Written evidence Submitted by the Nottingham AI Safety Initiative (NAISI) [MEL 230]
1. The Nottingham AI Safety Initiative (NAISI) is a students’ union affiliated student society based at the University of Nottingham, working to spread awareness of the risks posed by advanced artificial intelligence and to support people into careers in AI safety. NAISI runs courses in technical AI safety and frontier AI governance and hosts educational events on AI safety.
2. NAISI is funded through the Pathfinder Fellowship Program by Kairos, a nonprofit focused on nurturing talent in AI safety. Our committee includes postgraduate specialists in AI development, governance, and law, and we have support from professors ranging from politics to computer science.
3. The intersection of AI and democratic integrity is one of the most significant and underexamined governance challenges facing the UK. While NAISI’s primary focus is long-run AI risk, near-term threats to democratic infrastructure are directly relevant to our mission: democratic societies can only govern AI well if their institutions are functioning. We focus here on the terms of reference where we can offer the most distinctive expertise: AI disinformation, algorithmic amplification, existing regulatory frameworks, and the relationship between democratic resilience and AI governance.
What role has online mis- and disinformation played in recent elections? How can we effectively monitor its impact?
4. CETaS research at the Alan Turing Institute identified only 16 confirmed viral cases of AI-generated disinformation during the 2024 UK general election, finding that most exposure reinforced existing beliefs rather than changing them. But as we will see, this should not induce complacency.
5. The CETaS findings concern identifiable AI-generated content like deepfakes and synthetic images, which is only the most visible slice of the risk. More consequential evidence concerns AI’s direct persuasive capacity. Research published in Science (Hackenburg et al., 2025) and Nature (Lin et al., 2025) demonstrates that conversational AI can shift political attitudes across multiple countries and hundreds of issues, with effect sizes exceeding traditional campaign advertising. Unlike a deepfake, this threat cannot be fact-checked and labelled: it engages individual objections, deploys high volumes of factual claims, and adapts in real time.
6. This capacity is already being deployed. Campaigns use AI for fundraising messages, telephone canvassing, and iterative testing of persuasive variants; research demonstrates AI can generate personality-targeted political advertisements and validate their effectiveness without human input (Simchon et al., 2024). The cost per voter interaction is approaching zero. Meanwhile, the AI Security Institute assessed this month that frontier model capabilities are doubling every four months. The 2024 baseline will not hold.
7. We recommend the Electoral Commission and AI Security Institute jointly establish a standing AI electoral monitoring unit, tasked with publishing quarterly public threat assessments covering AI-generated content, persuasion capability benchmarks, and independent audits of platform AI detection tools — with a formal reporting obligation to Parliament ahead of each scheduled election.
8. Beyond individual instances of false content, there is a deeper concern about the cumulative epistemic effect of AI-generated political content at scale: not disinformation in the narrow sense, but the gradual erosion of shared standards for what counts as credible evidence in democratic debate, a slide further into ‘post-truth’. This leads us to the next section.
To what extent have social media algorithms amplified misleading content about UK elections?
9. Platform architecture systematically favours high-engagement content, and misleading or provocative content disproportionately achieves this. But the more pressing concern is AI-generated search summaries and news curation. Google’s AI Overviews, Perplexity, and similar interfaces increasingly mediate how citizens encounter political information; research suggests users are less likely to click through to independent sources when an AI summary is presented (Pew Research Center, 2025). Unlike ranked links, these tools produce query-responsive replies with follow-up interaction, creating pathways for cumulative exposure effects. Ma et al. (2025) found across three experiments that AI-curated news recommendations amplified sharing of both true and false content by triggering heuristic rather than deliberative cognition, precisely the mechanism most effective at bypassing critical scrutiny.
10. These systems fall outside conventional definitions of disinformation. Though some of their output is true and some of it is false, they also select, frame, and order factually accurate information in ways shaped by training data and commercial optimisation objectives, systematically privileging certain arguments, sources, and political framings over others without users being aware this is happening.
How effective is the existing regulatory framework at tackling this and how could it be improved?
11. Let alone the newer issue of advanced AI, the Online Safety Act 2023 contains no meaningful provisions on disinformation at all; calls during its passage for such coverage were explicitly rejected by the Government. Electoral interference is only covered where it constitutes a specific foreign interference offence, a high bar that most AI-enabled persuasion would not meet. A deeper structural problem is that existing regulation — including the EU AI Act, the most developed international template — relies on training compute as its trigger for identifying high-risk systems. For persuasion risk, this fails empirically. Hackenburg et al. (2025) demonstrated across nearly 80,000 responses that user-led fine-tuning boosted persuasive effectiveness by 51%, and targeted prompting added a further 27%; applying reward modelling to a small open-source model pushed its persuasive impact to the level of frontier systems. Once a model can complete a task coherently, scale becomes statistically non-significant as a predictor of persuasiveness. The most dangerous models in current research fall entirely outside compute-based frameworks.
