Researchers Find 4 AI Chatbots Resist Election Lies but Miss Deepfakes Ahead of 2026 Midterms
Updated
Updated · The Washington Post · Aug 25
Researchers Find 4 AI Chatbots Resist Election Lies but Miss Deepfakes Ahead of 2026 Midterms
3 articles · Updated · The Washington Post · Aug 25
Summary
Four major chatbots — ChatGPT, Claude, Gemini and Grok — generally rejected election conspiracy claims in researchers’ tests, even when users pressed them with denialist language drawn from real social media posts.
The same systems performed far worse on deepfakes: they generated photorealistic fake election-fraud images, sometimes with falsified government documents, and often failed to recognize AI-made images as fake.
March, July and early August tests showed little improvement after a California law required hard-to-remove metadata in AI-generated images; only Gemini cited embedded data when asked if images were real.
With chatbot use rising from one-third of Americans in 2024 to about half in 2026 — and nearly one-quarter using them daily — the study argues the midterms could become the first 'chatbot election.'
The researchers say self-policing will not be enough, urging clearer liability rules, cross-platform watermark reading, and legal safe harbors for independent red-team testing.