Ahead of the US midterms next week, OpenAI says ChatGPT now turns image requests that depict election interference into cartoon-style pictures instead of photorealistic ones, as Gizmodo reported on 27 October 2026. In Gizmodo's own test, ChatGPT then produced a photorealistic version when asked. If your product generates or publishes images of real people, do not rely on a model vendor's filter. Write your own rule for elections and real people, and enforce it in your product.
What did OpenAI change?
OpenAI's election deepfake safeguard is a style rule: certain election prompts still get an image, but a stylised one. According to Gizmodo, OpenAI said it would stylise images that contain "election interference or anything meant to make people less likely to vote." Those images come out "cartoonified," to "help reduce the risk that an imagined scene is mistaken for documentary evidence."
Josh Lawson, who leads OpenAI's election work, explained the goal to Gizmodo. He said OpenAI wants to "degrade the evidentiary value of an image" and to make ChatGPT "less useful to bad actors by adding another layer of friction." Gizmodo reports that OpenAI says users can still make illustrations and social commentary.
Separately, the Washington Examiner reported on 28 October, following Politico, that OpenAI is creating an "Election Operations Center" to monitor for misinformation and possible interference around the midterms.
Does the cartoon filter work?
Partly. Gizmodo's test is the one to read. ChatGPT gave a cartoon image for an election-interference prompt. When Gizmodo then asked for a photorealistic version, "it complied." Gizmodo also cites a CNET test in which ChatGPT refused that same follow-up. So the result depends on the session, the wording, or both.
That is not a scandal. It is what you should expect from a filter that sits on top of a general image model. Filters catch the obvious request. They struggle with the second message, the reworded prompt and the edit of an image that already exists.
I think the idea is clever. Turning a fake photo into an obvious drawing is a smarter move than a flat refusal, because it keeps satire and commentary possible. But clever and sufficient are different things.
Why is a style filter not an election policy?
Because a policy answers questions a filter cannot.
A filter decides what one model outputs for one prompt. A policy decides what your product allows, for whom, with what checks, and what happens when something slips through. Who reviews reports? How fast is a fake taken down? Do you tell the person depicted? Do you keep records? Is the rule the same the week before an election as the week after?
OpenAI's filter covers the moment of generation, in OpenAI's products. It does not cover the image after it leaves ChatGPT. It does not cover images made with another tool and uploaded to your platform. It does not cover the rest of the year, or elections in other countries.
There is also the question of who decides. A vendor sets its filter for its own risks, across millions of users and many uses. Your risks are narrower and more specific. A local news site, a campaign tool and a children's app need different rules. Only you know which one you are building.
If you publish images, your product is where those images land. The vendor's filter is upstream of you. The responsibility is not.
Who needs their own rule?
More products than you might think.
- Anything that generates images from user prompts, even as a side feature, such as avatars, marketing assets or "make a poster."
- Anything that hosts user images, such as community forums, marketplaces with profile photos and local news comments.
- Anything that edits photos of people, including face swaps, background changes and "improve this picture."
- Anything that sends messages at scale, where an image can be attached and forwarded thousands of times.
What should the rule say?
- Real, identifiable people. Decide whether your product generates photorealistic images of them at all. "No, unless they consent" is a strong default.
- Elections and voting. Ban content that misleads people about how, when or where to vote. Be specific, so moderators can apply it quickly.
- Labels. Mark generated images clearly, in the image and in the metadata, so that a copy keeps the label.
- Reporting. Give people a fast way to report a fake of themselves, with a named owner and a response time.
- Records. Keep what you need to investigate a report, and no more.
- The calendar. Tighten checks around elections you know about, and say so publicly.
We wrote about who owns rules like these in AI governance as an operating model. The point here is the same. A rule nobody owns is a rule nobody enforces.
Is the vendor filter still worth having?
Yes. Defence in layers is good practice. If the model you use already turns election fakes into cartoons, that is one less thing to catch downstream. Use it. Just do not mistake it for your policy, and do not assume it will behave the same next month. Gizmodo notes OpenAI's election post was first published in May and updated in October. Vendor rules change. Yours should change when you decide, not when they do.
What would I do this week?
If your product touches images of people, write one page: what you allow, what you block, how people report a fake, and who answers. Test it against the Gizmodo approach: ask for a cartoon, then ask for the real thing. If your product lets the second request through, fix that before the vote.
Where Alongside fits
Writing the rule for images of real people, and naming who enforces it, is the kind of work covered by Alongside's technical advisory.