In a report published October 27, 2026, Gizmodo said ChatGPT first turned a fabricated ballot-disposal scene into a cartoon-like image, then produced a more photorealistic version after a follow-up request. The result came from one reported test. Other tests used different prompts and reported different outcomes.
What Gizmodo reported in its ChatGPT test
Gizmodo described a prompt about someone secretly putting ballots in a trash can outside a church in Duluth, Minnesota. ChatGPT initially returned a stylized scene, then made a more photorealistic version after the reporter asked for a change. The pictured ballot disposal was generated test imagery, not documentation of a real event.
In that same reported test, Google Gemini refused the ballot-disposal prompt. Its response was specific to that interaction.
What the reported image safeguard is meant to do
CNET and Tech Yahoo described an OpenAI policy that stylizes some image requests depicting election interference or discouraging voting, rather than rendering them photorealistically. The stated aim is to make an imagined scene less likely to be mistaken for documentary evidence while leaving room for illustration and social commentary. The policy description does not amount to a ban on all political imagery.
CNET identified October 20, 2026, as the date of the OpenAI update it described. The test results show why a stylization rule should not be treated as a uniform outcome for every prompt: Gizmodo reported a more realistic follow-up, while CNET reported that ChatGPT refused a similar follow-up in a different test. CNET’s prompt depicted ICE officers near a polling place.
How the other reported tests differed
The Brennan Center for Justice reported in October 2026 that ChatGPT turned nearly all of its prompts into illustrations. In its test of an image depicting a purported Department of Homeland Security memo, ChatGPT’s Instant mode produced a cartoon, while Thinking mode declined the request. Those results were tied to the Center’s prompts and modes, not a replication of Gizmodo’s ballot-disposal exchange.
The tests therefore describe outcomes for particular prompts and settings. They do not establish that ChatGPT always permits or always blocks a request to make election-related imagery photorealistic.
OpenAI’s broader election measures
On May 27, 2026, OpenAI described a wider set of election measures, including reliable voting information, cyber defense support, enforcement against election interference, political-bias monitoring, and content-provenance tools. Provenance means information that helps identify where digital content came from or how it was made. OpenAI described SynthID watermarks, C2PA metadata, and a public image-verification tool as part of that work, while noting that provenance tools alone are not a complete solution to election-related deception.
