OpenAI Fights Election Deepfakes by Making Them Look Fake

The most unexpected part of OpenAI's plan to blunt election deepfakes is aesthetic. When someone asks ChatGPT for an image that depicts election interference or could discourage people from voting, the company now produces stylized, cartoon-like illustrations instead of photorealistic ones. The idea is that a drawing is much less likely to be mistaken for documentary evidence, while still leaving room for satire and social commentary. OpenAI says its ban on election interference and voter demobilization applies even when the request uses cartoonish imagery.

The change is one piece of an update the company published ahead of the U.S. midterm elections. OpenAI also points to tools aimed at the authenticity problem more directly. It is adding SynthID watermarks, developed by Google DeepMind, to images generated through ChatGPT, Codex, and the API. It continues to use the C2PA standard, which attaches cryptographically signed metadata to a file. The two methods are meant to complement each other: the watermark is invisible and can survive screenshots, while the metadata can carry more detail. OpenAI is also previewing a public verification tool that lets people check whether an image they found elsewhere was made with OpenAI's models.

Other measures in the update cover voting information and security. ChatGPT now links to official election sources for registration, deadlines, and polling places through a partnership with Democracy Works, and it will show live vote counts from The Associated Press in the U.S. and Brazil after polls close. OpenAI has also offered cyber defense access to registered voting-system manufacturers and is briefing state election officials.

Political advertising remains off the table. OpenAI says it will not run political ads on its platform this cycle, and it still bars using its products to produce scaled campaign messaging for or against a candidate, party, or ballot measure. The company has also said it supports the Protect Elections from Deceptive AI Act, which would make it illegal to knowingly distribute materially deceptive AI-generated content about federal candidates.

OpenAI is careful not to oversell the approach. It describes provenance tools as an important layer rather than a complete fix, and notes that most people see AI-generated content on social media and messaging apps, where platforms would need to act on those signals. Whether cartoonified imagery and watermarks are enough to slow real-world deception during a close election cycle remains to be seen.