AI washing

The term is modeled on greenwashing. It covers ad copy, sales pitches, pitch decks, press releases and securities filings that claim AI where none is at work, overstate what a system can do, or hide that the supposedly in-house technology comes from a third party or that people are doing the work behind the scenes. The range runs from marketing puffery to false statements of fact made to customers or investors.

The term gained traction through US enforcement. In March 2024 the Securities and Exchange Commission fined two investment advisers for advertising AI use that did not exist as described; law firms describe these as the first cases of their kind. In September 2024 the Federal Trade Commission bundled several cases against deceptive AI claims. Germany has no dedicated AI washing law; misleading AI claims fall under general unfair competition law, mainly Section 5 of the Act Against Unfair Competition (UWG).

For executives, the term cuts both ways. On the buying side, it decides whether a tool does what the sales team promised and whether budget goes into a label instead of a capability. On the selling side, for your own products, proposals and investor reports, leadership carries the risk when an AI claim cannot be backed up.

Where the term comes from and the cases that shaped it.

Since generative language models became widely available, "AI" has been a selling point. The label signals progress, supports higher prices and can prop up a startup's valuation. That made generous AI labels tempting, and regulators noticed.

On March 18, 2024, the SEC announced settlements with the investment advisers Delphia and Global Predictions. According to the SEC's findings, both had marketed AI use that did not exist as described. They paid a combined $400,000 without admitting or denying the findings. In January 2025 came a case against a public company: according to the SEC, Presto Automation presented drive-thru speech recognition as its own technology when, for a period, all of it was owned and operated by a third party.

In September 2024 the FTC grouped several actions under the name "Operation AI Comply," including one against DoNotPay. The company had marketed a "robot lawyer" without testing whether its output matched a lawyer's work. In April 2025 federal prosecutors in New York charged the founder of the shopping app Nate. The purchases the app described as automated were allegedly completed largely by people in call centers. He is presumed innocent unless proven guilty.

The forms AI washing takes.

The most common form is the plain false claim: a product is marketed as AI-powered but runs on fixed if-then rules, simple formulas or manual work. This often involves conventional automation such as robotic process automation (RPA), filter rules or text templates.

With exaggeration, there is AI, but it does less than promised. Text recognition that only works on clean templates gets sold as fully automatic, or an accuracy rate comes from a lab test on hand-picked data.

Another variant concerns origin. A vendor presents a technology as its own while calling a third-party model through an API in the background. That alone is not wrong; plenty of good products work this way. It becomes misleading when origin matters to the purchase decision or to a company's valuation and is misrepresented.

Harder to spot is hidden manual work: people do what the machine supposedly does, or correct every result, without disclosure. Human review is often a good idea. It then belongs in the product description.

The UK Advertising Standards Authority adds one more case in its guidance on AI as a marketing term: claims that an AI product beats conventional alternatives need solid evidence.

What counts as AI and what doesn't.

Whether an AI claim holds up depends on what AI means. There is no binding definition for advertising. The EU AI Act offers a usable yardstick. Under Article 3(1), an AI system is a machine-based system that operates with some degree of autonomy and infers from its inputs how to generate outputs such as predictions, content, recommendations or decisions.

The European Commission published guidelines on this definition on February 6, 2025. The deciding factor is the ability to infer outputs, for example through machine learning or through logic- and knowledge-based approaches. Software that automatically executes rules defined solely by people does not qualify. Neither does basic data processing, nor predictions based on a simple statistical rule such as an average.

For advertising law, that definition is not directly controlling. What counts is how the target audience understands a claim. A buyer who reads "AI" will usually expect a system that learns or infers on its own, not a set of fixed rules. The AI Act's line is still a useful test.

Many products are mixed. An RPA workflow may use a language model at one step, say to read incoming email. An AI claim is then accurate if it refers to that step, not the whole product.

The legal picture in the US, the EU and Germany.

In the United States, two regulators lead. The FTC treats unsupported AI claims to consumers as deceptive advertising under the FTC Act. The SEC pursues misleading statements to investors, whether in filings, on websites or on social media.

In the EU, the Unfair Commercial Practices Directive bans misleading claims about a product's main characteristics in every member state. Germany implements it in Section 5 of the UWG. The claimed technology can be such a characteristic if it plays a role in the purchase decision. Under Section 8 of the UWG, competitors, trade associations such as the Wettbewerbszentrale, consumer organizations and chambers of commerce can take action. The usual first step is a cease-and-desist letter demanding a binding undertaking backed by a contractual penalty. Injunctions, lawsuits and, for competitors, damages can follow.

With investors, securities law applies on top, for example to prospectuses and mandatory disclosures. A June 2026 report in the German legal journal beck-aktuell noted that, as far as known, no German court had yet ruled on AI washing. Assessments there rely on general case law on misleading advertising. This is not legal advice; specific cases need a lawyer.

How to spot AI washing when buying.

The best protection is a short list of plain questions to the vendor, asked and answered in writing.

The first question is about method: what exactly does the AI do, and with what technology? A model the vendor trained itself, a third-party language model called through an API, or a rule set? A credible answer names the method and the model provider.

The second is about share: which steps run automatically, and which do people handle or check? How often does someone step in, and is that work included in the price?

The third is about evidence: where do quoted accuracy rates or time savings come from, what data were they measured on, and do they hold for your use case?

The fourth is about the baseline: what does the product do without the AI feature, and what does that feature cost on top?

The most telling step is a trial with your own data and your own cases instead of a demo with hand-picked examples. Evasive answers such as "proprietary technology" with no further explanation are a warning sign.

How companies keep their own AI claims defensible.

If you market with AI yourself, every claim needs evidence that will still hold a year from now. A simple approach is a list of every AI claim on your website and in proposals, presentations and investor updates. For each claim, record which feature it refers to, what technology runs behind it and what backs it up, such as measurements, logs or technical documentation.

Before a new claim goes out, the business side and the technical team check together that it is true. For advertising with wide reach, add a legal review. Specific descriptions are safer than buzzwords. "A language model writes the first draft; a specialist reviews it" can be backed up. "Fully AI-driven" usually cannot.

The list needs upkeep. When the product changes, the model provider switches or a feature goes away, the claim has to follow. A review only covers the version reviewed.

Investor communications face a stricter standard. In its cases, the SEC explicitly looked at statements on websites and social media. Whatever goes into a pitch deck or annual report should therefore pass the same review as ad copy.

Where the term stops working

AI washing is about the claim, not the technology: a weak AI product that is described honestly is not AI washing. The term is distinct from the AI Act's transparency duties, such as labeling chatbots and AI-generated content under Article 50, and from greenwashing, which concerns environmental claims. AI washing is not a legal term; what decides a specific case is the ban on misleading advertising.