AI-washing

AI-washing is claiming more for artificial intelligence than a company actually has or does: capabilities it lacks, automation that is really people, accuracy it cannot show, or an AI strategy that exists mainly in the annual report. It is the AI version of greenwashing, and the incentive to do it has never been stronger: two thirds of S&P 500 earnings calls now mention AI, and investors have rewarded the companies that do. Regulators in the US have fined firms for it, and securities lawyers are suing over it at a growing rate. In the UK there has been no case yet, but the routes to liability already exist, and from 2027 boards will be declaring on the effectiveness of the controls behind what they report. For directors the question is simple to state: could we evidence every claim we make about AI?

Where the term came from#

The phrase was in use before regulators took it up, and nobody can be credited with coining it. It became an enforcement term in the United States. The then chair of the Securities and Exchange Commission, Gary Gensler, told an AI conference in Washington on 5 December 2023, as the D&O Diary and others reported, that "one shouldn't greenwash and one shouldn't AI wash."

On 18 March 2024 the SEC settled its first AI-washing cases, against two investment advisers. Delphia had claimed for years to use AI and machine learning on its clients' data, a capability the SEC found it lacked; Global Predictions called itself the "first regulated AI financial advisor". They paid $225,000 and $175,000. Gensler's line was blunt: "Such AI washing hurts investors."

The first public company followed on 14 January 2025. The SEC found that Presto Automation had said its AI ordering product "eliminated the need for human order-taking" when most orders needed human intervention, and had not disclosed that the speech recognition came from a third party. It imposed no penalty, citing Presto's cooperation. In April 2025 federal prosecutors in New York charged the founder of a shopping app, Nate, with raising more than $40 million on claims that purchases were completed without human intervention, when, prosecutors allege, the automation rate was effectively zero and the work was done by people in a call centre. Those are allegations, not findings.

Under its current chair the SEC has said its examiners will check the accuracy of firms' AI claims, but I have found no new AI-washing enforcement case in 2026. The risk has moved somewhere else.

Where the exposure is now: litigation#

Shareholder lawsuits have overtaken regulators as the main source of risk. Cornerstone Research's midyear assessment of 29 July 2026 counted 15 AI-related US securities class actions in the first half of 2026, against 16 in the whole of 2025. They were 13 per cent of core filings but 73 per cent of the losses investors alleged. As Stanford's Joseph Grundfest put it, a "modest share of total filings but an outsized share of alleged investor losses". Cornerstone's definition is broad: it covers any case where the allegations relate to AI or AI risk, not only claims that a capability was invented.

UK companies are not outside this. The D&O Diary counted a suit filed in New York against Endava plc at the end of September 2026 as the 28th AI-related securities class action of the year. The allegations concern accounting and controls, but AI sits at the centre of the company's strategy, and that was enough for the case to be counted as AI-related. A company that puts AI at the heart of its equity story invites its disclosures to be read against it.

Insurers are watching. Aon reported on 5 October 2026 that directors' and officers' cover for AI risk "remains broadly available" and that it had not seen insurers impose AI-specific exclusions, but that underwriters now examine "AI governance, controls, oversight, and disclosure practices". In a survey of insurers and brokers reported by the D&O Diary in August 2026, 94 per cent said poor AI governance makes a claim more likely. Expect the questions at the next renewal.

The UK position#

There has been no AI-washing enforcement action or court case in the UK so far. That is not the same as no exposure. Five routes already exist.

Statements to the market. Under Schedule 10A to the Financial Services and Markets Act 2000, a listed company is liable to investors for untrue or misleading statements in information it publishes if a person discharging managerial responsibilities "knew the statement to be untrue or misleading or was reckless" about it. The information need not be required by law, so announcements and presentations published through the usual channels are caught. Directors are liable only to the company, not directly to investors. Prospectus liability now sits in the Public Offers and Admissions to Trading Regulations 2024 rather than section 90, which covers listing particulars only. Separately, UK MAR Article 12 treats spreading information that gives false or misleading signals about a share as market manipulation where the person "knew, or ought to have known" it was false, a lower bar than recklessness.

The annual report. Under section 463 of the Companies Act 2006, a director who knew a statement in the strategic or directors' report was untrue or misleading, or was reckless about it, must compensate the company. And section 174 sets the duty of care, skill and diligence with two limbs: the knowledge reasonably expected of someone in the role, and the director's own knowledge and experience. The second limb only raises the bar, so a director appointed for AI expertise is expected to see through an AI claim that a lay director might miss.

Financial promotions. For FCA-authorised firms, a communication must be "fair, clear and not misleading" (COBS 4.2.1R). I have not found an FCA statement using the term AI-washing; the rule applies to AI claims as to any other. In the EU, ESMA said in May 2024 that information about firms' use of AI must be "presented in a clear, fair and not misleading manner".

Consumers. Since 6 April 2025 the Competition and Markets Authority can fine businesses directly for unfair commercial practices, up to 10 per cent of global turnover. Its guidance of 9 March 2026 on AI agents is the nearest thing to an official UK warning against AI-washing: do not "overstate the role" of AI in a service, "or what it can or cannot do".

Advertising. The Advertising Standards Authority's guidance of 27 August 2026 tells advertisers to "hold evidence to back up any objective claims about what your product can do". On 14 October 2026 it upheld a complaint against an AI sports prediction service that claimed "97% accuracy", finding the ad misleading and irresponsible.

