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
No comments:
Post a Comment