OpenAI Has a Surprising Defense Against Election Deepfakes

One of the biggest problems of the AI age is that AI deepfakes are now so realistic that they’re practically indistinguishable from real photos. That matters a lot during election season, when voters are looking for reliable, truthful information on their candidates and polling sites. OpenAI is implementing a new safeguard to prevent people from making political deepfakes.

The US midterm election is only a week away on Nov. 3. Tech issues are taking center stage in many races, especially as AI and data center concerns rise for folks across the country. (Check out CNET’s guide to spotting political scams.)

If you were to go into ChatGPT today and ask it to generate an image “depicting election interference or that may discourage voting,” as OpenAI put it in an Oct. 20 update, the chatbot would create the image, given it doesn’t violate any other content guidelines. But it will be in an illustrated or cartoon style — all requests to have the image appear photorealistic should be ignored.

More from CNET

For example, CNET asked ChatGPT to create an image of Immigration and Customs Enforcement officers standing outside a polling location. Every such prompt generated an image in a cartoon style. A photorealistic version of this kind of image would likely spark anxiety, if not outright panic, for voters.

The idea is to make it harder for bad actors to create realistic political deepfakes that could interfere with or question the integrity of an election without stymying people’s ability to create political content in general. In another example, CNET Managing Editor Jon Reed couldn’t trick ChatGPT into creating a photojournalist-style image that showed a voting machine recording the wrong choice.

“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, told CNET. The goal is to “make our platform less useful to bad actors by adding another layer of friction,” he said. 

This cartoonification of AI images is a practical way to make political deepfakes a little harder to create. Improving AI image models, like ChatGPT Images 2.5, is a double-edged sword. Newer models make fewer errors or hallucinations, making better content for creators. But removing those flubs makes it even harder for everyone else to identify AI-created content with the naked eye. Verification or AI labeling methods have tried to help, but those methods aren’t industry-wide nor foolproof.

When I asked Gemini maker Google about any precautions it is taking specifically for AI images, a spokesperson pointed me to its midterm election policy, which says the company is “rolling out additional protections to limit the creation of misleading election-related images” but doesn’t include any details on how. 

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

There’s no guarantee that OpenAI will keep this system in place once the election cycle is over. But it’s a solid sign from one of the biggest AI labs that it is taking its role in safeguarding American elections’ integrity at least somewhat seriously.

For more, check out what to do when you’ve been deepfaked and everything to know about the Take It Down Act.

Reporter 2

Katelyn is a reporter with CNET covering artificial intelligence, including chatbots, image and video generators. Her work explores how new AI technology is infiltrating our lives, shaping the content we consume on social media and affecting the people behind the screens. She graduated from the University of North Carolina at Chapel Hill with a degree in media and journalism. You can reach her at kchedraoui@cnet.com. See full bio