Since 20 October, OpenAI has automatically transformed requests for images of electoral interference or discouraging voting, even when the user requires photorealism.
In the run-up to the US midterm elections, OpenAI has chosen an intermediate route for sensitive political images. Since an update on 20 October, ChatGPT no longer systematically refuses to depict election interference or a scene that could deter voting, but it produces them in the form of stylised illustrations, ignoring any request for photorealistic rendering. The stated objective is to prevent an imaginary scene from being seen as an authentic document. CNET tested the safeguard by asking for ICE agents stationed in front of a polling station: each test resulted in a cartoon-like image.
The rest of the sector less cautious
The Brennan Center for Justice, which in October redid the tests conducted in the spring on six image generators, welcomes a clear progress by ChatGPT. The picture is less flattering elsewhere: Gemini produces more realistic images than before and without visible watermarks, Meta AI generates visuals that it previously refused, while Runway and Grok continue to manufacture credible disinformation content. Flux 3 is making slight progress.
OpenAI's protection is not watertight: NewsGuard claims that ChatGPT Images 2.5 created compromising images of candidates in 80% of the cases tested. The Brennan Center concludes that self-regulation is not enough and calls on state legislators, regulators and prosecutors to set rules and responsibilities.
Read the full article on the original source
Go to the source →