AI Washing: How to Tell If a Vendor Is Faking It

Every software product acquired an AI badge somewhere around 2024, and a fair number of them acquired nothing else. Regulators noticed. The FTC launched Operation AI Comply in September 2024, and according to a Holland & Knight review published this August, initiated more than a dozen AI washing cases in the following year. For a business shopping for AI tools, the useful part is not the enforcement. It is that the regulators had to develop a way of telling real capability from a claim, and you can borrow it.

The four patterns

A rules engine wearing a badge. Conditional logic written by a developer in 2019, relabelled. It works well within its configured cases, which is why demos go smoothly, and it fails hard on anything outside them rather than degrading. This is the most common version and often the least deliberate, since marketing renamed the feature and nobody in engineering was asked.

Proprietary claims over somebody else's API. The product calls a frontier model and the marketing implies in-house research. The wrapper is fine. Most good AI products are wrappers with excellent workflow around them. The claim of ownership is the problem, and it matters commercially: if the underlying provider raises prices or deprecates a model, the vendor's roadmap moves whether they admit the dependency or not.

Roadmap sold as present tense. The feature exists in a design document and a slide. Ask when you can use it in your account, in writing, and the tense in the answer changes.

Accuracy without a denominator. "99% accurate" with no test set, no date, no description of what counted as correct. Real evaluation numbers come with all three because the people who ran them know the numbers mean nothing otherwise, which is the same problem we went into in why AI benchmarks mislead.

The enforcement record

These are US regulators and none of it binds a Canadian vendor directly. In Canada, misleading representations sit under the Competition Act, and this is not legal advice. What the record establishes is that unverifiable AI claims are being treated as ordinary deceptive marketing rather than as puffery, which is a reasonable thing to say out loud in a procurement conversation.

Two developments that matter to buyers

The FTC has extended this scrutiny to business-to-business marketing rather than confining it to consumer claims. Pitches made to you, in a boardroom, are inside the same frame as advertising to the public.

More consequentially for anyone who resells, the agency has revived the means and instrumentalities doctrine, under which a company that supplies another with the tools to deceive, including marketing materials, sales pitches and prepared answers to customer questions, can face liability itself. If you white-label an AI product or repeat a vendor's claims to your own clients under your own name, their marketing becomes your marketing.

Seven questions that settle it

1. Which models does this use? A plain answer naming OpenAI, Anthropic, Google or an open-weight model is the healthy response. Deflection is the signal, not the dependency.

2. What is the accuracy figure, on what test set, measured when? All three parts. A number alone means nothing and the vendor knows it.

3. What happens on an input outside your examples? A model produces something imperfect. A rules engine produces an error or silence. The answer tells you which one you are buying.

4. Can I run my own documents through it during the trial? Yours, including the awkward ones. A refusal here is worth more information than any demo.

5. Show me a case where it got it wrong. Every real system has these and good vendors discuss them comfortably. A vendor who cannot produce one has either not looked or will not say.

6. Is this shipped or planned? Ask feature by feature, and get it in the contract rather than the deck.

7. Does it complete the task or assist with it? The distinction that decides whether you can change how work is staffed, covered in only 2.6% of agent tools finish a whole task.

What is not AI washing

Fairness matters here, because the accusation is easy to throw and expensive to a small vendor. Calling an OpenAI API is not fraud. Fine-tuning someone else's open-weight model is not fraud. A product that combines a model with conventional software and good workflow design is usually a better product than one that tries to build its own model, and the workflow is often where the value actually sits.

Nor is every rules engine a deception. Plenty of business problems are better solved by explicit rules, which are cheaper, faster, auditable and predictable. A vendor who says "this part is rules, this part is a model, here is why" is showing you good engineering rather than a weakness.

The thing to test is the gap between what is claimed and what runs. Our notes on making AI vendors prove it and vetting vendor claims on usage limits cover the rest of the diligence, and who checks the AI vendors covers who is doing this work on your behalf, which is mostly nobody.

If you sell AI yourself

Plenty of Canadian firms now put AI in their own marketing, and the safe position is straightforward. Describe what the software does rather than what technique it uses. Name the models you build on. Put a date and a denominator next to any number. Label roadmap items as roadmap. Customers who understand this space read plain description as confidence, and everyone else gets a claim you can still defend in two years.

Frequently Asked Questions