AI Deepfake Ads Test Whether Meta Values Trust

Mark Cuban says deepfake ads using his face and voice are pushing fake diabetes treatments to older adults. Internal Meta documents cited in reports raise a larger concern: how much fraud can an ad platform tolerate when the revenue keeps flowing?

Cuban described the problem during an August episode of the Pivot podcast, hosted by Kara Swisher and Scott Galloway. He said people contact him every week after seeing advertisements for a supposed $294 diabetes cure. Cuban has never sold that treatment, yet the ads make it appear that he has endorsed it.

The ads reportedly use artificial intelligence to imitate Cuban’s appearance and voice. Some include a disclaimer saying the person shown is AI-generated and that the video is intended as entertainment. Cuban’s view is that the disclaimer arrives after the central sales pitch has already done its work. A viewer sees a familiar face discussing a health product, follows the link, and may spend hundreds of dollars before checking whether the offer is genuine.

That makes the issue larger than a few misleading videos. The case puts pressure on Meta’s advertising systems, its enforcement thresholds, and the financial incentives behind them. A platform that distributes billions of ads each day has to decide how much evidence it needs before cutting off a suspected scammer. Those decisions shape whether users can treat an ad as a reasonable signal of legitimacy or as a piece of paid content that may have cleared only a weak automated test.

Deepfake fraud depends on borrowed credibility. Cuban is known as an investor, television personality, and co-founder of Cost Plus Drugs, an online pharmacy that sells generic medication at lower prices. That background gives a fake medical advertisement a ready-made story. Viewers may assume that a well-known business figure has entered the health market or is sharing a personal breakthrough.

Older adults can face particular pressure from health claims. A diabetes treatment promises relief from an ongoing condition, and a recognizable spokesperson can make an unfamiliar website appear safer. The price point also matters. A charge of $294 is large enough to cause harm while remaining within reach for some people who may hesitate to challenge a transaction after it happens.

Cuban said he hears from people who believe they bought a product he endorsed. Their messages show the practical result of the deception. They are not debating whether a video looks artificial. They are trying to find out whether a product they purchased will work.

The presence of an AI disclaimer does little to correct that confusion. A small label can be easy to miss, especially on a phone screen or inside a fast-moving social feed. It also may not explain that the product itself is fake, that Cuban never approved the advertisement, or that the seller has no connection to him. Disclosure can become a box-checking exercise when the main message remains visually convincing.

Mark Cuban’s account of AI-generated ads promoting a nonexistent diabetes cure illustrates how deepfake fraud turns celebrity familiarity into a sales tool. The reported ads expose a broader challenge for Meta: its automated systems must balance mistaken enforcement against the financial and health harms of letting suspicious campaigns continue. Internal estimates cited in reports, Meta’s disputed response, and losses reported by older adults all point to the same accountability question—whether the platform treats trustworthy distribution as a core responsibility or as a cost to manage after revenue is collected.

Reuters reported on internal Meta documents that estimated roughly $16 billion, or about 10 percent of the company’s 2024 revenue, came from advertisements linked to scams and prohibited products. Another internal document reportedly valued a category of higher-risk ads at about $7 billion annually. The figures are estimates, and Meta has challenged how they were calculated, but they point to a substantial business problem.

The same documents reportedly said users across Meta’s services encountered about 15 billion higher-risk ads each day. The category included advertisements with clear warning signs of fraud. A May 2025 presentation from Meta safety staff estimated that the company’s platforms played a role in roughly one-third of scams that succeeded in the United States.

Those numbers describe scale, not the outcome of every ad. They also show why enforcement can be difficult. Meta operates advertising systems that process enormous volumes of campaigns, accounts, landing pages, payments, and user reports. Fraudsters can create new identities quickly, shift between accounts, alter a product page, and change the wording of an advertisement after approval.

Still, scale cannot answer the accountability question by itself. An automated system may miss a dangerous ad because the advertiser stays just below a detection threshold. If the company earns money while collecting more data, a delay in enforcement can become part of the economic model rather than an occasional technical failure.

The reported documents state that Meta’s automated systems must determine that an advertiser is at least 95 percent likely to be running a scam before the account is banned. Suspected advertisers below that level can continue buying ads, even if they pay more for the opportunity.

A threshold like that may reduce mistaken bans against legitimate businesses. It also gives a scammer room to operate while the system gathers enough evidence. A deceptive advertiser may need only a short window to collect payments, capture personal information, or move a victim to another site.

The financial tradeoff deserves close attention. A lower threshold could remove more fraudulent ads, but it could also block legitimate advertisers and create disputes for small businesses. A higher threshold preserves access for more accounts while allowing more questionable campaigns to run. Meta’s challenge is to show how it measures those costs and whether it treats financial harm as a greater risk than temporary disruption to an advertiser.

The platform also has information that individual users lack. Meta can see patterns across accounts, payment methods, domains, ad copy, audience targeting, and previous enforcement actions. A single advertisement may look plausible in isolation. A network of related campaigns can reveal a pattern that a viewer cannot see. That data advantage increases the company’s responsibility to use it effectively.

