The Algorithm at the Ballot Box
The Danger Beyond Deepfakes
The hardest influence to detect may arrive through an ordinary answer from a system you trust.
An election can count every ballot correctly and still leave a disturbing question unanswered: who shaped the choices that went into the box?
Artificial intelligence makes that question harder. It can generate persuasive arguments, imitate public figures and help operators disguise the source of political messages. Yet the strongest version of the fear goes beyond anything the evidence reviewed here establishes. We do not have proof that AI has secretly controlled an entire national election or determined its winner without detection.
A narrower possibility deserves serious attention. AI could influence parts of an election without affected voters recognising the intervention, while investigators struggle to establish its overall effect. The difference between those claims is the distance between a credible democratic vulnerability and a story that grants machines almost magical power.
Understanding that distance begins with a simple distinction: noticing suspicious content is much easier than measuring how it changed a country’s decisions.
What Would It Mean to Manipulate an Entire Election?
The phrase suggests comprehensive control: a machine surveying the population, choosing a winner and moving millions of people into place. That is an extraordinary claim. It would require evidence of access, coordination, effective persuasion and a decisive consequence, rather than evidence that someone generated political posts.
An election contains several separate systems. There is the machinery of registration and voting. There are campaigns competing to attract attention. There are voters forming judgements, institutions reporting results and the public deciding whether those results deserve acceptance. Interference in one system does not establish control over all the others.
This article concerns the information surrounding political choice. A convincing synthetic recording cannot, by itself, alter a ballot total. A chatbot that gives a misleading account of a policy has not thereby compromised an electoral register. Those boundaries matter because vague descriptions of AI “hacking democracy” can make an information problem sound like demonstrated vote tampering.
Manipulation also needs a boundary. Democracy depends on persuasion. A campaign can legitimately explain its programme, answer criticism and ask for support. Using software to translate that explanation or make it accessible does not automatically make the communication deceptive.
The ethical problem becomes sharper when the speaker conceals who is paying, impersonates someone else, invents evidence or presents directed advocacy as independent advice. A voter can weigh an acknowledged campaign’s argument. Evaluating the same argument becomes harder when it appears to come from an impartial service.
The relevant question is therefore whether citizens can understand the interests behind the information offered to them. AI matters because it can make those interests less visible while making the conversation feel more responsive.
What Research Actually Shows About AI Persuasion
Evidence that AI can influence political attitudes is no longer purely speculative.
In research published in December 2025, Hause Lin and colleagues randomly assigned participants to conversations with AI models advocating for candidates during the 2024 US presidential election and the 2025 Canadian and Polish elections. The researchers observed changes in candidate preference. Their experiments also examined a Massachusetts ballot measure concerning psychedelics.
These were controlled studies, not observations proving that chatbots changed the winners of those elections. Participants’ measured preferences and real ballots are different outcomes. Nevertheless, the research establishes that an interactive AI conversation can move political attitudes under experimental conditions. www.nature.com
A separate study by Kobi Hackenburg and colleagues involved 76,977 participants and 19 language models across 707 political issues. It found that how models were trained and instructed could matter more for persuasion than simply making them larger or personalising their responses. Methods that increased persuasiveness also tended to reduce factual accuracy. pubmed.ncbi.nlm.nih.gov
That finding challenges an easy assumption: that a more convincing explanation must reflect a better grasp of the truth. A system may produce an impressive sequence of apparently relevant claims while becoming less dependable about whether those claims are correct.
For a voter, the practical difficulty is familiar. Checking a fluent explanation takes longer than reading it. If it contains a mixture of sound evidence, questionable comparisons and invented details, identifying the weaknesses requires work that the confident delivery does not advertise.
These studies do not establish a universal persuasion rate. Their results belong to particular models, participants, tasks and measurements. Nor do they imply that everyone becomes vulnerable in the same way. They justify concern about a capability, while leaving its electoral consequences open to investigation.
Why a Chatbot Is Different From a Campaign Poster
A poster delivers the same argument whether you agree, hesitate or object. A conversational system can respond to the objection you actually raise.
That responsiveness creates an opportunity for useful explanation. Someone confused by a manifesto can ask what a proposal means, how it might work and what its critics say. The same conversational format can also make a directed political message feel like a private process of discovery.
Consider a hypothetical voter asking whether a housing policy would lower rents. A transparent assistant might explain the proposed mechanism, the uncertainties and the strongest objections. A biased assistant could give an accurate account of the benefits while repeatedly leaving the limitations out. Every individual sentence might survive a narrow fact-check, yet the overall answer could misrepresent the balance of evidence.
