● What changed: photorealistic political imagery restricted in ChatGPT
● Instead returns: illustrated or cartoon-style images
● Timing: ahead of the US midterms
● Scope: OpenAI's own surface; open-weight models unaffected
OpenAI is restricting photorealistic political deepfakes in ChatGPT ahead of the US midterms. Election-related image requests now tend to return illustrated or cartoon-style output rather than photographs.
The implementation is the interesting choice. Rather than refusing outright, the model redirects to a style that cannot be mistaken for evidence - which preserves legitimate uses like illustration and satire while removing the output that does the damage.
It is also a constraint on one vendor only. Open-weight image models have no equivalent restriction and cannot be given one retroactively, so the policy shapes what the largest consumer surface produces rather than what is possible.
Sources
The Neuron, 27 October.