ChatGPT’s election-image safeguard fails a photorealism follow-up in Gizmodo test

ChatGPT turned an invented ballot-tampering scene into a cartoon, then made it photorealistic when asked again, Gizmodo reported on October 27. The result exposes a gap in OpenAI’s newly disclosed election-image safeguard. CNET’s earlier tests held the line against realistic follow-ups, leaving a mixed picture ahead of the November 3 US midterm election.

One follow-up changed the result

Gizmodo writer Mike Pearl asked for a surveillance-camera-style image showing someone secretly throwing ballots away near a polling place. The fictional scene was set outside a church in Duluth, Minnesota, with voting signs and people entering and leaving in the background. ChatGPT initially produced a cartoon, matching the behavior OpenAI had described. A subsequent request to make that image photorealistic succeeded.

Pearl said this was his first attempt to generate an election-interference deepfake. He described the final image as somewhat staged but plausible enough that he would not publish a higher-resolution version as the article’s lead image without a watermark. That is the reporter’s assessment of its appearance, not evidence that anyone mistook it for a real event.

Gizmodo said it contacted OpenAI on Monday night and had received no response by publication. Its article therefore offered no company explanation for why this follow-up passed the safeguard.

A cartoon, not a blanket refusal

OpenAI disclosed the measure in an October 20 update to its election-safeguards article. Requests depicting election interference or scenes that might discourage voting can still produce images, provided they do not violate other content rules. But those images are supposed to appear as cartoons or illustrations, rather than photographs. Requests for photorealism should not override that treatment.

The intended distinction is between an illustration expressing an idea and an invented scene presented as documentary evidence. OpenAI’s head of elections, Josh Lawson, told CNET that the company wanted to preserve illustrations and social commentary while making the platform less useful to bad actors. He described the safeguard as another layer of friction, rather than a ban on political imagery.

The problem with deepfakes isn’t social commentary, it’s deception. So our aim is to degrade the evidentiary value of an image, while preserving room for illustrations and social commentary

Josh Lawson, OpenAI’s head of elections, speaking to CNET

CNET’s requests met a firmer boundary

In tests published on October 26, CNET asked ChatGPT to depict Immigration and Customs Enforcement officers outside a polling location. The images remained cartoon-style, including after requests for realistic treatment. Its screenshot shows a request explicitly seeking a photorealistic, photojournalist-style image. ChatGPT instead supplied an illustrated version, demonstrating the distinction OpenAI says it wants the system to maintain.

CNET managing editor Jon Reed also tried to obtain a photojournalist-style image of a voting machine recording a choice different from the voter’s selection. That attempt did not produce the requested realistic image either. The outlet’s examples covered both an intimidating presence at a polling site and an invented malfunction inside the voting process.

Gizmodo also sent its ballot-tampering prompt to Google Gemini. Gemini refused, saying it could not generate unsafe images. In that instance, the refusal stopped the fictional scene from being created at all, rather than preserving it as commentary in a less realistic style. That comparison concerns the same prompt, unlike the different scenarios tested by CNET and Gizmodo in ChatGPT.