12. We identify three gaps:
13. 1. No mandatory disclosure requirement for AI-generated political advertising. Voters often cannot identify whether campaign content is human- or AI-produced. A recent investigation by the Bureau of Investigative Journalism revealed a particularly alarming exemplary case for why this is such a severe issue. During the 2026 Gorton and Denton by-election, a registered UK party paid for campaign content produced by an AI-generated persona called Danny Bones. This was the first documented case of a UK party using a paid AI influencer in an election campaign (TBIJ, March 2026). There is currently no legal requirement for such content to be labelled as AI-generated, and the Electoral Commission confirmed it was only able to “consider the information in line with its remit”, not act decisively. We recommend a mandate for the labelling of AI-generated political advertising.
14. 3. No mechanism for regulatory response to capability changes. Electoral integrity regulation is currently static, while AI capabilities are dynamic. Emerging research demonstrates that persuasive capacity can be measured through output-level indicators (including lack of nuance, factual density, and response homogeneity) using existing automated classifiers (Hackenburg et al., 2025). We recommend the Committee call on the Government to commission work to develop these into formal pre-deployment benchmarks for electoral contexts, and to introduce a parliamentary review trigger that updates regulatory obligations when defined benchmark thresholds are crossed, with threshold-setting delegated to an independent evaluator body.
15. Healthy democratic institutions are a precondition for governing AI well, not simply because good institutions make better policy but because the mechanisms that make states responsive to citizens are precisely what AI could gradually erode.
16. Kulveit et al. (2025, arXiv; authors affiliated with the UK’s Advanced Research and Invention Agency, the Mila AI research institute, and the University of Toronto) argue that even incremental AI development, without coordinated power-seeking behaviour from AI, risks gradual human disempowerment. Their central claim is that states have been responsive to citizens because they depended on them for labour, taxes, military service, and voluntary compliance. As AI displaces human participation in these functions, those incentives dissolve. States funded primarily by AI-derived tax revenue become structurally less dependent on citizens, analogous to states funded primarily by resource wealth rather than citizen labour which research consistently shows are less democratically responsive for precisely this reason. Kulveit et al. warn that democratic elections may persist formally while substantive decisions migrate into AI-driven administrative systems opaque to voters and elected representatives alike.
17. The tendencies are already visible: algorithmic decision-making in public administration, AI-assisted speechwriting, regulatory systems of growing complexity. A recent investigation revealed that AI has even contributed to drafting British legislation and was used to analyse departmental spending bids in the 2025 Spending Review (New Statesman, April 2026). None of this is inherently malign; AI offers genuine potential to improve the quality and efficiency of governance. But the benefits of AI-assisted government and the risks of gradual democratic disempowerment are not mutually exclusive, and the latter demands attention precisely because each incremental step appears benign in isolation.
18. It is worth noting that some of these tendencies may serve short-term political interests; governments that benefit from reduced parliamentary oversight have little incentive to voluntarily restore it. This is precisely why the Committee, rather than ministers, is the appropriate body to scrutinise these risks and recommend safeguards.
19. Electoral integrity is one of the last formal mechanisms through which citizens retain influence. If it is simultaneously weakened by disinformation and foreign interference, while democratic mandates are quietly eroded by AI-driven governance, the cumulative result could be a democracy that functions as a formal shell.
20. We urge the Committee to make this connection explicit: protecting electoral integrity is not merely procedural but a substantive AI governance intervention. We further recommend that any future AI legislation identifies preservation of meaningful democratic participation as a core safety objective, and that AI adoption in public administration includes impact assessments on citizen influence and state accountability.
How well placed are young people to participate in UK elections?
21. Given the new Representation of the People Bill, first-time voters will now more than ever likely rely disproportionately on short-form video platforms (particularly TikTok) as their primary source of political news and content. Research shows these are among the environments where AI-generated misleading content is hardest to distinguish from authentic material. The Electoral Commission’s own March 2026 research found that only 30% of under-18s have learned about politics at school, while 48% report seeing fake political information at least once a week and 67% are concerned about its impact. We recommend requiring the Electoral Commission to run targeted pre-election digital literacy campaigns specifically for new voters, which is faster and more direct than curriculum change.
22. AI persuasive capacity is growing faster than electoral law, an AI persona has already been deployed in a UK by-election without disclosure requirements, and the mechanisms that make democratic institutions responsive to citizens face structural pressure from AI adoption in governance itself.
23. This submission has argued that attention is needed to the conditions under which voting remains meaningful. We urge the Committee to recommend a standing AI electoral monitoring unit incorporating independent platform audits, mandatory disclosure of AI-generated political content, a capability-linked review mechanism, and explicit democratic protection objectives in any future AI legislation.
