AI washing: what it means now, what has been punished, and how to test a claim

The word was built on greenwashing, and its first meaning is the same shape: a company claims more AI than it has. The SEC’s then-chair gave the plainest definition when the agency brought its first cases in March 2024: “Investment advisers should not mislead the public by saying they are using an AI model when they are not. Such AI washing hurts investors.” Canada’s securities regulators defined it formally later that year as inaccurate, false, misleading, or embellished claims about the use of AI systems.

The second meaning arrived in early 2026. As companies attributed large layoffs to AI, the term was turned on them: AI washing as crediting AI with cuts that had other causes. Reporting in February noted that AI had been the stated reason for more than 50,000 layoffs in 2025, and quoted a Forrester report finding that many of the companies announcing AI-related layoffs did not have mature AI applications ready to fill those roles. Later that month OpenAI’s chief executive said there was “some AI washing where people are blaming AI for layoffs that they would otherwise do,” and, in the same breath, some real displacement.

The two meanings look opposite and are the same act. In one, AI is claimed in a product where it is absent. In the other, AI is claimed in a decision where it was absent. Both borrow the credibility of the technology for something it did not do, and both are aimed at investors.

No statute mentions AI washing. Every case so far has been brought under laws that predate the technology: securities fraud, false advertising, wire fraud. What follows is what regulators have actually done, read from their own documents.

The SEC

The first cases, in March 2024, were small and deliberate. Delphia, a Toronto investment adviser, said it used client data to make its AI smarter at picking investments; the SEC found it did not have the capabilities it claimed. Global Predictions called itself the first regulated AI financial advisor. They paid $225,000 and $175,000 without admitting or denying the findings. Three months later the agency charged the founder of Joonko, an AI recruiting startup, with defrauding investors of at least $21 million, which its enforcement director called “old school fraud using new school buzzwords.”

Presto Automation, in January 2025, involved a listed company and a shipping AI product, and it remains the most instructive case because nothing about it was exotic. The product existed. It was deployed in real restaurants. The misstatements were about who owned the technology and how much of the work people were doing. Presto received a cease-and-desist order and no fine, with the SEC citing its cooperation and remedial efforts.

In April 2025 the Justice Department and the SEC charged the founder of Nate, a shopping app marketed as completing purchases with no human involvement. The indictment alleges the founder told an investor that success rates ran from 93% to 97%, while the app’s actual automation rate was effectively zero and hundreds of contractors in a call center in the Philippines completed the purchases by hand. Prosecutors say he raised more than $40 million; the SEC’s complaint puts it at over $42 million. These are allegations, and the case had not been resolved when this guide was written.

Institutionally, the SEC folded AI into a new Cyber and Emerging Technologies Unit in February 2025, with fraud involving AI and machine learning first on its list of priorities, and its 2026 examination priorities say staff will review the accuracy of registrants’ representations about their AI capabilities. We found no new SEC case against a public company over AI capability claims between Presto and September 2026. The posture is watchful rather than busy.

The FTC

The FTC polices advertising, so its cases are about what customers were told. Operation AI Comply, in September 2024, announced five actions at once. The best known was DoNotPay, which promoted itself as the world’s first robot lawyer; the FTC said it had never tested its output against a human lawyer’s and employed no attorneys. It settled for $193,000.

The cases since have followed a pattern: Workado for an accuracy claim that did not survive testing, Air AI in March 2026 for selling “conversational AI” said to replace customer-service staff, and Cox Media Group in May 2026 for an AI service with no AI in it. The agency has also drawn a line. In December 2025 it set aside its own 2024 order against Rytr, an AI writing tool, saying the complaint had not met the requirements of the law. Guides that still cite Rytr as precedent are out of date. The FTC is pursuing claims that are false, not tools that could be misused.

Shareholders

The larger financial exposure is private. Cornerstone Research and Stanford count securities class actions with AI-related claims: seven in 2023, 15 in 2024, 16 in 2025, and 15 in the first half of 2026 alone. Those 15 were a modest share of all filings and accounted for 73% of the alleged investor losses in the period. “AI-related” is broader than AI washing, and other trackers count slightly differently, but the direction is not in doubt: for a public company, the realistic penalty for an AI claim that does not hold up is a shareholder suit, not a regulator’s fine.

Outside the United States

Canadian securities regulators have been the most explicit, telling fund managers that vague and unsubstantiated statements that incorporate jargon to attract investors should not be made. In Britain the Advertising Standards Authority found about 16,000 paid ads using “AI” in a single quarter of 2024 and advised advertisers not to imply capabilities that do not exist.

Through 2026 a long line of large employers cut staff and cited AI, among them Block, Atlassian, PayPal, Cloudflare, Cisco, Intuit, and Meta, according to TechCrunch’s running list. Whether any one of those announcements is AI washing is exactly what an outsider cannot tell, and that is the point. A product claim can be tested: you can run the system. A claim that AI made four thousand people unnecessary cannot be, because the counterfactual is not available to anyone outside the company.

What can be checked is consistency. A company that has really replaced work with AI will show it elsewhere: in what it hires for afterward, in the AI spending it reports, in output per employee over the following quarters. If similar roles reopen within a year, the explanation deserves a second look. Anyone evaluating such a claim, as an investor or as someone deciding whether to take a job there, should look at the twelve months after the announcement, not the press release.

Most AI washing is not a scheme. It is a marketing page that got ahead of the engineering, a fundraising deck nobody on the technical team reviewed, or a metric that started with a footnote and lost it. The SEC’s enforcement director described the standard simply: if you claim to use AI, make sure the representation is not false or misleading. In practice that comes down to five habits.

• Say whose model it is. “Built on” is an honest phrase and customers do not mind it.
• Define every number. Accuracy on what data, automation as a share of what. If the footnote would embarrass you, the number is the problem.
• Disclose the people. Human review is a feature in most serious deployments. Describe it as one.
• Keep the tense straight. What the product does, what it will do, and what you hope it will do are three different sentences.
• Have an engineer read the deck. Before it goes to investors, and again after anyone edits it.

The same discipline inside the company is AI governance, and the employee-side version of claims drifting from reality is shadow AI.