OpenAI's Answer to Election Deepfakes: Make Them Look Like Cartoons

One week before the US midterm election on Nov. 3, OpenAI has switched on a new safeguard inside ChatGPT: any request for an image that could depict election interference or discourage voting now comes back as an illustration or cartoon, never as a photorealistic picture.

The change was announced in an Oct. 20 update, and it targets a specific weakness in the information ecosystem. AI-generated images have become so realistic that they're practically indistinguishable from real photographs, and that matters most during election season, when voters are hunting for reliable, truthful information about candidates and polling places.

How does the new restriction actually work?

The mechanism is simple, and that's the point. Ask ChatGPT to generate an image "depicting election interference or that may discourage voting," as OpenAI put it, and the chatbot will still create it, provided the request doesn't violate any other content guidelines. The catch is the rendering style: all requests to make such an image appear photorealistic are ignored.

CNET tested the system directly. Reporters asked ChatGPT to create an image of Immigration and Customs Enforcement officers standing outside a polling location. Every such prompt produced an image in a cartoon style. A photorealistic version of that scene could spark anxiety, if not outright panic, among voters who saw it out of context.

The logic cuts both ways. OpenAI isn't trying to block political speech or satire; it's trying to block deception that borrows the visual authority of a photograph. When CNET Managing Editor Jon Reed tried to trick ChatGPT into creating a photojournalist-style image showing a voting machine recording the wrong choice, the model wouldn't do it.

The takeaway: the safeguard works by degrading the image's realism, not by refusing the request outright.

What does OpenAI actually want to achieve?

Josh Lawson, OpenAI's head of elections, laid out the reasoning in plain terms. "The problem with deepfakes isn't social commentary, it's deception," Lawson told CNET. "So our aim is to degrade the evidentiary value of an image, while preserving room for illustrations and social commentary."

He framed the move as friction rather than a ban. The goal, he said, is to "make our platform less useful to bad actors by adding another layer of friction."

That distinction matters for policy watchers. A blanket refusal to generate political imagery would raise free-expression concerns and would be technically messy to enforce. Cartoonification lets illustrators, satirists and commentators keep working, while making it much harder to pass off a fabricated scene as documentary evidence. An image of ICE agents at a polling station drawn in a cartoon style simply can't be mistaken for a news photo.

The stakes are high because the underlying technology keeps getting better at exactly the wrong skill.

Why is this needed now, when image models are improving?

The safeguard lands at an awkward moment for OpenAI's own product line. Improving AI image models, like ChatGPT Images 2.5, is a double-edged sword. Newer models make fewer errors and hallucinations, which produces better content for legitimate creators. But removing those flaws makes it even harder for everyone else to identify AI-created content with the naked eye.

In other words, the closer image generators get to photographic fidelity, the less the human eye can be trusted as a verification tool. Verification and AI-labeling methods have tried to fill that gap, but those methods aren't industry-wide, and they aren't foolproof.

That's what makes style-forcing a pragmatic move. It doesn't depend on viewers checking metadata or spotting a watermark. It builds the defense into the image itself, in a way anyone can see at a glance.

The takeaway: better models demand blunter safeguards, because subtler ones no longer hold.

What are Google and Anthropic doing about election images?

OpenAI isn't the only lab facing this question, though its answer is currently the most concrete.

When CNET asked Gemini maker Google about precautions it's taking specifically for AI images, a spokesperson pointed to the company's midterm election policy. That policy says Google is "rolling out additional protections to limit the creation of misleading election-related images," but it doesn't include any details on how those protections work.

Anthropic doesn't have a traditional image or video generator, so its exposure is different. A spokesperson told CNET that any attempts to use Claude to deceive voters or disrupt elections violate its policies. The company also posted general election guidance earlier this month.

Compared with OpenAI's specific, testable mechanism, the other labs' responses remain broad statements of policy rather than demonstrated technical controls.

Will it last after the election?

There's no guarantee OpenAI will keep this system in place once the election cycle is over, CNET notes. The measure reads as an intervention tuned to a high-risk window rather than a permanent product feature.

But the signal matters. One of the biggest AI labs has shown it will alter core product behavior to protect election integrity — not just publish a policy page about it. Tech issues are taking center stage in many races this cycle, especially as AI and data center concerns rise for voters across the country, and regulators, researchers and rivals will all be watching how this experiment performs under real-world pressure.

If cartoonification works, expect the approach to be studied, copied or formalized before the next national election cycle; if bad actors route around it, the limits of model-level safeguards will be exposed just as clearly.