Meta should be judged not only by how many scam ads it removes, but by how long dangerous campaigns remain active and what the company does with the warning signals it already possesses. A celebrity deepfake selling a supposed medical cure is not merely misleading content; it combines identity theft, health misinformation, financial risk, and targeted distribution. Platforms that profit from delivering such ads should publish enough evidence about detection thresholds, response times, repeat offenders, and refunds for outsiders to evaluate whether their safeguards protect users or primarily protect ad volume.

Meta spokesman Andy Stone told Reuters that the internal documents offered a selective view and distorted the company’s approach to fraud and scams. He called the 10 percent revenue estimate rough and overly inclusive. Stone also said user complaints about scam ads had fallen 58 percent over the prior 18 months and that Meta removed more than 134 million scam ads in 2025.

Those enforcement numbers matter. Removing millions of ads suggests that Meta has active detection and review systems, and a decline in complaints may indicate progress in some parts of the platform. Reports from users, however, capture only the harm people recognize and choose to report. A victim may blame a bank, a product seller, or personal judgment without reporting the social platform that delivered the ad.

The documents also reportedly noted that the Securities and Exchange Commission was examining Meta over advertising connected to financial fraud. An investigation does not establish wrongdoing, but it adds regulatory pressure to questions about how Meta identifies risky advertisers and represents the quality of ads shown to users.

The most useful public debate would focus on measurements that outsiders can test. Meta could disclose how quickly it removes scam advertisements, how often banned advertisers return, how many reports lead to action, and how much revenue it refunds or forfeits after an ad is found to be fraudulent. The company could also explain how it handles celebrity likeness abuse and whether AI-generated ads receive additional review.

Advertising platforms often describe themselves as neutral marketplaces that connect businesses with audiences. Their systems also decide who can buy access, which users receive a message, how prominently it appears, and how long it stays live. Those choices give the platform a role in the transaction even when the seller is an outside company.

That role becomes especially important when a scam uses a public figure’s identity. The advertiser is purchasing more than screen space. It is purchasing distribution through a system that many users may associate with a basic level of screening. Approval does not guarantee that a product is legitimate, but repeated approval can create that impression.

Cuban’s criticism targets this part of the business. He called Facebook ‘the worst of the worst’ and argued that the company has accepted a culture in which scammers can target people’s grandparents. The statement is blunt, but it identifies a concrete issue: Meta’s priorities become visible through the harms it accepts while campaigns remain profitable.

The question for Meta is therefore broader than whether a particular deepfake violated a rule. It concerns the cost of keeping uncertain ads in circulation. If an advertisement earns money during the period before a system reaches its 95 percent confidence threshold, the company has a reason to move cautiously. Victims and advertisers operating honestly have a reason to expect faster action.

The FBI’s 2025 Internet Crime Report found that people age 60 and older reported $7.7 billion in losses across 201,266 complaints. That represented a 59 percent increase from the previous year. Investment scams caused about $3.5 billion in losses, while technology and customer support scams accounted for another $1.04 billion.

AI appeared in more than 3,100 complaints, with reported losses above $352 million. Those numbers reflect only cases reported to authorities. Shame, confusion, uncertainty about where to complain, and the belief that recovery is impossible can all keep victims silent.

Health scams can carry additional consequences. A person may lose money, delay legitimate treatment, share medical information, or give a scammer access to a payment account. Family members may see the financial loss first and miss the fear or embarrassment behind it.

The Cuban ads show why platform safeguards need to account for context. A deepfake of a celebrity selling a harmless novelty product presents one type of risk. A deepfake promoting a supposed diabetes cure presents another. The health claim, the target audience, the price, and the use of a stolen identity should all raise the level of review.

People who see a celebrity promoting a health product through a social feed should pause before clicking or paying. A few checks can expose many fake campaigns:

Cuban’s name should receive the same scrutiny as any other celebrity name. His public criticism does not make every advertisement featuring him legitimate, and scammers may use a real interview beside a fake sales pitch to make the deception harder to spot.

Meta has the technical ability to detect more than a viewer can. It can compare a video with known recordings, identify reused voices, track domains connected to previous scams, and examine payment patterns across accounts. Those tools will not eliminate fraud, but they can reduce the time a campaign remains active.

The central decision is whether stronger detection and faster review receive the same priority as ad growth. A platform’s public statements can emphasize removals and falling complaint totals. Users experience the system through the ads that reach them, the speed of reporting tools, and the outcome after money disappears.

For Meta, trust is measured in those ordinary moments. A grandparent sees a familiar face, believes a medical claim, and pays a stranger. The company may later remove the ad and point to enforcement totals. The financial loss, however, has already happened.

Cuban’s complaint gives that policy debate a human face, while the reported internal estimates give it a financial one. Together, they raise a direct question for Meta and other ad platforms: how much revenue should be allowed to pass through before the evidence of deception becomes clear enough to act? The answer will show whether trustworthy distribution is treated as a core responsibility or as a cost to manage after the money is counted.