35% of Startups Selling AI Do Not Actually Have Any

More than a third of the startups reviewed in Europe had no AI in them at all. Some were wrapping someone else's ChatGPT API and selling it back as their own proprietary product. That number did not come from a marketing article. The Trustworthy AI Lab at the University of Hamburg audited 100 German AI startups against a seven-dimension codebook. They also ran 63 interviews with investors, researchers, and the people doing the selling. The results went up on arXiv on October 8, 2026.

What they found: 35% of the startups had no AI whatsoever. Another chunk had nothing but a large language model wrapper, with no proprietary contribution. A total of 60% marketed AI prominently without owning proprietary AI. And 79% of the startups making numeric claims disclosed no methodology at all. And 38% were flagged for moderate or serious ethical concern. Every one of those sat in a domain the EU AI Act classifies as high-risk, or prohibits outright.

The startups are not the problem here. The problem is how cheap it has become to be believed.

The Same Pattern Since 2019, Only Bigger Now

This is not a new disease. In 2019, a venture-backed study cited by Forbes found something similar. Around 40% of European companies registered as "AI startups" showed no evidence of using AI in their products. The number got a large volume of retweets and a wave of articles. Then nothing changed.

Seven years later the pattern is back, but the stakes are far larger. What changed was not the frequency. It was the cost of being believed. In 2019 a startup could still bootstrap on a laptop. Now a single "AI-powered" claim can attract a large funding round. The market can accept that claim without proof. Same motive, different stakes.

A direct comparison between the two numbers is not possible, because the methods differ. The 2019 study looked at registered "AI startups" and asked whether their products contained AI. The 2026 audit scored marketing copy against a graded codebook, then interviewed the people writing it. What can be said is that the pattern held. The tooling for measuring it also got much better in between.

Why a Claim Can Explode Without Anyone Noticing

The mechanism is simple, and that is what makes it hard to stop. Investors reward branding they cannot verify. Consumers treat AI as a near-magical capability. Those two things meet in one place: the marketing page.

The strongest evidence comes from correlation rather than assumption. Across the whole codebook, the strongest pairwise relationship ran between two things. How substantive the underlying AI was, and how much weight AI got in the marketing. The correlation was +0.51, with p below 0.001. In plain terms, startups whose product is a third-party wrapper overstate what the AI does for them far more often.

The reason is straightforward. The thing being wrapped belongs to someone else. If you do not own a model, you cannot show a model. All you can show is a claim. A wrapper is like repackaging someone else's coffee beans in your own bag and charging five times the price. The contents are identical. Only the label changed.

And a claim nobody can check always wins a marketing contest. There is a related pattern in the numbers. Among the 39 startups with the highest composite severity score, 28 had no proprietary AI at all. No model of their own, nothing beyond a wrapper or a combination of existing tools. The wrapper is not a rounding error in the sample. It is the single largest cluster of overstatement.

What the Audit of 100 Startups Found

LOONIXCHART: Numbers with no method, Claims with no limits, No proprietary AI, Buzzwords only, No inspectable proof, Ethical concern
79
76
60
59
43
38

Two further numbers get less attention than they deserve. First, only 2% of startups offered evidence at the top tier. That is the tier a reader outside the company could actually check. Most of the rest offered client logos, anonymized case studies, or controlled product demos. A logo proves a customer relationship, not what the AI did for that customer. An anonymized case study cannot be checked with the customer who was named. A demo runs on inputs the seller chose.

Second, 76% of startups named no limitations anywhere in their material. Two went further and promised systems that never fail. A field sales tool promised to think ahead, learn from corrections, and never drop the ball. A knowledge management platform invited employees to ask anything, and receive a perfect validated answer instantly.

Three forms of exaggeration recur across the sample. Capability claims with no inspectable evidence behind them. Performance numbers with no disclosed methodology. Claims that acknowledge no limits at all. All three make a system sound more capable than it is.

How to Tell Whether a Tool Actually Uses AI

This is the most usable part of the audit, and the part almost nobody writes down. Five questions fall out of the codebook. You can apply them to any product page before you pay.

First, is a model named anywhere? If no model name and no technical description appears, it is probably an API wrapper with a UI on top.

Second, does the number have a baseline? "Saves up to 96% of production costs" is meaningless without a baseline. What it is compared to, which industry, and over what period. That is branding, not a measurement.

Third, is there a section explaining the limits? Honest tools almost always carry a note about what they cannot do. Tools that do not have one have probably never been tested against the awkward cases.

Fourth, is the AI weighted more heavily in the copy than in the product? If 80% of the page covers non-AI features while one small chat box says AI-powered, marketing weight beats product weight. This was the single strongest correlate of the severity score in the study.

