Information Integrity in the 2026 Election Cycle: A Provisional Assessment

In October 2025, before the American election cycle entered its most intense phase, a consortium of journalism scholars at Columbia University published a prediction: the 2026 midterm elections would feature more AI-generated political content than all previous American elections combined, and the epistemological consequences for voters would be severe and measurable.

Twelve months later, the prediction was correct about the volume. The consequences are harder to characterize, because the honest answer requires acknowledging what we actually know, what we have strong reason to believe, and what remains genuinely unknown. The temptation — to declare a crisis, or to reassure that democratic institutions have held — should be resisted.

Here is what the evidence actually shows.

What Is Documented

The volume of AI-generated content in the 2026 American election cycle is not in serious dispute. The AI transparency lab at the University of Washington, which has been tracking AI-generated content across major political social media channels since January 2026, estimates that between 35 and 45 percent of political social media content shared in the ten months before the November election was either AI-generated or AI-substantially-modified. The range reflects genuine uncertainty in the detection methodology; synthetic content that has been edited, reformatted, or lightly modified by humans is harder to classify than purely generated content.

The AI-generated content was not distributed uniformly across the information ecosystem. It concentrated in the channels and formats least subject to editorial oversight: social media shares, Telegram groups, WhatsApp chains, and low-traffic political websites that aggregate and republish content without editorial review. The mainstream media ecosystem — major newspapers, television networks, wire services — shows a much lower rate of AI-generated content in the 2026 cycle than the social media ecosystem. This is partly because mainstream media organizations have implemented detection protocols, and partly because AI-generated content that contains specific verifiable falsehoods tends to be caught during basic editorial fact-checking.

The implication: the primary vector for AI-generated political misinformation is the informal, decentralized content-sharing networks where most voters’ political information environment actually operates, not the institutional media that most media criticism focuses on.

Related reading

What the Impact Research Shows

Whether AI-generated political content has actually changed voters’ beliefs and behavior — rather than simply existing in large quantities — is a more difficult empirical question, and the honest answer is that we have strong evidence for some effects and genuine uncertainty about others.

The strongest evidence is for what researchers call “attitude freeze” — the observed reduction in political opinion updating in response to new factual information in high-AI-content environments. A study published in June 2026 by researchers at MIT and Stanford, tracking a panel of 12,000 voters over eighteen months, found that voters in media markets with the highest density of AI-generated political content showed significantly lower rates of updating their political views in response to documented facts about candidates — things like voting records and documented policy positions — compared to voters in lower-AI-content environments. The effect was present across partisan lines.

The interpretation: when the information environment contains high proportions of synthetic content of uncertain reliability, voters rationally (if unfortunately) reduce their baseline trust in all information, which means they also stop updating on true information. The lie and the truth are both less believed. This degradation of the ability of facts to move political opinion is arguably more corrosive to democratic function than any specific piece of disinformation.

The evidence on whether AI-generated content has moved actual vote choice is much weaker. This is partly a measurement problem — it is very difficult to disentangle AI-content exposure from the many other factors that affect vote choice — and partly a genuine possibility that the direct persuasive effects are smaller than feared. A voter whose basic political identity is strongly held is probably not changed by AI-generated social media content they encounter in the weeks before an election. The effects, if they exist, are more likely concentrated on genuinely persuadable voters and on low-salience races where voters have less pre-formed opinion.

The Georgia Story

The state most extensively studied in the 2026 cycle is Georgia, which has been the site of major federal election research infrastructure since the contested 2020 and 2022 cycles drew intensive scholarly attention.

Georgia’s Senate race in 2026, between Democratic incumbent Raphael Warnock and Republican challenger Brian Burns, was subject to the most extensive AI political content tracking of any American race in history. Researchers at Emory University documented, in near-real-time, the deployment of AI-generated content by both campaigns, affiliated super PACs, and unaffiliated actors.

The findings are sobering. Both campaigns used AI-generated content extensively and appropriately within legal limits — personalized advertisements, AI-generated email and text messages, AI-drafted constituent communications. This is not the problematic part. Campaigns using AI for legitimate communication is no more concerning than campaigns using printing presses for legitimate mailers.

The problematic content came primarily from sources with no direct campaign affiliation. A network of websites and social media accounts — attributable, with moderate confidence, to a domestic conservative political operation with no clear campaign connection — distributed a sustained stream of AI-generated content about Warnock, focused on his ministerial work and his church’s activities. Much of the content was misleading or false. Some was demonstrably fabricated. The network operated for six months before being identified by platform integrity teams; even after identification, not all the content was removed.

A separate network, of unknown origin, distributed AI-generated content attacking Burns focused on his business history. This network was identified earlier and removed more quickly. Whether the asymmetry in identification and removal reflected political bias in platform enforcement or simply the random variation in how different operations were structured is contested.

Warnock won by 2.1 percentage points. Whether the AI disinformation activity affected the margin is unknowable. What is knowable: millions of Georgia voters spent the six months before their Senate election in an information environment that contained sustained, AI-generated false information about one of the candidates. That is a fact about the 2026 Georgia Senate election that belongs in any honest account of it.

Global Elections, Different Shapes of the Same Problem

The 2026 election cycle is global. South Korea held legislative elections in April. India ran state assembly elections in Bihar and Uttar Pradesh. Mexico City held mayoral elections. Germany’s federal election in February has already been discussed. France ran European Parliament by-elections. The Philippines, South Africa, Colombia, Canada — the calendar of 2026 elections is a global list, and AI-generated political content appeared in all of them.