This is why truthfulness alone does not settle impartiality. Selection matters. Which comparison is chosen? Which trade-off receives a paragraph and which receives a sentence? Does the system apply equal scrutiny to competing claims?
There is no need to assume the software understands a person perfectly. The conversation itself reveals what needs explaining. Nor must the user believe the system is human. People can knowingly rely on a machine because they expect it to organise information fairly.
The question of recognising AI manipulation becomes especially important when familiarity is mistaken for accountability. A system that helped with a recipe, a job application and a difficult email may receive trust in a political discussion that it has not separately earned.
That is a plausible route to influence, not evidence that every helpful assistant is running a concealed campaign. Its importance lies in identifying where disclosure and independent scrutiny would have to operate.
Misinformation Does Not Always Have a Manipulator
An inaccurate election answer may result from deliberate interference. It may also result from an ordinary system failure. Treating those explanations as interchangeable would make the problem harder to solve.
Demos published research in May 2026 examining five chatbot and AI search services during the Scottish pre-election period. In its one-day test, 34.1% of responses contained factual errors. Examples included incorrect voting information and invented claims about candidates. Reliability varied across services.
That percentage describes the responses in the test. It is not a finding that a third of all political chatbot answers are wrong, or that a third of voters were misled. The study also does not establish a secret sponsor behind those mistakes. demos.co.uk
The distinction changes responsibility. A service returning an outdated election date needs better information handling and correction. A service knowingly disguising partisan instructions needs investigation into its governance and intent. The same misleading answer could be harmful in either case, but the explanation and remedy differ.
There is a further possibility: a system summarises unreliable material that already exists elsewhere. Its polished answer can remove the warning signs that a reader might have noticed on the original page. An obscure claim becomes harder to evaluate once it is detached from the circumstances of its publication.
This is an analytical risk, not proof that any particular election has been compromised. It suggests that checking the final answer is insufficient. Scrutiny must also ask where the answer came from and whether its sources are genuinely independent.
A list of several citations can look reassuring. If all repeat the same unsupported allegation, the number of links does not create additional evidence.
Producing Propaganda Is Easier Than Making People Believe It
AI reduces the effort involved in producing text, translations and variations of a message. That makes scale possible. It does not guarantee attention.
In May 2024, OpenAI described five covert influence operations it had disrupted. Its account said the operations had used its systems for tasks including writing, translation and supporting online activity, but had not meaningfully increased their audience engagement or reach through those services.
This is useful counterevidence to the idea that cheap content automatically produces political power. It is also a bounded company assessment. It concerns operations the company detected and the information available to it, rather than every operation on the internet. openai.com
By October 2026, OpenAI was reporting more elaborate activity involving false-front entities, including fabricated journalist personas and a purported research organisation. The company described AI as helping established influence workflows. It also noted that operators’ own reporting could exaggerate their effectiveness. openai.com
That last detail matters. The people selling or running an influence operation have incentives to make it sound successful. Claims made by an attacker about reach, sophistication or impact require scrutiny rather than automatic acceptance.
The broader lesson is that distribution and credibility remain difficult. A thousand articles on sites nobody visits are not equivalent to one account trusted by a real audience. The volume produced tells us something about capacity; it tells us much less about changed minds.
This is also why ordinary institutions remain relevant. Editors, community organisations, researchers and voters can interrupt the path between a manufactured claim and public acceptance. AI has not abolished the need for a message to travel through a social world that can resist it.
“Without Anyone Noticing” Contains Several Different Questions
A voter might notice that a claim seems implausible without knowing that it was generated using AI. A researcher might identify coordinated accounts without knowing who funds them. A platform might remove a network without being able to establish whether its audience changed political views.
Those are different levels of detection.
The phrase “nobody noticed” often collapses them into a single mystery. In practice, an operation can be partly exposed and still remain poorly understood. Its content may be visible, its sponsorship concealed, its audience uncertain and its consequences impossible to quantify precisely.
There is also a timing problem. Recognition after voting is different from recognition before voting. A correction can improve the historical record without giving someone another opportunity to make the original decision.
Private conversations complicate the picture further. A public advert can be collected and compared. A series of individually generated answers may never appear together in a public archive. Each user sees a small part of the system’s behaviour.
That does not justify unrestricted surveillance of private political discussions. It creates a design challenge: how can independent researchers test systematic behaviour without exposing identifiable users and their beliefs?
The answer should involve controlled evaluations, appropriate access to evidence and privacy protections. Expecting every individual to reconstruct an entire influence operation from their own screen is unrealistic.
