Are Politicians Using AI? What the Evidence Actually Shows

TL;DR

• Yes, politicians and political campaigns are already using AI. The most common uses are practical rather than futuristic: drafting messages, summarizing research, translating content, generating social posts, testing ads, analyzing voter data, and automating routine campaign work.
• There is no reliable global figure for what percentage of politicians use AI. Most research instead measures reported incidents, surveys political consultants, or studies specific elections. One 2025 U.S. survey found that 86% of political consultants had used AI for work, while direct use by elected officials appears much more uneven.
• AI-generated political content is already widespread internationally. A 2025 study covering the 2024 global election cycle documented 215 reported generative-AI incidents across 50 countries, with incidents found in 80% of those countries. India and the United States had the largest numbers in the dataset.
• The biggest concerns are deception, impersonation, microtargeting, privacy, and scale. Deepfakes and voice clones receive the most attention, but experts are also concerned about campaigns using AI to cheaply generate enormous volumes of personalized political messaging.
• Evidence that AI alone has decided major elections remains limited. Researchers and election experts generally see real risks, but there is much less evidence that AI-generated content has independently transformed election outcomes than headlines sometimes suggest.
• Governments are responding with disclosure and transparency rules. Brazil has election-specific AI and deepfake rules, the EU now requires transparency for certain synthetic content and political advertising, and several U.S. states have adopted their own AI-election laws.
• AI is affecting politics beyond campaigning. It is increasingly relevant to legislative research, policy analysis, government administration, political advertising rules, privacy regulation, election law, and international AI governance.

Executive summary

Yes. Politicians, political parties, campaign consultants, advocacy groups, and people trying to influence elections are already using artificial intelligence. But the reality is less cinematic than the phrase “AI election” suggests. Most documented legitimate use is mundane: drafting emails, brainstorming messages, summarizing research, translating material, analyzing voter data, creating social posts, producing graphics, testing advertisements, and reducing the cost of routine campaign work. The spectacular uses—voice clones, synthetic candidate videos, fake endorsements, AI avatars, and deepfakes—are real, but they are only one part of a much larger shift in political communications.

The strongest global measurement comes from the International Panel on the Information Environment's 2025 study of generative AI in the 2024 elections. Researchers built a database of 215 reported incidents across 50 countries with competitive national elections. They found AI-related election incidents in 80% of those countries. Ninety percent involved content creation, while political candidates and parties were identified as the source in 25% of incidents; 20% came from foreign actors and 46% could not be reliably attributed. India and the United States had the largest number of recorded cases, with 30 each.

That finding does not mean 80% of politicians use AI. Measuring actual politician-level adoption turns out to be surprisingly difficult. Campaigns rarely publish their prompts, AI invoices, internal workflow logs, or vendor contracts. Much of the use is deliberately invisible because AI may only produce a first draft that a human rewrites. Academic research therefore measures related things—reported incidents, interviews with campaign professionals, voter reactions, synthetic-media capabilities, and detected influence operations—rather than producing a reliable global percentage of politicians who use ChatGPT.

One of the clearest direct adoption measurements comes instead from an industry survey. In 2025, the American Association of Political Consultants Foundation surveyed 200 U.S. political consultants. Eighty-six percent said they had used AI for work at some point, 59% used it at least weekly and 34% at least daily. Respondents said AI was concentrated in internal workflows, research, writing and brainstorming rather than sophisticated voter targeting or fully automated campaigning.

Politicians themselves appear more uneven. A September 2026 Axios survey of more than two dozen members of the U.S. Congress found that many senior lawmakers rarely or never personally use AI, even when their staff sometimes do. By contrast, a 2026 survey of 40 Ohio legislators found that a majority reported using AI in legislative work. In other words, adoption varies substantially by institution, generation, staff culture and task.

The danger is similarly uneven. Evidence so far does not support the strongest prediction—that a single flawless deepfake or autonomous AI propaganda machine has routinely decided major democratic elections. For example, OpenAI reported disrupting more than 20 election-related influence attempts in 2024 without finding that any achieved meaningful audience traction, while a TIME review of the 2024 election super-cycle similarly concluded that the feared “AI election” largely failed to materialize in its most apocalyptic form.