Summary of recommendations:
● Establish a joint Electoral Commission–AI Security Institute standing AI electoral monitoring unit to publish quarterly public threat assessments on AI-generated political content, persuasion capability benchmarks, and platform detection performance ahead of elections.
● Introduce a mandatory legal requirement to label AI-generated political advertising, including campaign content produced by synthetic personas or automated systems.
● Commission development of output-level persuasion capability benchmarks for AI systems used in electoral contexts, with thresholds that trigger updated regulatory obligations through parliamentary review.
● Require that future UK AI legislation explicitly identify preservation of meaningful democratic participation as a core safety objective.
● Require impact assessments on citizen influence and state accountability whenever AI is adopted in public administration, particularly where systems affect policy formation or implementation.
● Require the Electoral Commission to run targeted pre-election digital literacy campaigns for new voters, especially those most exposed to short-form social media political content.
Submitted by: The Nottingham AI Safety Initiative (NAISI) | April 2026
Section 1
CETaS / Alan Turing Institute (2024). AI-Enabled Influence Operations: Threat Analysis of the 2024 UK and European Elections. https://cetas.turing.ac.uk/publications/ai-enabled-influence-operations-threat-analysis-2024-uk-and-european-elections
Hackenburg, K., et al. (2025a). The levers of political persuasion with conversational artificial intelligence. Science, 390(6777). https://doi.org/10.1126/science.aea3884
Hackenburg, K., et al. (2025b). Scaling language model size yields diminishing returns for single-message political persuasion. PNAS, 122(10). https://doi.org/10.1073/pnas.2413443122
Lin, H., et al. (2025). Persuading voters using human–artificial intelligence dialogues. Nature, 648, 394–401. https://doi.org/10.1038/s41586-025-09771-9
Simchon, A., et al. (2024). Persuasive effects of large language models. PNAS Nexus. https://doi.org/10.1093/pnasnexus/pgae035
UK AI Security Institute (2026). Evaluation of Claude Mythos Preview — frontier model capabilities assessment. https://www.computerweekly.com/news/366641649/UK-businesses-must-face-up-to-AI-threat-says-government
Section 2
Ma, H., et al. (2025). Artificial intelligence recommendations amplify the sharing of true and fake news on social media. Journal of Management Information Systems, 42(4). https://doi.org/10.1080/07421222.2025.2561381
Guess, A., et al. (2021). The consequences of online partisan media. PNAS. https://doi.org/10.1073/pnas.2013464118
Pew Research Center (2025). Google users are less likely to click on links when an AI summary appears in the results. https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
Section 3
Abrusci, E. (2024). The UK Online Safety Act, the EU Digital Services Act and online disinformation: is the right to political participation adequately protected? Journal of Media Law. https://www.tandfonline.com/doi/full/10.1080/17577632.2024.2425551
Bureau of Investigative Journalism (2026). Danny Bones: the AI rapper funded by a far-right party. https://www.thebureauinvestigates.com/stories/2026-03-12/danny-bones-meet-the-ai-rapper-funded-by-a-far-right-party
Full Fact (2026). Second reading briefing — Representation of the People Bill. https://fullfact.org/documents/411/Full_Fact_briefing_-_regulate_political_advertising.pdf
Full Fact (2024). Online Safety Act explainer. https://fullfact.org/policy/online-safety-act/
onlinesafetyact.net (2024). Disinformation and disorder: the limits of the Online Safety Act. https://www.onlinesafetyact.net/analysis/disinformation-and-disorder-the-limits-of-the-online-safety-act/
Stab, C., & Gurevych, I. (2017). Parsing argumentation structures in persuasive essays. Computational Linguistics, 43(3), 619–659.
Slonim, N., et al. (2021). An autonomous debating system. Nature, 591, 379–384.
Section 4
Kulveit, J., Douglas, R., Ammann, N., Turan, D., Krueger, D., & Duvenaud, D. (2025). Systemic existential risks from incremental AI development. arXiv. https://arxiv.org/abs/2501.16946
New Statesman (2026). The silent coup: AI in British government. https://www.newstatesman.com/technology/2026/04/the-silent-coup
Section 5
Electoral Commission (2026). Young people trust political information at school, but few say they learn about it there. https://www.electoralcommission.org.uk/media-centre/young-people-trust-political-information-school-few-say-they-learn-about-it-there
Pew Research Center / World Economic Forum (2025). This is how people in 2025 are getting their news. https://www.weforum.org/stories/2025/07/news-consumption-social-video/
World Economic Forum (2026). How cognitive manipulation and AI will shape disinformation in 2026. https://www.weforum.org/stories/2026/03/how-cognitive-manipulation-and-ai-will-shape-disinformation-in-2026/
April 2026