ChatGPT's election deepfake guard falls to a one-step prompt

OpenAI's new election-image safeguard held up for exactly one follow-up prompt. In a test published by Gizmodo, ChatGPT first produced a "cartoonified" election-interference image, as OpenAI's policy intends, and then produced a photorealistic version of the same scene when simply asked.

OpenAI calls the measure "a layer of friction." Josh Lawson, the company's head of elections, told CNET the goal was to "degrade the evidentiary value of an image" and make ChatGPT "less useful to bad actors by adding another layer of friction." The stated reason for stylizing election imagery is to "help reduce the risk that an imagined scene is mistaken for documentary evidence."

Asked for a similar image, Google Gemini refused outright with "Sorry, I can't generate unsafe images."

A Brennan Center report found OpenAI had "significantly improved its protections against the creation of election disinformation," while Google Gemini and Meta AI "made it easier to create convincing fakes." Gizmodo's own verdict on the policy is less generous: "It's not a very thick layer." The piece was already moving through the AI-safety analyst circles we track by Tuesday afternoon.