But that should not be mistaken for safety. AI makes persuasion, translation, impersonation, experimentation and content production dramatically cheaper. That matters because politics is a scale business: reducing the cost of producing one personalized message is mildly interesting; reducing the cost of producing ten million of them is potentially transformative. The most serious long-term risk may therefore be less “one perfect deepfake changes history” and more a permanent flood of inexpensive, personalized, difficult-to-attribute political communication.

Also read: Free AI Detectors With No Word Limit: What Actually Works in 2026?

What research can actually tell us about politicians' AI use

There is an important evidence problem at the center of this subject. We have plenty of research about AI and elections, but very little research that can answer the seemingly simple question, “What percentage of elected politicians use generative AI?”

That is partly because AI usage is difficult to observe. A campaign employee may paste a polling memo into an LLM, ask it for ten message ideas, rewrite one of them, send it to a designer and never disclose that AI was involved. No deepfake exists to count. No watermark survives. An outside researcher cannot infer the workflow reliably from the final advertisement.

So it helps to separate four kinds of evidence: direct surveys of political professionals; interviews with campaigns; databases of observable AI incidents; and experiments measuring how voters respond to AI political communication. They answer different questions and should not be mixed together.

The most defensible conclusion is therefore not “X percent of politicians use AI.” It is that AI has moved from experimental novelty into normal political-professional workflows, while adoption by elected officials themselves remains much more uneven.

The AAPC data are especially revealing because the glamorous applications are not the most common ones. Consultants primarily describe AI as a productivity system. The same survey identifies drafting presentations, brainstorming strategy, writing emails and fundraising appeals, summarizing large volumes of text, automating routine work, and creating or refining talking points as practical uses. Sophisticated voter segmentation and personalized voter journeys were seen more as growth opportunities than as already-dominant applications.

Other academic research helps explain why visible usage is hard to police. Radivojevic and colleagues tested more than 1,000 people on their ability to distinguish LLM-generated from human social-media writing and found that human identification is unreliable. More recent work by Dawkins, Fraser and Kiritchenko, using more than half a million generated social-media posts, found that fine-tuned systems can produce text that is particularly hard to distinguish from human writing. These experiments do not show that politicians generated a particular message; they show why trying to identify political AI merely by “reading carefully” is becoming increasingly unreliable.

A useful rule for interpreting future headlines is therefore: counting detectable deepfakes substantially undercounts political AI use, while counting every AI-related incident substantially overstates direct politician use.

Also Read: Are Ebooks on Amazon and Other Ebook Stores AI-Written?

How campaigns and politicians are using AI

The phrase “politicians using AI” covers several very different activities. Some are little different from using spell-check or Photoshop. Others can impersonate a candidate or optimize persuasive messages for individual voters. Treating them as one category makes sensible regulation almost impossible.

Campaign operations and research

This is probably the least controversial and most widespread category. Staff use LLMs to summarize news coverage, policy papers, polling transcripts and opposition research; turn meeting notes into action items; brainstorm questions; clean data; write code; generate first drafts; and create alternative message formulations. The AAPC survey strongly suggests that these behind-the-scenes uses currently outweigh futuristic autonomous campaigning.

This also helps explain an apparent contradiction. many U.S. members of Congress told Axios in September 2026 that they personally seldom use AI, while political professionals surrounding campaigns use it routinely. A senator does not have to open ChatGPT for AI to affect the research memo, donor segmentation, social copy or draft talking points arriving on the senator's desk.

Speeches, talking points and written communication

Generative AI is well suited to producing a rough draft from a few policy positions, shortening a speech, rewriting an answer for a different audience, or turning a briefing note into a fundraising email. The LSE analysis by Florian Foos identifies scripts and campaign communication as established assistant-style uses, while the AAPC survey explicitly records the creation and refinement of ad scripts, talking points and speeches among consultants' applications.

The important distinction is between assistance and authorship. A politician may use AI to suggest ten versions of a sentence and personally choose or rewrite one. At the other extreme, a campaign might publish an entire AI-produced statement without checking it. Both can truthfully be called “using AI,” but they present very different accountability risks.

Social media and content generation

This is where adoption becomes much more visible. AI can produce dozens of captions, memes, images and short videos in the time previously needed to design one. The IPIE global election database found content creation in about 90% of documented generative-AI election incidents. That includes synthetic audio, video, images, text and social posts.