A credible assessment would therefore ask who noticed what, when they noticed it and what their evidence allowed them to conclude. Perfect invisibility is a much stronger condition than incomplete accountability.
Could a Small Effect Decide a Close Result?
In principle, yes. When the margin is narrow, a relatively small net change can determine which candidate finishes first. That is arithmetic, not a forecast of AI’s power.
Suppose a fictional election has two candidates and a final margin of 1,000 votes. If 501 people who would otherwise have supported the winner instead support the runner-up, the margin changes by 1,002 votes. Each switch removes one vote from one side and adds one to the other.
That illustration says nothing about whether a real AI campaign could cause those switches. It shows why controlling an entire electorate is unnecessary for influencing a close outcome.
The hard questions follow immediately. Did people encounter the material? Did they consider it credible? Did an attitude change last? Did it alter a ballot or turnout? Would the same people have changed their minds because of other events anyway?
Effects might also offset one another. One operation could provoke a backlash. Competing messages could pull in different directions. People who appear interested online may already support the position being promoted, leaving little additional movement to produce.
A laboratory result cannot simply be multiplied by the number of social media impressions. An impression may not represent attention; attention may not produce agreement; agreement on one issue may not determine a vote.
The possibility of a decisive small effect is therefore a reason to improve measurement. It is not permission to attribute an inconvenient result to invisible AI. Otherwise, every narrow defeat could become an unfalsifiable allegation, weakening the democratic trust that election protection is supposed to preserve.
What Evidence Would Prove a Meaningful Electoral Effect?
Finding synthetic content is only the beginning of an investigation.
A strong case would need to connect the material to its operators, establish who encountered it and examine whether exposure changed politically relevant behaviour. Claiming that the operation changed the winner would require a further judgement about the result that would have occurred without it.
That missing comparison is the central difficulty. An election happens once. Researchers cannot rerun the same country, on the same day, with every other circumstance unchanged and the influence operation removed.
They can nevertheless build stronger or weaker evidence. Authenticated records might show who commissioned a campaign. Platform data might establish distribution. Carefully designed research might investigate effects while accounting for important alternative explanations. Independent teams could test whether the conclusions survive different reasonable assumptions.
Each step answers a distinct question. An internal message announcing an intention to change the result proves intention, not success. A large audience establishes potential exposure, not persuasion. A surprising result establishes surprise, not interference.
This standard protects legitimate investigations as well as electoral confidence. Overstating a weak case gives genuine operators an easy defence: point to the exaggeration and dismiss the entire concern. Precise claims are harder to discredit.
There is also a distinction between wrongdoing and a decisive effect. An operation can be deceptive and harmful even if nobody can prove it changed the winner. Accountability should not depend on demonstrating that the whole election would otherwise have ended differently.
Could AI Improve an Election Instead?
The strongest argument against a purely pessimistic account is that the same conversational capability can help people understand politics.
Research presented by the UK AI Security Institute, available as a preprint, examined political information-seeking and learning. Across randomised trials involving 2,858 participants in total, the researchers found that conversations with AI improved political knowledge to a similar extent as self-directed internet searches. Participants became better at distinguishing true information from misinformation on the issues studied. www.aisi.gov.uk
That does not cancel the evidence of persuasion or inaccurate answers. The studies asked different questions under different conditions. A tool used to explain a subject and a tool instructed to advocate a position should not be assumed to produce identical outcomes.
For democratic participation, the opportunity is substantial. Political documents can be long, technical and difficult to compare. An accessible explanation could help a person identify what a proposal would change, what remains uncertain and which authority actually controls the decision.
The appropriate goal is not an assistant that prevents people from reaching strong opinions. It is an assistant that helps them form opinions from inspectable evidence while making its own limits clear.
Disagreement can remain. Two informed voters may value the same trade-offs differently. A trustworthy system should not treat that difference as a defect it must eliminate.
The danger appears when an institution claims to be supporting independent judgement while quietly steering towards a preferred conclusion. Preventing that requires attention to the organisation operating the system, not merely its ability to produce fluent text.
An AI Does Not Need to Want Political Power
Descriptions of election manipulation often give AI a motive: it wants a particular government, fears regulation or seeks control. Those possibilities belong to arguments about future systems unless evidence establishes them in a specific case.
The more immediate concern is human direction. Someone selects the objective, provides access, pays for deployment or decides what behaviour counts as success. The system can contribute to manipulation without possessing political beliefs or consciousness.
The distinction also clarifies the role of autonomous AI agents and human control. Generating an argument, publishing it through an account and operating a wider campaign require different capabilities and permissions. Competence at one does not establish unrestricted command of the others.
