AI-Generated Video Falsely Showing Trump Watching Girls Dance With Epstein

The source image is a real photograph taken April 28, 1997, at a Victoria's Secret "Angels" party in New York City, credited to photographer Sonia Moskowitz and held by Getty Images. Its caption names Donald Trump, Jeffrey Epstein, and Belgian model Ingrid Seynhaeve — no children appear in it ^[2].

The fake video kept the real photo as its opening frame and used AI to animate it and insert fabricated scenes of children, according to Italian fact-checking outlet Open ^[1].

This was not an isolated fake. Full Fact separately debunked a still image claiming to show Trump dancing with a 13-year-old on Epstein's island, citing an unnaturally elongated forearm, a distorted thumb, and — per digital forensics expert Hany Farid — a reflection present in one eye but missing from the other ^[3].

A NewsGuard test of AI image generators found that Grok Imagine (xAI) produced "convincing fakes in seconds" of Epstein with several politicians, including a realistic image of a younger Trump with Epstein and young girls; Google's Gemini refused that specific combination but generated other politicians with Epstein, and ChatGPT declined entirely. No authentic photo of Trump with Epstein in the presence of underage girls is known to exist ^[4].

Separately, viral images claiming to show Trump and Bill Clinton with Epstein in "compromising positions" were confirmed AI-generated by multiple detection tools, including Google's SynthID watermark detector ^[7].

Fabricated images of New York Mayor Zohran Mamdani and his mother, filmmaker Mira Nair, alongside Epstein also circulated, several bearing a Google SynthID watermark confirming AI origin ^[5].

Other fabricated claims rated false include a viral audio clip claiming Ellen DeGeneres was a cannibal who "ate children's flesh," rated Pants on Fire by PolitiFact after deepfake-detection analysts found the clip was likely AI-generated ^[8].

X's Grok chatbot amplified the crisis. Reviewing requests from January 30 to February 5, 2026, Bellingcat found Grok generated images in response to 27 of 31 user requests asking it to "unblur" or identify women and children in Epstein-file images — about 87% — though quality "ranged from believable to comically bad." In a follow-up review from February 7–9, Grok did not fulfill any of 16 similar requests ^[6].

The episode drew a rapid, uncoordinated regulatory response. Malaysia and Indonesia blocked Grok in early January 2026 over non-consensual sexualized imagery, and the Philippines followed on child-protection grounds; the UK's Ofcom opened an Online Safety Act investigation on January 12, the EU opened Digital Services Act proceedings on January 26, and 35 U.S. state attorneys general sent xAI a joint letter ^[9].

Separately from the social-media fakes, digital forensics expert Hany Farid, working with WIRED, analyzed nearly 11 hours of DOJ-released prison surveillance footage from outside Epstein's cell. Metadata showed the file was modified — likely with Adobe Premiere Pro — and assembled from at least two source clips rather than exported directly from the prison's system ^[13].

Farid said the metadata raises chain-of-custody concerns independent of any AI question: "If a lawyer brought me this file and asked if it was suitable for court, I'd say no. Go back to the source. Do it right. Do a direct export from the original system—no monkey business" ^[13].

The TAKE IT DOWN Act, signed May 19, 2025, criminalizes posting nonconsensual intimate images — including AI-generated ones — and requires covered platforms to remove flagged content within 48 hours of a valid request, enforced by the FTC ^[10].

The Senate passed the DEFIANCE Act by unanimous consent on January 13, 2026, which would let victims sue over nonconsensual AI-generated intimate images; a House companion bill has not advanced ^[11].

The C2PA content-provenance standard and Google's SynthID watermarking are the leading technical responses. Google joined C2PA's steering committee and helped develop a more tamper-resistant version of the standard, and plans to extend SynthID — already used to help identify several of the fakes above — to more generative tools and media types ^[12].

The volume of AI-generated Epstein content outpaces manual verification. Meta reportedly spent roughly $65 million on AI classifiers for election-related misinformation and deepfakes, but Stanford Internet Observatory researchers found synthetic media still circulated for hours or days before removal, with a "stubbornly high" false-positive rate ^[15].

Globally, the Duke Reporters' Lab counted 443 active fact-checking organizations as of June 2025, a number that has been declining as funding — including Meta's since-ended U.S. fact-checking partnership — has dried up ^[17].

AI tools help fact-checkers but have real limits. Norwegian outlet Faktisk Verifiserbar uses AI for geolocation and translation, but Ghanaian and Georgian fact-checkers report the tools are built from a Western perspective and lack training in local languages and context ^[16].

The viral video of Trump and Epstein watching girls dance is AI-fabricated, built from a genuine 1997 photo that shows neither man with children. It is one of many AI-generated Epstein images that spread after the DOJ's 2026 document release, exploiting a wave of renewed public attention faster than platforms or fact-checkers could respond.