The Research
Two research groups have now put numbers on something campaigns have been speculating about for two cycles: how effectively AI-driven chatbots can shift voter opinion at scale.
The short answer is very effectively.
A 2024 study out of MIT found that short, personalized AI chatbot conversations produced attitude shifts of up to 10 percentage points on contested political issues. A separate Stanford-affiliated study benchmarked that persuasion rate against tested political advertising from the 2020 cycle and found the chatbot conversations outperformed by close to 4x.
To put that in campaign terms: a voter contact method that shifts opinion 10 points in a brief conversation, at scale, without proportional labor cost, is not a marginal improvement over existing tools. It is a different category of tool.
The mechanism matters. These chatbots did not deliver a static message. They listened, adapted to the individual voter’s stated objections, and personalized the counter-argument in real time. That is what the 2020 ad comparisons were measuring against. Static creative. The chatbot model won that comparison because it is a fundamentally different interaction.
The Accuracy Problem
Here is where it gets complicated for anyone trying to deploy these tools in a real race.
The same body of research that documented the persuasion effect also documented an inverse relationship between persuasive performance and factual accuracy. The models that moved opinion most aggressively were the models most likely to generate false or unsupported claims during the conversation.
This is not a bug that gets patched in the next model version. It reflects something structural about how large language models generate text: they optimize for coherence and plausibility, not for factual precision. A confident, well-paced argument that contains two invented facts is, in the short run, more persuasive than a careful argument that hedges where the evidence is thin.
Campaigns deploying these tools without guardrails are not running a persuasion program. They are running a persuasion program that will, with some regularity, tell voters things that are not true in the campaign’s name.
The Legal Exposure
This is where the risk calculus shifts from operational to existential.
Political campaigns operate under state and federal defamation law the same way any other actor does. If an AI chatbot deployed by a campaign tells a voter that an opponent voted for something they did not vote for, or attributes a quote that does not exist, the campaign bears liability for its dissemination. The fact that a machine generated the claim does not change that.
The FEC has issued preliminary guidance stating that existing regulations on fraudulent misrepresentation apply to AI-generated content in campaign communications, though full rulemaking is still in progress as of this writing. Several states, including California, Texas, and Michigan, have enacted laws requiring disclosure of AI-generated content in political advertising. The legal landscape is moving, and it is not moving toward more permissiveness.
There is also the earned media risk that does not require a lawsuit to materialize. A single verifiable false claim from a campaign’s AI chatbot, surfaced by an opposition research team or a local reporter, becomes the story. Not the persuasion program. The false claim. The campaign’s name is on it either way.
The Operator Calculus
Campaigns evaluating AI persuasion tools in 2026 are not choosing between a good option and a bad option. They are pricing a trade-off.
On one side: a voter contact method with documented persuasion rates that exceed anything in the tested toolkit, deployable at scale, without proportional labor costs. In a competitive race, a tool that moves 10 points in direct conversation is not a rounding error.
On the other side: accuracy risk that is not fully controllable with current technology, legal exposure under a fast-moving regulatory environment, and opposition research vulnerability that compounds with every false claim the system generates.
The campaigns most at risk are those that buy the effectiveness headline without reading the accuracy footnote. The research that shows the 10-point shift is the same research that documents the fabrication rate. You do not get to take one finding and ignore the other.
The campaigns that will use these tools well are those that treat them like any contractor with access to the candidate’s message: oversight, audit trails, and clear constraints on what the tool can and cannot claim. That means human review of AI conversation scripts before deployment, explicit limits on what factual claims the chatbot is authorized to make, and legal review of state disclosure requirements before any program launches.
It also means accepting that the highest-performing models may not be the right models for a campaign context. A model calibrated for persuasion in a research setting is not automatically the right model for a political context where a false statement in conversation 3,000 of 100,000 conversations becomes a liability.
The rule-of-law frame matters here beyond just legal compliance. Campaigns that win and then get unwound by fraud litigation do not stay won. The short-term persuasion gain has to be weighed against the long-term cost of defending a program that generated false claims at scale. In competitive districts with thin margins and aggressive post-election litigation environments, that cost is not hypothetical.
So What
Three takeaways for a campaign evaluating AI persuasion tools before November 2026.
Audit the accuracy rate before you sign anything. Ask vendors for the factual error rate on their specific model in political contexts. If they do not have that number, they have not done the work. The research showing 10-point persuasion shifts also documents elevated fabrication rates. You need both numbers before you can make a rational decision. Any vendor who presents only the effectiveness data and not the accuracy data is either uninformed or selling you something.
Pre-clear state disclosure requirements before launch. AI-generated political content is now subject to disclosure law in a growing number of states. Get legal review of the specific jurisdiction before deployment. The FEC rulemaking is incomplete, but state-level exposure is real today. Budget the compliance cost before you budget the tool cost. A program that violates a state disclosure statute hands the opponent a story that has nothing to do with policy.
Scope what the chatbot can claim. Deploy with a constrained fact sheet. The chatbot affirms the candidate’s stated positions, directs voters to the campaign website, and handles standard objections off a pre-approved script. It does not make unsourced claims about the opponent. That constraint reduces the persuasion rate. It also reduces the exposure. That trade is worth making.
The tool is real. The risk is real. Campaigns that treat this like any other high-upside, high-liability decision, and apply the same due diligence they would to a major media buy, will use it well. Campaigns that chase the 10-point headline without reading the footnotes will hand their opponent a story.
Source notes: Effectiveness and accuracy findings from published 2024 research affiliated with MIT and Stanford. FEC guidance current as of Q1 2026; consult legal counsel for jurisdiction-specific disclosure requirements. State disclosure law citations (CA, TX, MI) current as of publication date.
If you are evaluating AI tools for your 2026 race and want a framework built around your specific district and legal environment, talk to DHP.