And from 2027 the boards of companies applying the UK Corporate Governance Code will make their first declarations under Provision 29 of the UK Corporate Governance Code on the effectiveness of material financial, operational, reporting and compliance controls. Controls over what the company says about AI in its reporting fall naturally within that scope, although the FRC has not said so in terms. See Provision 29 and AI and directors' duties and AI.

Why the pressure to overclaim is rising#

FactSet counted 331 of 493 S&P 500 earnings calls citing AI in the quarter to mid-September 2026, 67 per cent, the third quarter running above 65 per cent. Aon, citing FactSet data, reported that companies discussing AI had seen share prices rise 12.7 per cent on average since the end of March 2026, against 2.6 per cent for those that did not. In the UK, Deloitte's review of FTSE 100 annual reports, as summarised by the Management Consultancies Association, found 93 mentioning AI in 2025 against 50 in 2023. When the market rewards the word, every incentive points towards saying more than can be shown.

What AI-washing looks like#

The cases so far fall into five patterns, and each one has a version inside ordinary companies.

Capability that does not exist. Delphia's claimed use of AI on client data. Internally: an "AI-powered" product label on a rules engine.

People presented as automation. Nate's call centre, Presto's human order-takers. Internally: describing a process as automated when staff check and correct most of its output. This is the mirror image of meaningful human involvement: a company may claim more automation to investors than it admits to a regulator, and both claims cannot be true.

Undisclosed dependence. Presto did not say its core technology was a third party's. Internally: proprietary AI that is a thin layer on a vendor model, with the concentration risk unstated. See what it would take to leave.

Accuracy that cannot be shown. The ASA's "97% accuracy". Internally: performance figures from a pilot quoted as if they described production.

Adoption as performance. Strategy language, licence counts and "AI-first" slogans standing in for evidence of use and value. This is what I call usage theatre, or AI pretend: the gap between what an organisation performs about AI and what it operates. Usage theatre aimed at investors is AI-washing. See also is AI-first a strategy or a slogan?

What a board should do#

Keep a register of AI claims. Every statement the company makes about its AI, in the annual report, results presentations, investor days, the website, marketing and sales material, listed in one place with the owner of each.

Hold the evidence before the claim. The ASA's standard works for every audience: for each claim, the evidence that supports it, dated, and the person who checked it. A claim nobody can evidence comes out.

Bring AI claims inside disclosure and reporting controls. Treat them as part of the Provision 29 work: who drafts them, who verifies them, and how the board knows the verification happened. See how a board knows management's claims about AI are true.

Make the investor story and the operating reality match. If the company tells investors a process is automated and tells regulators a human reviews every decision, one of those statements is wrong. Reconcile them.

Prepare for the D&O renewal. Underwriters are asking about AI governance, controls and disclosure. A board with a claims register and evidence file has an answer.

Check the board's own competence. Section 174 judges directors partly by what they know. A board that cannot test an AI claim cannot say it took reasonable care over it. See what a non-executive director needs to know about AI.

Questions for the next board meeting#

What are we saying about AI to investors, customers and staff, and is it all in one place? For each claim, what is the evidence, and who checked it? Where do we describe something as automated that people still check or correct? Does anything we tell investors about automation contradict what we tell regulators about human review? Which of our AI depends on a third party we have not named? And if a short-seller read our last annual report against our operations tomorrow, which sentence would they quote?

Where this sits in my own argument#

AI-washing is drift reaching the outside world. Inside a drifting organisation, AI language spreads faster than AI practice: a pilot becomes a capability in a slide, a capability becomes a strategy in a speech, and the strategy becomes a claim in the annual report, without anyone deciding that it should. The first of my four receipts, Hand, asks who stands behind a decision and will answer for it. Applied to disclosure, it asks who stands behind each AI claim. If nobody can, the claim is AI-washing whether or not anyone meant to mislead. The wider position of this research applies too: what a board controls is which decisions machines make, who can stop each one, what people must remain able to do, and how anyone would know if it went wrong. A company that can answer those four questions has little to wash.

I have seen the other side of this line. When I was building EtonX, copycats appeared across Asia within months, cheaper and plainer and sometimes under names so close to ours you had to read them twice. Copying the idea felt like part of the game. What crossed the line was when they started lifting our actual content. I wrote about where that line sits in Winning Ugly: "Principled innovation asks you to decide where your line is before the game starts, while you are calm, nobody is watching and the scoreboard still says nil-nil. If you draw it afterwards, you will draw it wherever you happen to be standing." AI claims work the same way. A board that decides what it will and will not say about AI before the results call will say less it has to take back.

What this does not show#

This page is not legal advice, and directors should take advice on their own disclosures. There is no UK AI-washing case, so how UK courts and regulators will treat AI claims is untested. The US enforcement record is small: four matters described here, one settled with no penalty and one at the allegation stage, and no new SEC case found in 2026. Cornerstone's AI-related category is broader than AI-washing, and the D&O Diary's count uses its own definition. Gensler's December 2023 remark is reported by secondary sources rather than an SEC transcript. The FTSE 100 figures come from a summary of Deloitte's review rather than the review itself, and the share price comparison is Aon's reading of FactSet data and shows correlation, not cause. The inference that AI claims fall within Provision 29 reporting controls is mine, not the FRC's.