Some uses are openly playful or satirical; others sit in an uncomfortable middle ground. Donald Trump, for example, shared apparently AI-generated images in 2024 portraying Taylor Swift supporters and a false-looking Swift endorsement. In 2026, California Governor Gavin Newsom's online operation was also using AI-generated political memes. Such material may be intended as joke, persuasion or provocation rather than documentary evidence, but repeated use can make synthetic political imagery feel ordinary.

Cheap video and localization

Generative AI is particularly attractive to smaller campaigns because it reduces production costs. An underfunded candidate can create graphics, voiceovers and localized videos without hiring a full creative team. That is one reason the Associated Press found substantial experimentation in lower-profile U.S. campaigns. For small campaigns, AI can narrow the professional-production gap with richer opponents.

India provides the clearest example of scale. During the 2024 general election, WIRED documented politicians and political vendors using AI avatars, voice clones and multilingual personalized messages. Vendors described producing enormous volumes of personalized calls and videos. The IPIE database separately records an example in which local technology companies produced 25 million personalized AI calls to voters in two southern Indian states.

This is politically important because translation used to be expensive. A candidate wishing to address many linguistic communities might need translators, actors, studios and repeated recording sessions. Voice cloning and AI translation allow one politician's message—and sometimes a synthetic version of the politician—to speak many languages quickly. That can broaden democratic access, but it also creates a difficult question: is the voter hearing the politician, an approved simulation of the politician, or something made by an unknown third party?

Targeted advertising and A/B testing

Political campaigns were microtargeting long before ChatGPT. Voter files, consumer data, Facebook targeting and statistical models are not new. AI changes the economics and speed of the process rather than inventing it.

The Center for Media Engagement interviews found political professionals already using generative AI to accelerate data analysis and A/B testing. Instead of writing two fundraising subject lines, a consultant can generate 100, sort them by demographic group, test a subset and quickly produce variations of the winner.

Former Trump campaign manager Brad Parscale's technology company Campaign Nucleus has promoted precisely this broader vision. An Associated Press investigation into Parscale's AI-oriented campaign technology described systems aimed at personalized content, donor and voter targeting, influencer amplification and rapid media production. The important caveat is that the capabilities advertised by a vendor should not automatically be treated as evidence that every associated campaign deployed every capability in practice.

Also Read: How LLMs Can Degrade Academic Papers?

Voter microtargeting and hyper-personalization

This is the application experts disagree about most. In principle, AI can combine voter-file attributes, survey answers, geography, browsing-derived interests and previous responses to generate individualized persuasion. A climate-conscious suburban voter could receive one version of a candidate's economic argument while a small-business owner receives another.

That sounds extremely powerful. Empirically, the picture is less clear. Foos notes that evidence for political microtargeting's effectiveness was mixed even before generative AI. The Center for Media Engagement likewise found practitioners divided: some regarded AI hyper-personalization as a major advance; others warned that campaigns could produce so many tiny audience segments that results cease to be statistically meaningful.

This distinction matters. AI clearly makes personalization cheaper. It has not yet been established that unlimited personalization makes political persuasion proportionately more effective.

Chatbots and synthetic candidates

Campaign chatbots can answer policy questions, explain voting procedures, collect voter concerns or imitate a candidate's style. An unusually revealing case occurred during the 2024 U.S. Democratic primary, when a super PAC supporting Dean Phillips funded a chatbot called Dean.Bot designed to converse as an AI version of the candidate. OpenAI subsequently suspended the developer responsible under the company's political-use policies at the time.

That example also shows why attribution matters. It was a super PAC project, not simply “the candidate using a chatbot.” Political ecosystems contain campaigns, independent expenditure groups, parties, vendors, influencers and activists, and news coverage often collapses those actors into one.

Deepfakes and voice clones

This is the highest-profile category because it attacks a basic assumption of political communication: that seeing or hearing someone is evidence they actually spoke.

A famous 2024 example was the AI-cloned Joe Biden robocall used ahead of the New Hampshire primary. Separate testing reported by the Associated Press found that researchers could coax six popular voice-cloning services into producing convincing fake election audio in roughly 80% of 240 attempted tests involving political figures. That experiment measured technical vulnerability, not the prevalence of actual election attacks, but it illustrates how low the barrier to impersonation had become.

Deepfakes do not have to come from national campaigns to matter. former Shreveport mayor Adrian Perkins described being targeted in 2022 by a local satirical advertisement that digitally placed his face onto another person's body. Local contests can actually be attractive targets because candidates often lack national press teams, forensic experts and rapid-response networks.

