 May 2026 update describes four lines of work: directing users to reliable voting and results information; supporting cyber defenders; improving provenance through approaches such as C2PA metadata and SynthID watermarks; enforcing policies against election interference and scaled deceptive advocacy; and evaluating political bias (OpenAI, 2026). These are relevant controls, especially when they are connected to official sources and security operations. OpenAI also acknowledges that provenance markers are not a complete solution. The appropriate interpretation is that these are provider-reported safeguards and commitments, not independent proof that errors or manipulation have been eliminated.

Anthropic’s 14 October 2026 report describes a related approach: official voting information, political-evenhandedness evaluations, misuse classifiers, enforcement, and external pressure-testing with independent organisations. Its reported finding that multi-step harmful operations can evade request-level safeguards is especially important, because it maps directly onto the need for campaign-level red-teaming (Anthropic, 2026). Its 99% political-evenhandedness score is a result on the company’s own evaluation, and should be interpreted only with the published test design, scope, limitations and independent replication in view. High scores on one benchmark cannot, by themselves, demonstrate factual accuracy, complete candidate coverage or absence of privacy harms.

For regulators and election-management bodies, neither company policy should become the benchmark merely because it is current or well resourced. The standard should be outcome-based and provider-neutral: can an ordinary voter receive timely and accurate official information; are candidates treated comparably; can source contamination or staged abuse be detected; are personal data protected; can independent auditors reproduce the findings; and are failures corrected before they can do material harm?

6. A research and implementation agenda for Malaysia

Malaysia should move from extrapolation to local measurement. A minimum programme would combine independent technical audits, preregistered voter experiments, longitudinal surveys and incident monitoring.

The work should be coordinated across election administrators, universities, civil society, media researchers and AI providers, with clear separation between independent oversight and political actors that stand to benefit from the findings.

• A multilingual benchmark: evaluate leading chatbots on matched prompts in Bahasa Melayu, English, Mandarin and Tamil, including state-level and local contests. Test accuracy, source quality, party coverage, candidate omission, voting-procedure answers and consistency when prompts are paraphrased.

• A controlled voter study: use preregistered designs to compare neutral, source-linked policy comparison with direct recommendation, persuasive dialogue and misinformation exposure. Measure immediate attitude change, confidence calibration, source recall and persistence over time. Do not treat stated vote intention as equivalent to an actual vote or election result.

• A source-poisoning exercise: create a controlled red-team environment with duplicated articles, fake interviews, altered screenshots and misleading social posts, then measure whether models detect provenance problems or repeat false material. Publish aggregate performance without amplifying harmful content.

• A privacy and profiling audit: assess whether information shared in ordinary conversations or inferred user attributes are used to steer political answers, personalise advocacy or infer turnout. Include data flows, third-party access, retention and deletion controls.

• A public reporting mechanism: establish rapid contacts between the Election Commission, platforms, independent fact-checkers and civil-society monitors. Log reports, classify severity, publish aggregate response times and preserve evidence for independent review.

This programme should be preregistered where experiments are involved, follow independent ethics review, use representative sampling where feasible, report confidence intervals and null findings, and publish reproducible methods. Researchers should distinguish four different endpoints: (1) factual performance of the system; (2) short-term changes in beliefs or stated preferences; (3) real-world exposure and sharing; and (4) measurable effects on turnout or election results. Evidence at one level does not establish effects at the next.

7. Conclusion

The evidence supports a careful but decisive policy position. AI systems can assist civic understanding, translate political information and lower barriers to participation. They can also produce persuasive arguments, give wrong election instructions, omit local parties, repeat unreliable source material and use personalisation in ways that undermine voter autonomy. The risk is not confined to spectacular deepfakes: it can arise from ordinary-looking answers delivered through a trusted interface.

