A deepfaked diabetes cure and $16B: the scam-ad problem Meta's own documents describe

Mark Cuban says he keeps getting emails about a product he has never sold. “Every single week I get emails from little old men and ladies saying, ‘Will my cure for diabetes work?' or ‘I just ordered your cure for diabetes for $294,'” the investor said on an August episode of the Pivot podcast, hosted by Kara Swisher and Scott Galloway. The ads, he says, are deepfakes built with an AI copy of his face and voice, and Meta runs them anyway: “Facebook knows the little old ladies are getting ripped off.” He called the platform “the worst of the worst.” These are Cuban's accusations; the report we have does not include a response from Meta.

The impersonation works because it is plausible. Cuban co-founded Cost Plus Drugs, an online pharmacy selling low-cost generics — so a video of him pitching cheap medicine lands inside an existing, true story about him. That is the craft of modern ad fraud: it does not invent a persona, it borrows a real one that the target already trusts.

Cuban flags one detail that deserves more attention than it gets. Some of the ads, he says, carry a disclaimer admitting the person on screen is AI-generated and the clip is meant as entertainment. By the time anyone reaches it, he argues, the pitch has already worked — “and to Facebook, that's good enough.” That is compliance theatre: a label that satisfies a policy checkbox while leaving the harm fully intact. Disclosure placed after persuasion is not disclosure. It is a receipt.

The numbers behind the anecdote come from Reuters, which reported on 6 November 2025 on internal Meta documents showing the company expected roughly $16 billion — about 10% of its 2024 revenue — to come from ads pushing scams and prohibited products. By Meta's own internal count, users across its apps were seeing some 15 billion “higher risk” ads a day, the ones carrying obvious fraud markers. That is roughly two such ads per human being on Earth, every day. A late-2024 document valued that category at about $7 billion a year, and a May 2025 presentation by Meta's own safety staff estimated the company's platforms play a part in a third of the scams that succeed in the United States. The most damning figures here are not an outside critic's; they are the company's internal estimates as reported by Reuters.

Generative AI did not create this economy, but it changed its unit economics. A credible celebrity endorsement used to require a production budget and a legal risk; now its marginal cost is close to zero and it can be regenerated the moment a creative gets blocked. Detection, meanwhile, costs money, and removal costs revenue. That asymmetry — cheap to produce, expensive to catch, profitable to miss — is the entire story. It is also the clearest example of our standing thesis that the near-term danger from AI is not a distant superintelligence but the industrialisation of ordinary fraud, aimed at the people least equipped to spot it.

Our reading: this will not be fixed with better classifiers, because the binding constraint is not technical. An ad is not speech a platform merely hosts; it is a product the platform sells, targets and optimises, and it should carry the liability that comes with selling something. The venue where that pressure is actually being applied right now is not new AI legislation but old law in the hands of courts and state attorneys general — fraud and consumer-protection statutes that already exist and already bite. Longer term, we remain genuinely optimistic: the same model class that forges a face can flag one, and an agent that sits between an 80-year-old and a fake pharmacy ad is one of the most valuable consumer products nobody has shipped properly yet. But that future does not arrive on its own while the fraud is still a revenue line. Make it a cost line, and the engineering follows within a quarter.