There is also a more subtle danger known as the liar's dividend: once everyone knows that convincing fakes exist, authentic recordings become easier to dismiss as AI. The political payoff is no longer limited to fabricating something that never happened. A politician caught by genuine evidence may simply claim that it was generated.

That is why provenance—where a file came from and how it was edited—is ultimately more useful than asking citizens to become human deepfake detectors.

Where AI disclosure is becoming part of election law

Governments are converging on a few basic ideas: synthetic political media should sometimes be labeled; deceptive candidate impersonation deserves special treatment; automated interactions should not pretend to be human; and sensitive personal data should not be freely available for political profiling.

But there is no uniform global rule. Some countries mandate labels. Some prohibit deepfakes during sensitive election periods. Some rely on general privacy, defamation or advertising law. In the United States, much of the action is at state rather than federal level.

Brazil is especially worth watching because it is attempting to regulate not merely obvious deepfakes but the broader political AI environment. Ahead of Brazil's 2026 election, Reuters reported that the Superior Electoral Court was enforcing mandatory AI labels and a deepfake ban while also restricting certain AI systems from ranking or recommending candidates. Nineteen AI-related election cases had reportedly been filed in the first eleven days of campaigning.

Independent monitoring suggests why regulators are worried. According to a 2026 survey of political AI material by Brazil's Observatório IA nas Eleições, reported by Folha de S.Paulo, researchers identified 413 AI-related political publications between January 1 and August 15. President Lula appeared in 112; 81% of the monitored material was classified as deepfake content and 63% reportedly lacked a label. Politicians or parties accounted for about 35% of the posts. This is a monitoring dataset rather than a complete census, but it shows the enforcement problem plainly: disclosure rules are only as useful as compliance.

The European approach is broader. The EU AI Act is not an election law; it creates horizontal rules for AI systems across the economy. Its Article 50 transparency regime matters politically because people deploying deepfakes generally must disclose that the content was artificially generated or manipulated. The rule includes tailored treatment for artistic, satirical and similar material rather than simply banning synthetic expression.

The EU Regulation on the Transparency and Targeting of Political Advertising, most of whose provisions began applying in October 2025, attacks another part of the problem. Political ads must be identifiable and accompanied by information about who sponsored them, while political targeting using personal data is much more tightly constrained. The rule matters even when the advertisement itself is human-written because AI targeting systems can be politically consequential without generating a single fake image.

Europe also benefits from the Digital Services Act, which requires ad transparency on large platforms and restricts targeting based on sensitive personal information. Taken together, the rules move regulation away from the narrow question “Is this a deepfake?” toward the larger questions “Who paid for this? Why did I receive it? What personal information was used? Was AI involved?”

The United States remains much more fragmented. However, one important national intervention came from the Federal Communications Commission. After the New Hampshire Biden-clone incident, the FCC clarified that AI-generated voices fall within the Telephone Consumer Protection Act's rules governing artificial or prerecorded voices. That does not outlaw every synthesized political voice, but it means AI does not magically escape existing robocall consent and identification requirements.

There is another model besides disclosure: prohibition. Some jurisdictions have chosen to ban particular forms of deceptive candidate deepfake close to an election rather than assuming a small label will neutralize a powerful fake. This is an important policy debate because labels work well for benign, authorized synthetic content, whereas deliberate disinformation operators may simply ignore them.

What political and technology experts think: useful tool, democratic danger, or both?

Experts are not divided into a simple “AI good” and “AI bad” split. The more useful disagreement is over where the danger lies and how much genuinely new power AI gives political actors.

The last point deserves emphasis. Generating propaganda and persuading people are not the same problem. AI has unquestionably made the first easier. Evidence that it has made the second easy is far weaker.

A beautifully generated fake video that nobody sees has no electoral effect. A poorly produced human conspiracy theory promoted by a major influencer may reach millions. Distribution networks, partisan identity, trusted messengers, recommendation algorithms and pre-existing political grievances can matter more than the production technology.

The Alan Turing Institute's Centre for Emerging Technology and Security analysis of the 2024 UK and European elections makes this broader point useful: generative AI should be understood as one component of influence operations rather than a magical replacement for organization, access and audience building.

