Why Every Company That Claims to 'Do AI' Is Lying to You in Exactly the Same Way

In the spring of 2019, a British startup called MMC Ventures published a study examining 2,830 European AI startups. The researchers applied a specific test: they looked at whether the companies actually used artificial intelligence in the way that was material to their products or value propositions. The result was striking. Forty percent of the companies studied could not be verified to use AI in any meaningful sense. They had the word “AI” in their marketing materials, their pitch decks, and their investor communications, but the underlying products didn’t use machine learning, neural networks, or any other technique that would qualify as AI in any meaningful technical sense.

The study made headlines briefly, and then the AI industry continued in exactly the same direction, at greater speed.

This is worth examining carefully, because the lie has a specific structure — and specific consequences — that make it more than just a marketing problem.

The anatomy of AI-washing follows a consistent pattern, once you know what to look for. The first form is the simplest: rule-based systems called AI. A company that applies fixed business rules to route customer service tickets — “if the email contains the word ‘refund,’ send it to the billing team” — is using automation. If that company describes itself as “using AI to optimize customer service workflows,” it is misleading anyone who takes the claim at face value. The distinction matters because rule-based systems and machine learning systems have fundamentally different properties: the former require humans to specify every decision rule; the latter learn decision patterns from data. They have different costs, different maintenance requirements, different failure modes, and different scalability characteristics. Calling a rule-based system “AI” doesn’t just misrepresent the technology; it misrepresents the company’s actual capabilities and limitations.

The second form is more sophisticated: using AI tools is not the same as being an AI company. A startup that wraps GPT-4 in a user interface, adds a vertical-specific prompt, and sells it to HR departments as an “AI-powered hiring tool” is using AI — but it is not an AI company in any meaningful sense. Its differentiation comes from the user interface, the go-to-market approach, and the customer relationships, not from any AI capability it has developed. When the OpenAI API becomes more expensive, or when a competitor releases a better API, the company’s “AI capability” changes overnight. This is a fundamentally different business than one that has trained its own models on proprietary data and has genuine AI research as a source of competitive advantage.

The third form is the most dangerous for investors and acquirers: retrofitting AI into existing products without changing the underlying logic. A company notices that “AI” commands a valuation premium — the data is clear; AI-labeled companies raised money at higher multiples than comparable non-AI companies throughout the early 2020s — and adds an AI feature to its existing product. The feature might be genuinely useful and might genuinely involve machine learning. But when the company’s core value proposition, the reason customers pay it money, has nothing to do with that AI feature, describing the company as an AI company is misleading in a way that distorts capital allocation.

The historical parallel with the 1999 dot-com era is instructive but imprecise. During the internet bubble, established companies added “.com” to their names or announced internet strategies to capture the market’s enthusiasm for internet business models. Brick-and-mortar companies rebranded as e-commerce plays. The mismatch between the valuation implied by internet-company pricing and the underlying business fundamentals became apparent when the bubble corrected.

The AI analog is similar in structure but different in one important respect: AI is genuinely more technically complex than having a website, which makes the washing harder to detect. In 1999, it was relatively easy to check whether a company was actually selling online. In 2026, checking whether a company actually uses AI in a way that constitutes a genuine competitive advantage requires technical due diligence that most investors, journalists, and customers are not equipped to perform.

The specific lie — and it is worth being precise here — is not usually an outright falsehood. Companies don’t typically claim to have built AI capabilities they don’t have. The lie operates at the level of implication: positioning that allows audiences to draw incorrect inferences from technically true statements. “We use artificial intelligence to improve our recommendations” is technically true if you apply any form of statistical model to your recommendation engine, but it implies technical sophistication and competitive moats that a simple collaborative filter does not provide. “Our platform is powered by advanced machine learning algorithms” sounds impressive and may be true in a minimal sense, while describing something that any junior data scientist could replicate in a week.