At the same time, the evidence must not be inflated. Current studies do not prove that generative AI has changed Malaysia’s election outcomes. The meta-analytic literature does not show that LLMs are universally more persuasive than humans. Malaysian evidence is strong enough to demonstrate a real misinformation and multilingual-information challenge, but still insufficient to estimate the causal impact of AI chatbots on Malaysian vote choice. Good policy should be proportionate to what is known and designed to reveal what is not.

Livne’s principle—that AI should help people understand their choices without making those choices for them—is therefore a defensible foundation. The necessary improvement is to make it enforceable: authoritative and dated sourcing; a narrow prohibition on voting recommendations rather than political discussion; privacy protections against political profiling; source-poisoning and multi-step abuse tests; multilingual, matched-prompt audits; independent evaluation; transparent incident response; and due-process safeguards against overreach. In Malaysia and elsewhere, the objective is not to promise a perfectly neutral machine. It is to build an electoral information environment in which errors, omissions and manipulation can be detected, challenged and corrected while voters still retain control over the decision.

-----------------

References

1. Anthropic. (2026, October 14). Preparing for the 2026 US midterms. https://www.anthropic.com/news/2026-us-midterms

2. Autoriteit Persoonsgegevens. (2025, October 21). AP warns: Chatbots give distorted voting advice [Dutch-language report]. https://autoriteitpersoonsgegevens.nl/actueel/ap-waarschuwt-chatbots-geven-vertekend-stemadvies

3. Autoriteit Persoonsgegevens. (2026, March 12). AI chatbots ignore local parties when giving voting advice. https://autoriteitpersoonsgegevens.nl/en/current/ai-chatbots-ignore-local-parties-when-giving-voting-advice

4. Bai, H., Voelkel, J. G., Muldowney, S., Eichstaedt, J. C., & Willer, R. (2025). LLM-generated messages can persuade humans on policy issues. Nature Communications, 16, Article 6037. https://doi.org/10.1038/s41467-025-61345-5

5. Balakrishnan, V., Kaity, M., Abdul Rahim, H., & Ismail, N. (2021). Social media analytics using sentiment and content analyses on the 2018 Malaysia’s general election. Malaysian Journal of Computer Science, 34(2), 171–183. https://doi.org/10.22452/mjcs.vol34no2.3

6. Brennan Center for Justice. (2026, August 11). Does AI fight or fuel election disinformation? https://www.brennancenter.org/our-work/research-reports/does-ai-fight-or-fuel-election-disinformation

7. Conrad, C., Nissen, A., Masoumi, K., Ramchandani, M., Braga, R. F., & Newman, A. J. (2025). Do truthfulness notifications influence perceptions of AI-generated political images? A cognitive investigation with EEG. Computers in Human Behavior: Artificial Humans, 5, 100185. https://doi.org/10.1016/j.chbah.2025.100185

8. Daud, M. (2020). Fake news in the Malaysian 14th General Election: Shall the net be free forever? IIUM Law Journal, 28(S1), 303–324. https://doi.org/10.31436/iiumlj.v28iS1.587

9. European Parliament, Special Committee on the European Democracy Shield. (2025, April 29). Working document on protecting European democracy and our values: Findings and recommendations of the Special Committee on the European Democracy Shield (PE 771.957v01-00).

10. Gehringer, F., Nehring, C., & Łabuz, M. (2024, May). The influence of deep fakes on elections: Legitimate concern or mere alarmism? Konrad-Adenauer-Stiftung.

11. Hancock, J., & Moores, A. (2026, May 20). Electoral hallucinations: Safeguarding UK elections in the world of LLMs and AI chatbots. Demos. https://demos.co.uk/research/electoral-hallucinations-safeguarding-uk-elections-in-the-world-of-llms-and-ai-chatbots/

12. Helmus, T. C., & Chandra, B. (2024, April). Generative artificial intelligence: Threats to information integrity and potential policy responses. RAND Corporation. https://doi.org/10.7249/PEA3089-1

13. Hölbling, L., Maier, S., & Feuerriegel, S. (2025). A meta-analysis of the persuasive power of large language models. Scientific Reports, 15, Article 43818. https://doi.org/10.1038/s41598-025-30783-y