So, do experts consider political AI dangerous? Broadly, yes—but mostly under specific conditions. The strongest concern is not somebody using an LLM to shorten a speech. It is AI combined with deception, sensitive data, impersonation, anonymous distribution, highly personalized targeting or automated scale.

At the same time, responsible uses can plausibly improve democratic participation. Real-time translation can let a candidate answer people previously excluded by language barriers. Small campaigns can afford professional-looking communication. Constituents can query long policy documents conversationally. Election administrators can translate voter information faster. AI can help journalists process enormous advertising libraries.

The difficult policy goal is therefore to preserve those benefits without treating consent-free cloning, covert manipulation and mass personalized propaganda as ordinary innovation.

AI is already reshaping policy-making and regulation

The most striking thing about AI's political impact may be that its effect on the policy agenda is currently deeper than its direct adoption by many elected politicians. The September 2026 Axios interviews found numerous senior members of Congress barely using AI themselves even as legislatures around the world spend enormous amounts of political attention deciding how it should be governed.

Regulation is developing in layers.

Election-specific rules deal with deepfakes, candidate impersonation, political advertisements, robocalls and synthetic-media disclosure. Brazil's election rules and U.S. state laws are examples.

Economy-wide AI law governs AI regardless of whether the user happens to be a campaign. The EU AI Act is the clearest example. Its risk-based system addresses prohibited practices, high-risk systems, general-purpose AI and transparency. For politics, Article 50's deepfake and synthetic-content rules are particularly relevant.

Privacy and advertising regulation tackles the input side of AI persuasion. Europe's GDPR, Digital Services Act and political-advertising regulation constrain how campaigns and intermediaries can exploit personal information and require greater advertising transparency.

Telecommunications rules can be extended to new AI behavior rather than rewritten from scratch. The FCC's treatment of AI-generated voices under robocall law is a useful example of a regulator saying, in effect, “the technology is new, but the underlying regulated behavior is not.”

Soft law and agency guidance fill gaps where legislation moves more slowly. The U.S. National Institute of Standards and Technology's AI Risk Management Framework and its Generative AI Profile provide practical risk-management guidance rather than election law. Their influence is broader: governments and companies can use such frameworks to structure testing, documentation and human oversight.

The United Kingdom initially pursued a deliberately sector-based approach through its AI Regulation: A Pro-Innovation Approach white paper. That model asked existing regulators to apply common AI principles rather than immediately creating a single EU-style statutory framework. Subsequent British AI policy has continued evolving, illustrating a wider global debate between horizontal legislation and regulator-by-regulator governance.

At international level, the Council of Europe's Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law goes even further conceptually. Opened for signature on September 5, 2024, it is the first international legally binding treaty specifically addressing AI in this field. It calls for principles including human dignity, nondiscrimination, privacy, transparency, oversight, accountability and safe innovation. It also requires mechanisms for risk and impact assessment and, in relevant circumstances, notice that a person is interacting with AI rather than a human.

In other words, “AI policy” has already grown from a technology-industry question into an issue of election law, consumer law, privacy law, human rights, telecommunications, platform regulation, administrative procedure and international treaties.

There is another policy implication that receives less attention: regulation can itself create data. Mandatory labels, political-ad libraries, sponsor disclosures and provenance metadata can eventually make academic measurement much better. Right now researchers often rely on newspaper articles precisely because campaigns and platforms disclose too little. Good transparency law could therefore improve both accountability and political science.

The ethical, legal and practical problems are bigger than deepfakes

Authenticity and consent.

The easiest ethical case is unauthorized impersonation. A candidate should not be able to make an opponent appear to confess to a crime or endorse a policy through a cloned voice. But harder questions appear when the candidate authorizes the clone. India's 2024 election showed that synthetic candidate voices can be used with campaign approval to reach people at scale. The message may be politically legitimate while still leaving the listener with the false impression that the candidate personally recorded it.

A sensible transparency principle is therefore broader than “fake versus real”: did the depicted or imitated person actually perform the words or actions the audience is seeing?

Hallucinations and factual error.

Political work is full of dates, legal claims, voting records, budget figures and quotes. LLMs can produce fluent mistakes. A hallucination copied into a speech can become a news story; copied into official constituent guidance, it can become a governance problem. This is one reason some lawmakers interviewed by Axios cited unreliability as a reason to avoid generative AI.

