ChatGPT rejected 250K deepfake requests during US election

OpenAI’s election playbook just got its first real-world test, and the early numbers tell an interesting story. In the month leading up to the 2024 US election, ChatGPT batted away over 250,000 requests to generate DALL·E images of major candidates — Trump, Harris, Biden, Vance, and Walz. That’s not a rounding error. It’s a signal that people are probing the system’s guardrails at scale, and that those guardrails mostly held.

The company also leaned hard into redirecting users toward authoritative information rather than attempting to answer political questions itself. Roughly 1 million ChatGPT responses in the pre-election month pointed people to CanIVote.org, a partnership with the National Association of Secretaries of State. On Election Day and the day after, around 2 million responses nudged users toward the Associated Press and Reuters for results. The approach is straightforward: don’t guess, don’t editorialize, just route people to sources that actually know what they’re talking about.

That philosophy extended across the Atlantic, too. Ahead of the European Parliament elections, OpenAI rolled out a similar redirect system pointing users to elections.europa.eu for voting questions — a mirror of the NASS collaboration in the US. And back in May, the company joined the steering committee of C2PA, the Coalition for Content Provenance and Authenticity, while giving researchers early access to a tool that identifies DALL·E 3 outputs. The provenance play is a long game, but it’s becoming table stakes for any AI firm that wants to be taken seriously during election cycles.

What didn’t happen might be just as telling. OpenAI’s threat intelligence teams monitored for covert influence operations using their models and found no evidence of election-related campaigns gaining viral traction or building sustained audiences. That doesn’t mean nobody tried — the company disrupted several operations earlier in the year. But it suggests that the combination of detection systems and outright refusals to generate certain content is making these tools less useful for bad actors than some feared. The real question now is whether other platforms can say the same.