14. Jang, H., McGregor, S. C., & Neill, L. (2026). Inaccuracies, omissions, and bullshit: 2024 U.S. elections through the eyes of ChatGPT. Information, Communication & Society. Advance online publication. https://doi.org/10.1080/1369118X.2026.2685121

15. Juneja, P. (2024, April 29). Artificial intelligence for electoral management. International IDEA. https://www.idea.int/publications/catalogue/artificial-intelligence-electoral-management

16. Kruschinski, S., & Votta, F. (2026). Efficiency tool or manipulation machine? Assessing dominant claims about the quantity, quality, and persuasive power of visual generative AI in political campaigns. Zeitschrift für Politikwissenschaft, 36, 79–89. https://doi.org/10.1007/s41358-026-00457-w

17. Lee, Y.-H., & Yuan, C.-W. (2025). The authenticity paradox of political AI chatbots: Effects on candidate credibility, social presence, and voting intentions. Telematics and Informatics, 103, 102339. https://doi.org/10.1016/j.tele.2025.102339

18. Lin, H., Czarnek, G., Lewis, B., White, J. P., Berinsky, A. J., Costello, T., Pennycook, G., & Rand, D. G. (2025). Persuading voters using human–artificial intelligence dialogues. Nature, 648, 394–401. https://doi.org/10.1038/s41586-025-09771-9

19. Livne, D. (2026, October 14). Israel’s election overseer on how to stop AI distorting votes. The Economist; republished by Mint. https://www.livemint.com/opinion/israels-election-overseer-on-how-to-stop-ai-distorting-votes-11791989553163.html

20. Murphy, G., Ching, D., & Meehan, E. (2025). An average Joe, a laptop, and a dream: Assessing the potency of homemade political deepfakes. Applied Cognitive Psychology, 39(2), e70061. https://doi.org/10.1002/acp.70061

21. National Institute of Standards and Technology. (2024). Artificial intelligence risk management framework: Generative artificial intelligence profile (NIST AI 600-1). https://doi.org/10.6028/NIST.AI.600-1

22. Neyazi, T. A., Tan, K. E., & Kuru, O. (2026). The interplay of conservatism and religiosity on misinformation susceptibility: Evidence from the 2022 Malaysian General Election. International Journal of Public Opinion Research, 38(1), edaf061. https://doi.org/10.1093/ijpor/edaf061

23. OpenAI. (2026, May 27). Election information and safeguards in 2026. https://openai.com/index/election-safeguards-2026/

24. Park, S., & Nan, X. (2026). Generative AI and misinformation: A scoping review of the role of generative AI in the generation, detection, mitigation, and impact of misinformation. AI & Society, 41, 1501–1515. https://doi.org/10.1007/s00146-025-02620-3

25. Sabri, F. A. A. (2024). Narrative review on detonation of AI-generated deep fakes and misinformation: A threat to politicians in Malaysia. e-Journal of Media and Society, 7(2), 71–88. https://ir.uitm.edu.my/id/eprint/97612/

26. Suherman, A., Mulia, F., Hasrullah, & Maulana, H. F. (2026). Algorithmic emotions and digital deception: AI’s emerging role in political communication in Southeast Asia. Social Sciences & Humanities Open, 13, 102996. https://doi.org/10.1016/j.ssaho.2026.102996

27. Tapsell, R. (2023). Social media and Malaysia’s 2022 election: The growth and impact of video campaigning. Pacific Affairs, 96(2), 303–322. https://doi.org/10.5509/2023962303

28. Zailani, Z. N. H. M., Pirus, M. S. H. M., & Adnan, W. H. (2026). Literasi digital: Peranan aplikasi Facebook dan AI dalam membentuk sokongan politik dalam konteks Malaysia dan antarabangsa. Journal of Human Development and Communication, 15, 111–119. https://doi.org/10.58915/johdec.v15.2026.3446