The consequences of this misrepresentation are distributed across different stakeholders. For investors, the consequence is capital misallocation: money flows to companies whose AI claims imply competitive moats that don’t exist, depriving genuinely differentiated companies of capital and investors of returns. For employees, particularly engineers, the consequence is disillusionment: people join “AI companies” expecting to work on interesting technical problems and find themselves maintaining dashboards or cleaning data for models built entirely by a third-party API. For customers, the consequence is misplaced trust: they buy products under the assumption that the AI label implies a level of sophistication and reliability that the underlying technology doesn’t support.

How do you identify real AI integration versus marketing claims? Several tests are revealing. The first is the data question: does the company have proprietary data that would allow it to train or fine-tune models better than its competitors? Companies whose AI depends entirely on third-party APIs have no data moat. Companies with years of proprietary behavioral, transactional, or domain-specific data might have a genuine AI advantage. The second is the model question: does the company employ AI researchers or ML engineers who are actually building and maintaining models, or does its “AI team” consist of product managers and engineers who are primarily working on user interfaces and API integrations? The third is the switching cost question: if OpenAI or Google raised their API prices by 10x, would the company’s core value proposition survive? If not, its AI capability is a dependency rather than a competitive advantage.

The correction, when it comes, will probably not look like the 2000 dot-com crash, which was dramatic and rapid. It will more likely be gradual as the market develops better instruments for evaluating AI claims. This is already happening in sophisticated investor circles, where due diligence processes now routinely include technical AI audits. The companies that will be most exposed are those in the middle — not pure rule-based-system frauds (those will simply fail to grow), but companies that have genuine software businesses but have raised money on AI valuations they don’t deserve.

There is a harder question lurking beneath the AI-washing problem, which is whether the distinction between “using AI tools” and “being an AI company” matters as much as the analysis implies. Some observers argue that in a world where AI capabilities are increasingly available as commodity infrastructure, the competitive advantage lies in application, distribution, and market understanding rather than in the AI itself — exactly as competitive advantage in internet businesses eventually lay in brand, network effects, and execution rather than in “being an internet company” versus using the internet. On this view, the AI-washing critique is making the same mistake as people in 1999 who thought that having a website was a competitive differentiator: the technology itself commoditizes, and the sustainable advantage moves elsewhere.

This is a reasonable argument, but it requires the companies making it to be honest about what they are actually differentiating on — and to price accordingly. A company that is fundamentally a vertical software business that uses AI tools, pricing at a premium justified by “AI capabilities” it doesn’t actually possess, is still misrepresenting itself. The argument that AI will commoditize doesn’t justify claiming AI differentiation today.

The tell, for anyone learning to read AI claims, is specificity. Genuine AI companies talk about their AI in specific ways: the data they have, the problems they’ve solved, the benchmarks they’ve beaten, the limitations they’ve discovered. Companies performing AI-washing tend to talk about AI in general, abstract, and non-falsifiable ways. “Powered by AI.” “AI-first platform.” “Machine learning at the core.” Sentences that could be appended to almost any technology product without modification should trigger skepticism. Sentences that describe specific capabilities with specific evidence should not. It is a surprisingly effective filter, and it is almost never applied by people who should know better.

The regulatory response to AI-washing is still nascent, but it is developing. The Securities and Exchange Commission issued guidance in 2023 warning that “AI washing” — making materially misleading statements about AI capabilities to investors — would be treated as fraud under existing securities law. The FTC has similarly indicated that consumer-facing AI claims are subject to existing truth-in-advertising standards. These are meaningful signals, but enforcement has been sparse. The market correction, when it comes, will likely be more consequential than regulatory action: as enterprise buyers develop more sophisticated evaluation criteria, the gap between claimed and actual AI capability will become a commercial liability rather than a marketing asset.

The employees of AI-washing companies face their own version of this problem. Engineers hired to work on “our core AI platform” who discover they are maintaining API integrations and building dashboards leave quickly, and their departure concentrates the AI talent in companies that are actually doing what they say. This talent sorting is a market mechanism that, over time, pushes genuine AI capability toward companies that have it and away from companies that merely claim it. The correction may be slow, and it will certainly be uneven, but the fundamental dynamic is one where misrepresentation becomes increasingly costly as the people best positioned to evaluate the claims — technical talent in the labor market — become more discerning. The specific lie of AI-washing works only as long as the audience cannot tell the difference. That audience is learning.