OpenAI 如何遏制选举中的深度伪造干扰

OpenAI plans to build tools that keep bad actors from using deepfakes to meddle in US elections. The November 2026 midterms are approaching, and AI image generators have never been more convincing.

What OpenAI is rolling out

The most concrete move came on October 11, 2026. That day, OpenAI launched a ChatGPT feature designed to restrict realistic images that could mislead voters.

The targets are specific. The feature focuses on imagery about polling locations, ballots, and election-day logistics.

The October feature sits inside a broader 2026 election integrity strategy. OpenAI laid out that plan in May 2026, describing a mix of partnerships and technical safeguards.

On the partnership side, OpenAI teamed with Democracy Works to point users toward accurate voting information. It also worked with the Associated Press on live election results.

Watermarks, metadata, and legislation

OpenAI is deploying SynthID watermarking technology in collaboration with Google. OpenAI is also following C2PA metadata standards.

OpenAI has committed to supporting the Protect Elections from Deceptive AI Act. It is also providing cybersecurity tools to election officials.

OpenAI has actively disrupted foreign influence operations linked to Iran and Russia, from 2024 through October 2026.

The guardrails are not airtight

In September 2026, the Brennan Center for Justice tested AI models, including ChatGPT. Its assessment found that existing safeguards could be circumvented. Testers were able to produce convincingly fake election imagery despite the restrictions.

That testing came roughly a month before the October 11 feature launched. Whether the new restriction closes the gaps the Brennan Center found remains an open question.

What this means for AI companies and voters

For voters, the practical takeaway is caution. Labels, watermarks, and chatbot refusals reduce risk, but the Brennan Center findings show they do not eliminate it.

For the AI industry, OpenAI backing federal legislation signals that at least one major player sees rules as inevitable and wants a voice in shaping them. Building watermarking, provenance standards, and abuse detection is not cheap, which could raise the bar for smaller developers.

The near-term test is the November 2026 midterms themselves. OpenAI has stacked up partnerships, watermarks, policy commitments, and a new ChatGPT restriction. The Brennan Center has already shown that determined users can find a way around similar defenses.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.