Privacy

Hyper-personalized campaigning requires data. That makes political AI inseparable from debates over voter files, consumer data, inferred religion, ethnicity, health conditions, sexuality, economic vulnerability and political beliefs. The AI model may be new, but the danger of political actors knowing too much about individual citizens is old.

Generative AI adds an important twist: the system can turn that data into individualized language automatically. Traditional microtargeting chooses which prewritten advertisement a person receives. Generative targeting can theoretically write a different argument for every person.

Manipulation without lying

A political AI message does not have to contain a factual falsehood to be ethically troubling. Imagine a system inferring that one voter is anxious about crime, another is struggling with debt and a third recently lost a job, then framing the same candidate differently to exploit each vulnerability. Every statement might technically be true. The ethical concern lies in invisible psychological tailoring.

This is one reason transparency about targeting can matter as much as transparency about content.

Discrimination

AI-supported voter models can reproduce demographic biases in their training data or campaign databases. A turnout model might tell a campaign that certain neighborhoods are “low probability” and therefore not worth contacting. Generative systems can also produce stereotyped messages aimed at racial, religious or linguistic groups. Such harms can be difficult to see because every audience receives different material.

Defamation, publicity rights and intellectual property

Deepfakes can implicate defamation law when they falsely portray someone engaging in damaging conduct. Voice and likeness cloning may trigger publicity or personality rights in jurisdictions that recognize them. AI campaign material may also raise copyright questions involving training inputs, cloned performance styles, music, photographs and generated output. Which legal theory applies varies considerably by country.

Campaign-finance accountability

AI complicates the boundary between campaigns and supposedly independent organizations. The Dean.Bot example illustrates the issue: a super PAC can build something that behaves like an interactive representation of a candidate without the candidate's official campaign necessarily being the publisher. Regulators need to know who paid, who instructed the system and who is legally responsible for its statements.

Cybersecurity and confidentiality

Campaigns hold donor records, strategic research, opposition files, draft policy positions and voter information. Staff who paste sensitive documents into consumer AI systems can create confidentiality or data-governance risks. An AI-use policy for a serious political organization therefore needs to cover data handling, not merely deepfake ethics.

The spam problem

Even accurate AI content can degrade politics if it creates virtually unlimited low-cost communication. A campaign that once sent three fundraising emails a week can generate thirty variations a day. Interest groups can create thousands of pseudo-grassroots comments. Social networks can fill with synthetic supporters. The danger becomes volume rather than falsity.

The accountability gap

Politicians can blame an employee, employees can blame the vendor, vendors can blame the model and model providers can say the user violated policy. Democratic accountability requires that a human or legal organization remain responsible for a political communication regardless of how much of it was automated.

The liar's dividend

Better deepfakes can paradoxically damage trust even when nobody is fooled by a particular fake. Once synthetic evidence is common, real evidence becomes contestable. This is why digital provenance standards are valuable. The Coalition for Content Provenance and Authenticity, for example, develops technical standards intended to preserve information about the origin and editing history of digital media. Provenance is not foolproof, but it tackles a more durable problem than teaching everyone to spot strange fingers or unnatural blinking.

Detection is an arms race

A detector trained on yesterday's generator may perform poorly on tomorrow's model, and ordinary compression, screenshots, cropping and re-recording can further weaken forensic signals. The academic studies on human detection reach a similar conclusion from another direction: humans should not be treated as reliable AI detectors either.

Inequality cuts both ways

AI can democratize campaign production by giving a local candidate inexpensive design, translation and research assistance. But the richest campaigns can buy proprietary voter data, custom models, experimentation infrastructure and high-end consultants. Cheap generic AI may therefore narrow one resource gap while advanced data-driven AI widens another.

Those tensions explain why a blanket “ban AI in politics” is both unrealistic and poorly targeted. It would treat automated translation like a fraudulent voice clone and spreadsheet analysis like covert psychological profiling. The more defensible approach is risk-based: ordinary productivity assistance receives light rules; automated persuasion, sensitive-data profiling and realistic impersonation receive much stronger ones.

What voters, journalists and policymakers should do

For voters, the most important change is to stop treating realistic audio or video as self-authenticating evidence. A surprising recording of a candidate should be checked against the candidate's verified channels, reputable reporting and the original source before it is shared. Look for provenance information or a synthetic-media label, but do not assume the absence of a label proves authenticity. Conversely, do not assume a politician's claim that something is a “deepfake” proves it is fake.

