A fake photo of Tom Holland and a pregnant Zendaya exposes the limits of AI detectors: only 55% certainty

By Zendoric · September 26, 2026. AI or Not, a platform specialized in detecting AI-generated content, catalogued on September 25 an image showing two people resembling Tom Holland and a pregnant Zendaya getting out of an SUV on a New York street. According to the source, it does not specify the scores the tool assigns, nor whether it separates the probability that the image was AI-generated from the probability that it is specifically a deepfake designed to impersonate real people; that distinction is revealing in itself, because detecting a synthetic image is not the same as certifying an intent to deceive about a specific person. Nor are the specific generators that might be behind it detailed. The record itself provides no information on who created or spread the image, its actual reach on social media, or reactions from those involved.

As industry context, this kind of deepfake tracker has multiplied over the past year alongside the proliferation of photorealistic image generators: they catalogue suspicious cases, assign them probability scores and build public databases, but they are no substitute for a full forensic verification. Fake celebrity pregnancy photos are already a subgenre of their own within this visual disinformation: they fit the demand for gossip content, are shared fast and rarely reach the end user with a clear label indicating they are synthetic.

That a detector cannot commit 100% matters more than the image itself. Verification systems work on probabilities, not binary certainties, and that leaves a grey area which anyone sharing the image —or any outlet reproducing it without checking— can exploit, with or without intent to deceive. The industry's response, rather than chasing after-the-fact detection, increasingly runs through provenance at the source: standards such as Content Credentials (C2PA), which cryptographically sign the file at the moment it is generated, so as not to depend on a classifier guessing later whether something 'looks like' AI.

Our take: this particular case is minor —one image, with no documented spread and no known harm—, but the underlying pattern is not. In the short term, the ease of fabricating hyperrealistic images of real people in intimate situations, such as a pregnancy, compounds a very human problem: the erosion of our ability to know what is true without outside help, and the reputational and emotional risk for anyone portrayed without having consented to it. In the medium term, that same pressure is what is pushing the industry toward more solid provenance infrastructure —signing at the source, not detection afterwards— and toward tools that are honest about their own margin of error instead of selling a certainty they do not have. It is uncomfortable but necessary progress: the abundance of creative tools that generative AI puts within anyone's reach will only be sustainable if the infrastructure for telling the real from the fabricated grows at the same pace.