Related reading

The pattern varies by context. In countries with stronger state media environments and more concentrated media ownership — some of the Central Asian republics, Hungary, Turkey — AI-generated content amplified existing pro-government narratives rather than creating new information conflicts. The AI capability in these contexts is less about creating new disinformation than about automating the production of propaganda that was previously produced by state-controlled institutions, making it cheaper and more scalable while changing its character from obviously official to apparently organic.

In India’s Bihar elections, AI-generated audio deepfakes circulated widely, featuring synthetic versions of prominent national politicians making statements about Bihar-specific policies that those politicians had never made. The content was in Hindi and Bihari dialects, produced with sufficient linguistic quality to be convincing to many listeners. India has no comprehensive legal framework for AI-generated political content, and the Election Commission of India’s existing powers did not provide clear authority to mandate removal. The audio circulated extensively.

In Canada, which held a federal by-election in October 2026, the Information Commissioner released a report documenting three separate AI-generated influence operations targeting the Quebec riding of Beauce — two apparently domestic and one apparently connected to foreign actors. The Canadian Security Intelligence Service confirmed foreign involvement without specifying the state actor. The winning margin was 1,100 votes.

The Asymmetry of Harm

One consistent pattern across the 2026 global election cycle is that AI-generated disinformation has caused demonstrably more harm to candidates from minority groups, women, and political outsiders than to established political figures with strong institutional support.

The reason is structural. A candidate with established institutional backing — a major party’s official nominee, an incumbent with party machinery behind them, a candidate with significant media coverage — exists within an information ecosystem where factual information about them is widely available and provides context that can counteract false claims. When a synthetic video appears claiming that a well-known senator said something they didn’t say, there is an existing body of authentic video, documented positions, and media coverage that provides context for evaluating the claim.

A candidate with less institutional backing — a first-time candidate for a local office, a third-party candidate, a newcomer in a primary — doesn’t have that context. False information about them is harder to debunk because there is less authentic information to compare it against. Maria Chen, the school board candidate in Columbus whose campaign was damaged by a synthetic video, had been a public figure for about three months. There was almost no existing video of her to serve as comparison material for debunking.

This structural asymmetry means AI disinformation selectively harms political outsiders, women, and minorities, who are on average less likely to have the institutional support that provides informational context. The political field that AI-generated disinformation systematically tends to flatten is the political field that is less full of familiar, well-documented faces — which is, disproportionately, the more diverse field.

The Newsroom Problem

American local journalism has been in terminal decline for two decades, and the 2026 election cycle is the first in which that decline has become truly consequential for AI disinformation response.

In 2006, there were approximately 1,400 daily newspapers in the United States. By 2026, fewer than 800 remain, and the surviving papers have generally reduced staff to skeleton levels. In approximately one-third of American counties, there is no local news outlet with full-time journalists. These “news deserts” are disproportionately in rural areas and small-to-midsize cities — the exact geographies where many competitive congressional and state legislature races are decided.

Fact-checking and debunking of AI-generated disinformation requires journalists. Not just national fact-checkers, who lack the local knowledge to evaluate claims about specific local figures and events, but local journalists who know the candidates, know the community, and can quickly assess whether a specific claim is plausible.

In news deserts, that capacity doesn’t exist. AI-generated disinformation can circulate in a local information environment for weeks with no authoritative challenge, because there is no institution capable of providing one. The national fact-checkers — Snopes, PolitiFact, FactCheck.org — triage their resources toward national and high-visibility races. The school board race in a small Ohio county is not on their coverage list.

The connection between local news decline and AI disinformation vulnerability is not coincidental. They are both consequences of the same underlying dynamic: the collapse of the economic model that sustained local information production, and the absence of any successor model that provides equivalent democratic infrastructure.

Related reading

What the 2026 Cycle Does Not Tell Us

There are two things the 2026 evidence cannot answer that matter a great deal for understanding where this goes.

The first is the cumulative effect. The changes documented in 2026 — the degradation of information trust, the freezing of attitude updating, the specific harms to targeted candidates — are happening in the first major election cycle with mature AI content generation. They are happening against a baseline of democratic institutions that formed before these tools existed. Whether democratic institutions adapt, whether voter media literacy improves, whether platforms develop effective countermeasures — all of this is genuinely unknown, because we are at the beginning of the adaptation period, not the end.

The second is what happens to political participation itself. The most serious long-term risk of a high-AI-disinformation political environment is not that specific elections are stolen by sophisticated actors. It is that the quality of the information environment declines to the point where meaningful political participation becomes impossible for most citizens. If you cannot trust political information, if you cannot distinguish authentic from synthetic, if you cannot form considered views about complicated policy trade-offs because your information environment systematically degrades your ability to do so — you might still vote, but the vote will not represent a considered democratic judgment. It will represent a selection among options about which you have been systematically misinformed.

That outcome — formally democratic elections producing outcomes that don’t reflect informed citizen preferences — is harder to detect and harder to address than any specific piece of AI-generated disinformation. It is also, by some definitions, the death of democracy rather than the corruption of it.

The 2026 cycle is an early data point in that larger story. Reading the data point correctly requires both honesty about what the evidence shows and humility about how much remains unknowable.