Fifth, can an outsider audit the proof? Client logos are not proof. Anonymized case studies are not proof. A demo on seller-selected inputs is not proof. What counts is something a reader can check without the seller's cooperation.

None of this is theoretical. It came from measuring 100 real companies, and it applies to any tool. Including the ones you are using right now.

Regulators Were Six Months Late, and What Filled the Gap Was Not Regulators

This is the most ironic part of the story. When AI claims started multiplying, the FTC opened Operation AI Comply in September 2024, with more than a dozen cases. The SEC charged two investment advisers over false statements about their use of AI in 2024.

Then, in the first half of 2026, the SEC brought zero cases focused on AI claims. Six months of nothing. The new cases only arrived in August and September 2026, three of them.

What filled the gap turned out to be two things, and neither was a regulator.

The first is private litigation. Class action securities suits alleging AI washing against public companies have been filed this year. They name officers and directors, not just the company. That pressure lives outside the SEC docket, and it moves faster than the agency does.

The second is self-regulation. The BBB National Programs National Advertising Division is not a regulator. Its decisions still become a public record that plaintiff's counsel cites. The NAD has consistently flagged three patterns. Availability claims for technology that has not shipped. Demonstrations that mislead. Productivity claims resting on user perception rather than measurable data.

The FTC has also changed direction. A growing share of its cases now target business-to-business claims rather than consumer advertising. The agency has revived the means and instrumentalities doctrine. Vendors who hand a client deceptive marketing materials are now liable too. Not only the company making the claim.

Why Credibility Becomes the New Currency

One finding from the 63 interviews matters more than the rest. People who had previously encountered an exaggerated AI claim trusted the next piece of AI marketing less. Not marginally less. It is a market-clearing mechanism. It is also the only part of the study suggesting the problem fixes itself.

Picture the effect. If everyone becomes more suspicious every time they see the phrase AI-powered, the most outrageous claims lose their value fastest. And outrageous claims are usually the easiest to copy. A wrapper with one prompt takes an afternoon to reproduce. What takes years to build is the slowest thing on the market to displace. That means an owned model, with owned data and owned evaluation.

This kind of structural shift has happened before, in markets with nothing to do with AI. In music, once streaming made raw listening volume worthless, a label with genuine artists became more valuable. In ecommerce, a seller with authentic reviews holds an asset a seller who bought reviews can never acquire.

What is not settled is how far this travels. The audit covers German startups only. The people interviewed were AI-literate professionals, the group best equipped to check a claim. For a general public with lower AI literacy, the effect could be larger. Nobody has measured that yet.

What This Means If You Are About to Buy an AI Tool

Put the audit to work and it becomes a buying checklist rather than a scandal. Four red flags cover most of what the study found. Each one has a concrete tell on the page in front of you.

The cheapest test is the first row. Open the pricing page, open the security page, and look for a model name. If there is none, the thing is a front end. That does not make it useless, and it is not fraud. It just means you are paying a margin for someone else's inference, and you should know that going in.

The second cheapest test costs one written question to the vendor. Which model, which version, and what did you measure the 96% against? Vendors with a real product answer in a day. Vendors with a marketing problem go quiet. Silence is itself a data point, and it costs nothing to collect.

One more thing worth remembering for anyone in a market where buyers cannot easily verify. The study found that AI-literate professionals were the ones best able to spot the gap. Where buyers are individuals or small businesses without the time to audit a vendor's model, the asymmetry is wider. The discount on a wrapper should be larger there, not smaller.

What Is Still Open

Three things remain unresolved, and none should be answered with a guess.

Whether the gap closes on its own. The interviews show that buyers burned by exaggerated claims stop trusting the next seller. There is no data yet on whether the market will reward accuracy in the next two years. Or keep rewarding the boldest marketing instead.

Whether disclosure becomes mandatory. If AI labels have to be documented, the gap could close from the regulatory side. Which components were used, how performance was measured, and what the limits are. What is happening now leans on private litigation, which moves slowly.

Whether AI washing that works at the startup level spreads to consumer products. This audit focused on business-to-business, where claims are harder to enforce. Buyers are assumed sophisticated enough to check for themselves. Consumer claims are easier to police, and much less often audited.

Sources

• arXiv:2610.11788, Auditing AI-Washing in German Startups, Trustworthy AI Lab, University of Hamburg, October 8, 2026
• Holland & Knight, Operation AI Comply 2 Years Later, August 18, 2026
• DLA Piper, Three new SEC cases show continued scrutiny of AI-washing claims, October 2026
• SEC Press Release 2024-36, Charges Two Investment Advisers with False AI Statements
• Forbes, Nearly Half Of All AI Startups Are Cashing In On Hype, March 4, 2019