OpenAI blocked 250k election deepfakes

Heading into the 2024 U.S. election, everyone braced for the AI deepfake apocalypse. Now that the dust has settled, OpenAI pulled back the curtain on November 8 to show what actually happened behind the scenes.

Turns out, ChatGPT's content filters were working overtime.

250,000 blocked portraits and 3 million redirects

The most striking number came from DALL·E.

In the single month leading up to Election Day, ChatGPT shot down more than 250,000 requests to generate images of public figures, specifically targeting President-elect Donald Trump, Vice President Kamala Harris, Vice President-elect JD Vance, President Joe Biden, and Governor Tim Walz.

It wasn't just images. ChatGPT was strictly programmed to refuse political endorsements, even when users pushed hard to make it pick a favorite candidate.

When it came to hard voting facts, OpenAI decided the safest play was passing the microphone.

Roughly 1 million queries about voting logistics—like where to cast a ballot—were pointed straight to CanIVote.org through a partnership with the National Association of Secretaries of State.

Then on Election Day and the day right after, around 2 million responses told users checking on live ballot counts to go read the Associated Press and Reuters instead.

Covert campaigns failed to go viral

Beyond everyday users trying to generate spoof portraits, OpenAI was hunting for covert influence networks.

The company has been publishing disruption reports since May, tracking state-backed and malicious campaigns attempting to use GPT models for covert propaganda.

According to OpenAI, their threat intelligence teams monitored systems closely right up through Election Day and found no evidence of election-related influence ops gaining viral traction or building sustainable followings using their tools.

Backing that up was tech like C2PA cryptographic metadata credentials baked into DALL·E 3 and restricted API policies that outlaw political campaigning apps or bots impersonating politicians.

Why it matters

This election served as the premier live-fire stress test for frontier AI, and OpenAI's playbook was surprisingly simple: know when to shut up.

For everyday users, that meant friction. You couldn't generate candidate memes or get real-time election forecasts. But that friction was intentional.

By punting 3 million high-stakes election questions to newsrooms like AP and Reuters, OpenAI dodged the deadly hallucination trap. A single hallucinated swing-state call could have sparked a bipartisan regulatory nightmare.

We think this ultra-defensive posture worked remarkably well for OpenAI's brand safety. But let's be real: locking down ChatGPT only secures ChatGPT.

Malicious actors looking to flood social feeds with synthetic audio and fabricated clips will simply shift to uncensored, open-weight models where no corporate red team is watching.