AI Disclosure to Customers: What the Law Requires, and What the Trust Data Says

Disclosing AI use to customers is now a legal duty in the European Union, not a brand choice. Article 50 of the EU AI Act applied from 2 August 2026, and fines reach €15 million or 3% of worldwide turnover. The harder question is what the law leaves open. AI disclosure carries a measured cost, and hiding it carries a larger one.

The legal floor is lower and narrower than most policies assume

If you sell into the European Union, you must tell customers when they are interacting with an AI system. You must also mark synthetic audio, image, video and text. Outside the EU, disclosure is mandatory only in narrow cases. Everywhere else it is a commercial decision, not a compliance one.

That distinction gets lost fast. Most internal AI policies I have read treat disclosure as one undifferentiated duty. The statutes do not. They target four specific situations, and a company can be fully compliant while saying nothing about how a report was drafted.

What Article 50 actually catches

The European Commission's own guidance splits the duty between providers and deployers. Providers must design systems that interact directly with people so those people know they are talking to an AI system, unless the interaction is obvious to a reasonably observant person. Providers of generative systems must also mark outputs in a machine-readable format.

Deployers carry a shorter list. They must label deepfakes, disclose emotion recognition and biometric categorisation, and label AI-generated text published to inform the public on matters of public interest.

The exemptions matter more than the rules. Marking is not required for short sequences of symbols, source code, machine-to-machine output, closed-loop industrial use, or standard editing assistance. Published text that went through human review or editorial control is outside the labelling duty. That last carve-out is the one most B2B teams will rely on, and it is the reason a designed review step is worth more than a disclaimer. The design question behind it is covered in the piece on building the review loop before you buy the agent.

The United States is disclose-on-request, not label-everything

There is no federal AI disclosure statute. The state rules that exist are aimed at different targets, and two are worth knowing.

California's AI Transparency Act was moved by AB 853, signed in October 2025. The operative date shifted from 1 January 2026 to 2 August 2026, and the scope widened to large online platforms, model hosting platforms and camera manufacturers on later dates. It binds providers of generative systems with more than a million monthly users reachable in California. It does not bind the average software vendor.

Utah took the opposite approach. Under SB 226, effective 7 May 2025, a supplier using generative AI in a consumer transaction must say so only when the customer makes a clear and unambiguous request. Proactive disclosure is required for regulated occupations in high-risk interactions, meaning health, financial and biometric data, or advice on financial, legal and healthcare matters. There is a safe harbour if the system itself says it is not human at the outset and throughout.

Spain approved a draft bill in March 2025 proposing fines of up to €35 million or 7% of global turnover for unlabelled AI content, supervised by AESIA. I could not confirm its final enactment status, so it is excluded from the table and should not be planned against yet.

Above the floor, your buyers set the rule

Compliance answers a small question. The commercial question is what a procurement team will ask you, and that has moved faster than the statutes.

Enterprise buyers now route AI questions through the same channel as security questions. The Cloud Security Alliance publishes an AI extension to its consensus assessment questionnaire, mapped to an AI controls matrix. Vendors submit it to the same registry that already holds their security self-assessments. Once a question sits in a standard questionnaire, answering it is not optional in any practical sense.

The contract is where disclosure becomes binding

A marketing page can be vague. A master services agreement cannot. Buyers are asking for named clauses covering model change notice, sub-processor disclosure, training-data warranties and whether prompts and outputs count as customer data. Those clauses convert a soft transparency posture into an enforceable one, which is a different risk profile. The negotiation detail is set out in the breakdown of the clauses a CFO should insist on, and the remedies question in the piece on liability caps for agent errors.

My position is that inconsistency here is the real risk, not disclosure itself. A vendor whose sales deck, questionnaire answer and MSA describe three different levels of AI involvement has created a documented misrepresentation. That is worse than any honest answer would have been.

Your own staff have already answered the question

The KPMG and University of Melbourne study surveyed more than 48,000 people across 47 countries between November 2024 and January 2025. The employment findings are the ones that should worry a founder. 60% of employees said they use AI intentionally at work. 57% said they hide that use and present AI-generated work as their own. 48% admitted using AI against company policy.

Read those three numbers together and the disclosure policy problem inverts. You are not deciding whether to tell customers. You are deciding whether you know enough about your own delivery to make a truthful statement at all. That measurement gap is the same one behind the cost of undisclosed AI inside the organisation. It shows up again in the gap between how much engineers use these tools and how much they trust them.

Disclosure has a measured price, and it is not trivial

The advice that transparency always builds trust is comfortable and wrong. Two controlled studies point the other way, and both used decent samples.

Naming AI lowers purchase intent

Researchers at Washington State University and Temple University ran experimental surveys with more than 1,000 US adults across eight product and service categories. The work was published in the Journal of Hospitality Marketing & Management in July 2024. Descriptions were identical apart from the phrase "artificial intelligence". The group that saw it was less likely to buy.