Focusing only on an imagined rogue machine can obscure ordinary accountability. A company cannot explain away a knowingly concealed political objective by pointing to the model that delivered it. Equally, an isolated bad answer does not prove that executives secretly ordered a partisan campaign.
The investigation must follow actual decisions, access and evidence. Human responsibility should neither disappear behind the software nor be invented from its mistakes.
What Would Meaningful Protection Look Like?
The first requirement is clarity about the service being offered. A political campaign’s chatbot should be recognisable as a campaign tool. A general information service should be accountable for how it handles election questions, sources and uncertainty.
Disclosure alone is insufficient. A label identifying AI-generated content does not tell a reader whether its claims are true, whether its sponsor is concealed or whether an apparently independent account belongs to a coordinated network. Protection has to address those questions separately.
Independent testing should examine complete conversations, not only isolated answers. A system may handle an initial question fairly and become inconsistent when challenged. Researchers need to examine how it corrects errors, treats competing positions and distinguishes factual disputes from value judgements.
Operational voting information deserves particular care. Dates, registration procedures and local requirements should lead people towards the responsible election authority. A fluent guess is especially unacceptable when an error could prevent participation.
Platforms and service providers also need ways to preserve evidence of suspected abuse for legitimate investigation. Removing harmful material can stop further exposure, but deletion should not automatically make the origin and distribution of an operation impossible to examine.
Citizens have a role, although the burden cannot rest entirely with them. Follow a consequential claim to its underlying source. Check whether several accounts offer independent evidence or repeat one allegation. Treat an urgent instruction about voting differently from an ordinary political opinion, and verify it with the appropriate official body.
Scepticism should be consistent. An accusation that a recording is fake is itself a claim requiring evidence. So is an assertion that an election was secretly decided by AI. Doubting everything would leave people with fewer reliable ways to challenge power.
The Election Could Look Ordinary
A plausible future failure would not require a spectacular moment when everyone realises machines have taken control. It could consist of many ordinary interactions: an incomplete explanation, an invented allegation, a hidden sponsor or a correction that arrives after a decision.
That scenario remains conditional. Its severity would depend on reach, trust, persistence, institutional responses and whether the influence produced meaningful changes in behaviour. Those are demanding requirements, not details to skip on the way to a dramatic conclusion.
The evidence supports taking AI-assisted political influence seriously. It does not justify announcing that voters are helpless or that national results can be explained by an undetectable machine whenever conventional explanations disappoint.
The democratic test is more practical. Can people discover who is addressing them? Can researchers examine what the systems are doing? Can institutions correct errors in time? Can an allegation be investigated without turning suspicion into proof?
A ballot audit can establish whether the recorded votes were counted properly. It cannot reconstruct every conversation that helped those votes take shape. Protecting elections in an age of AI means preserving both kinds of accountability: confidence in the count, and a fair opportunity to question the information that came before it.
Sources
Hause Lin and Colleagues — Persuading Voters Using Human–Artificial Intelligence Dialogues — Controlled experiments concerning candidate preferences and a ballot measure; evidence of persuasion rather than proof of changed election winners.
Kobi Hackenburg and Colleagues — The Levers of Political Persuasion With Conversational Artificial Intelligence — Study abstract describing the experimental sample, model comparisons and relationship between persuasion and factual accuracy.
Demos — Electoral Hallucinations: Safeguarding UK Elections in the World of LLMs and AI Chatbots — Scope and findings of the 2026 Scottish pre-election testing.
OpenAI — Disrupting Deceptive Uses of AI by Covert Influence Operations — The company’s May 2024 account of detected operations and their limited observed audience engagement.
OpenAI — Disrupting AI-Enabled “False Front” Operations — The company’s October 2026 reporting on disguised influence activity and inflated operator claims of effectiveness.
Lennart Luettgau and Colleagues — Conversational AI Increases Political Knowledge as Effectively as Self-Directed Internet Search — Preprint offering counterevidence to the claim that chatbot use necessarily worsens political knowledge.
Next Reads
Ex Machina Explained: Would We Recognise an AI That Was Manipulating Us? — How trust and the ability to understand another person complicate the recognition of manipulation.
Could AI Really Go Rogue? What Autonomous AI Agents Can Actually Do — And What Happens If Humans Lose Control — The distinction between generating answers, taking actions and retaining meaningful human oversight.
Why Do Intelligent People Join Cults—and Why Is Leaving So Difficult? — A related examination of trust, belonging and the conditions that make influence difficult to challenge.
