Deepfake scam ads aren't a moderation failure: Meta's own documents turn them into a revenue line, says Reuters

Mark Cuban says the scam has become routine. "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 are deepfakes built with an AI copy of his face and voice, Cuban said, and Meta runs them anyway. "Facebook knows the little old ladies are getting ripped off," he said, calling the platform "the worst of the worst." These are Cuban's accusations; the material we have does not include Meta's response to them.

One detail makes the fraud land harder than it otherwise would: Cuban genuinely co-founded Cost Plus Drugs, an online pharmacy selling low-cost generics. The victims are not being fooled by something implausible. They are being fooled by a believable extension of a real business, voiced by a real face. Cuban also noted that some of these ads carry a disclaimer admitting the person on screen is AI-generated and the clip is meant as entertainment. By the time anyone reads it, he argued, the pitch has already worked. "And to Facebook, that's good enough."

Then there are the numbers, which come from somewhere else entirely. On 6 November 2025, Reuters reported on internal Meta documents indicating 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 per day: the ones carrying obvious markers of fraud. Put those side by side and the shape of the problem changes. One is an anecdote about a celebrity's face. The other is a line item.

That is the thesis. Deepfake fraud has stopped being a lab demo about what generative models might enable and become a measurable component of an advertising P&L. And once it is a revenue stream, the economics run backwards from what users assume. Every point of enforcement removes revenue. Every classifier tuned tighter raises false positives on paying advertisers. Fifteen billion flagged-risk impressions a day is not a backlog any human team can work through — it is a volume only machines can police, at the exact moment the same class of machines has driven the cost of producing a convincing fake video to roughly nothing. Generation scales and gets cheaper. Verification scales badly and stays expensive. That asymmetry is the whole game.

Our read: this is the short-term AI harm we keep saying deserves more attention than distant superintelligence scenarios. The industrialization of fraud is here, it is automated, it targets the people least equipped to spot it, and it now has a dollar figure attached courtesy of the accused company's own paperwork. But we would push back on framing it as a moderation failure, because that implies someone is trying and failing. A better classifier does not fix an incentive this clean. What fixes it is making fraudulent ad revenue a cost center rather than a profit center — through liability for proceeds, mandatory refunds, advertiser identity verification with real consequences — so that the detection budget stops competing with the revenue target and starts defending it.

And hold the long view, because it is still intact. The same model architectures that clone a face can detect a cloned face; provenance and verification are tractable engineering, not science fiction. Elder-targeted ad fraud is a bounded, solvable problem. It just will not be solved by the party booking the revenue until the incentives flip. That is a governance failure, not a technology failure — and conflating the two is how we end up regulating the capability instead of the conduct.