Voters should also ask a second question that is less obvious: Why am I seeing this particular political message? A perfectly truthful AI-generated advertisement can still be highly targeted. Political-ad transparency regimes such as the EU's political-advertising rules point toward a useful norm: people should be able to identify the sponsor and understand the basis on which they were targeted.

For journalists, provenance should become a standard reporting beat. When a suspicious recording appears, preserve the highest-quality original file possible rather than analyzing a compressed social-media repost. Establish who first published it, whether the depicted person consented, what metadata survives, whether an independent forensic expert has examined it and whether the campaign itself has disclosed AI assistance.

Journalists should also resist turning every synthetic artifact into a dramatic “AI threatens democracy” headline. Doing so can unintentionally advertise low-reach propaganda and contribute to the liar's dividend. Coverage should distinguish four separate questions: Was AI actually used? Who used it? How many people encountered the material? Is there evidence it changed anyone's behavior?

The fourth question is frequently missing. A deepfake with 400 views is technologically interesting but politically different from one seen by 40 million people.

Newsrooms should broaden their investigations beyond synthetic media. Vendor contracts, campaign technology spending, voter-data brokers, AI-generated fundraising, automated canvassing, ad-variant production and chatbot deployments are likely to tell us more about long-run political AI adoption than collections of strange AI memes.

For policymakers, the priority should be accountable transparency rather than a technologically impossible promise to eliminate synthetic media. A workable framework would require clearly visible disclosure when realistic synthetic media materially represents a real political figure saying or doing something they did not actually say or do; require automated political agents to identify themselves; preserve sponsor and funding information; and establish fast remedies for unauthorized impersonation during election periods.

Rules should also cover the inputs to political AI. Sensitive personal information should receive strong protection against covert political profiling. Citizens should be able to understand why they were targeted. Platforms should maintain useful political-ad archives that preserve different versions of personalized ads instead of showing researchers only the final campaign aggregate.

Regulators should avoid assuming that a disclosure sticker solves everything. A malicious foreign influence operator is unlikely to obediently write “AI generated” across a fake candidate confession. Disclosure works best for lawful campaigns and legitimate synthetic media. Deliberately deceptive activity still requires enforcement, platform response, provenance investigation and ordinary criminal or civil law where applicable.

Campaigns themselves should be required—or strongly encouraged—to maintain internal records of consequential AI use: which systems were used, what data entered them, whether candidate likenesses were synthesized, who approved public output and what human review occurred. Such records would make post-election investigations far easier than attempting to reconstruct AI use from finished social posts.

Governments should also create safe access for independent researchers. The IPIE study repeatedly runs into a fundamental problem: researchers are trying to understand global political AI through public news stories because platforms, vendors and campaigns expose limited underlying data. Regulation that produces structured transparency datasets would dramatically improve our ability to distinguish genuine systemic problems from a few highly publicized anecdotes.

Finally, lawmakers should preserve benign uses. Translation, accessibility, document summarization, routine research, coding assistance and clearly disclosed creative work should not be governed as though they were consent-free candidate impersonation. The evidence from Jungherr, Rauchfleisch and Wuttke suggests the public itself recognizes these distinctions: deceptive AI is viewed much more severely than ordinary operational uses.

The central policy principle can be stated simply: the closer AI gets to secretly pretending to be a person, secretly exploiting intimate information about a person, or secretly making political decisions for a person, the stronger the transparency, consent and accountability requirements should become.

So, are politicians using AI? Absolutely. But the most important story is not that a few candidates can now make synthetic videos. It is that AI is quietly becoming part of the political production system—the research assistant, copywriter, translator, analyst, designer, targeting engine and, increasingly, conversational interface between campaigns and voters.

The available evidence suggests that this transformation is already broad but still shallow in many places: lots of experimentation, lots of productivity assistance, relatively fewer sophisticated autonomous systems, and limited proof that generative AI by itself has swung major elections. That is actually the moment when democratic institutions have the best chance to act. Waiting until personalized AI agents, indistinguishable voice clones and automated persuasion are fully embedded in election infrastructure would make sensible rules much harder to design.

The right question for the next election is therefore no longer simply, “Was this made by AI?” It is: Who used the AI, what did it do, what data did it use, did the people involved consent, was the audience told, and which human being remains accountable?