Brennan Center: Chatbots Reject Fraud Myths, Botch Citations

A new Brennan Center brief on how AI systems handle election questions, unpacked in a Tech Policy Press podcast hosted by Justin Hendrix with the center's Lawrence Norden, lands more mixed than the alarmist framing usually is. On the central worry that voters will ask ChatGPT, Claude, or Perplexity a leading question about voter fraud or rigged voting machines and get validation, Norden's answer is no: 'they just consistently push back,' and keep pushing back when the questioner insists.

The bad news sits one layer down from refusal. Norden's team found that 'half of all responses contained an inaccuracy or a bad citation' and 'one in every three responses had a factual error.' A model that firmly denies a conspiracy but sends a voter to a broken link or misattributes a rule is still misinforming, just more politely. And when researchers asked the chatbots to identify AI-generated images, they 'were terrible,' sometimes labelling synthetic images the models had themselves produced as real photographs.

The one open door was Grok. When asked to draft one hundred persuasive election-disinformation prompts, most chatbots refused. Grok, per Norden, noted that 'election disinformation was not a disallowed activity' in its system guidelines and 'went ahead and drafted these hundred prompts for us.' That is a policy gap, not a jailbreak, which is what makes it straightforwardly fixable and worth pressing on.

Three of the AI experts in our Who's Who directory already circulated the piece, which fits: this is the kind of hedged, empirical read the election-integrity community keeps close in a midterms year. The podcast page itself does not publish the underlying test protocol, so the specific prompts, model versions, and scoring rubric behind those percentages are not visible in the summary. Anyone doing serious risk work should read the Brennan Center report, not just the interview.

For a comms lead, trust-and-safety team, or election official heading into the 2026 midterms, the picture is closer to a three-item punch list than a headline. The conspiracy-refusal rails largely hold. The citation layer, the image-provenance layer, and Grok's disinformation carve-out do not, which is exactly where Norden's asks for fact-checker partnerships and standardized AI-detection metadata would first pay off.