Lead author Mesut Cicek framed the mechanism plainly. Mentioning AI lowers emotional trust, and lower emotional trust reduces purchase intent. The effect was stronger for high-risk categories such as expensive electronics, medical devices and financial services, where failure costs the buyer money or safety. The release did not publish effect sizes, so treat the direction as established and the magnitude as unknown.

The person who used the tool takes a hit too

A 2025 paper in the Proceedings of the National Academy of Sciences ran four experiments with over 4,400 participants. People who disclosed using AI were judged as less competent and less motivated by their evaluators. The authors found the penalty was both anticipated by users and real in how evaluators actually rated them.

That result explains the 57% who hide it far better than any story about laziness. Employees are responding rationally to a penalty that exists. If you want honest internal reporting, you have to remove the penalty first, and a policy memo will not do that on its own.

What silence costs when it fails

Set against that, the case for disclosure rests on asymmetry. The cost of disclosing is a small, continuous discount on trust. The cost of concealment is rare, large and public.

Deloitte Australia produced a report on the welfare compliance system for the Department of Employment and Workplace Relations under a contract worth roughly A$439,000. Academics found citations to studies that did not exist and a misattributed legal quote. The firm republished the report in September 2025 with a disclosure that a large language model had been used. It also agreed to refund the final instalment of A$97,000, confirmed by the department in October 2025.

Look at what the refund actually priced. It was not the AI use. It was the combination of unverified output and the absence of a disclosure at delivery. Had the methodology note been in the first version, the story is a quality complaint. Without it, the story is a credibility failure that ran in national press for two months.

Three postures, and the one most companies drift into

Every company lands in one of three positions. Naming yours is more useful than debating principles.

The on-request model is the worst of the three, and it is the one most companies reach by default. It carries the trust cost of eventual discovery without the operational discipline of a written policy. It also guarantees the inconsistency that turns an awkward answer into a misrepresentation claim.

When disclosure is genuinely a differentiator

Some vendors market their transparency. Most of that is theatre, but there are two situations where it converts.

The first is a regulated buyer. If your customer is a bank, an insurer or a public body, they will be asked how their supply chain uses AI. A vendor who hands over a complete answer removes work from their procurement team. That is a real purchase reason, and it is worth more than any trust messaging aimed at end users.

The second is a formal signal. The EU published a Code of Practice on Transparency of AI-generated Content in June 2026. The Commission and the AI Board confirmed it as an adequate route to demonstrating Article 50 compliance. Around 190 organisations signed before the obligations took effect, roughly half of them small and recent companies. Signatories get a presumption of conformity and a lighter enforcement posture. Non-signatories have to demonstrate compliance some other way.

Treating a general disclosure page as a trust asset is mostly wishful thinking, because almost no buyer reads it. A named signature on a recognised code is different. It is checkable, it is dated, and a procurement team can verify it without asking you.

Where this argument is weakest

Three places, and the first is the one that undercuts most of the section above.

Nobody has measured this in a B2B sale

Both experimental studies I cited test consumer framing. One used product descriptions, the other used evaluations of individuals. Neither tested what happens when a software vendor discloses AI use in a proposal to a procurement committee. I have extended a consumer finding into a business buying context, and that extension is an assumption, not evidence.

There is a decent argument it runs the other way in B2B. Professional buyers have their own AI mandates and may read disclosure as competence rather than as a shortcut. I would not bet either way without data.

The trust penalty may be decaying

The purchase intent study ran before mid-2024. Everyday exposure to these tools has risen sharply since. A penalty grounded in unfamiliarity should shrink as familiarity grows, and a 2024 result may already overstate the 2026 effect. Any post quoting these findings in 2027 without re-checking them, including this one, should be treated with suspicion.

Writing the policy so it survives contact with a buyer

The useful output is not a public statement. It is an internal map of which surfaces carry which duty, agreed once, so that three different people give the same answer.

The row I expect argument about is marketing copy. Volume publishing without disclosure is a live reputational question, and the search side of it is covered in the analysis of what content saturation does to visibility. My view is that a label on every post is noise, and that the accountable version is a named author who checked the facts. The EU drew the same line by exempting text under editorial control.

One more test before you publish anything. If your AI use would embarrass you when a customer discovered it, disclosure is not your real problem. Fix the use, then decide what to say about it.

Frequently asked questions

Where to start on Monday

Two tasks, both small, both prerequisites to any public statement.

First, find out what your team actually does. Ask each function which AI tools touch customer-facing output, and write the answers in one place. Given that 57% of employees globally say they hide this, run the question as an amnesty rather than an audit, or you will collect a fiction.

Second, write one paragraph and put it in three places: the security questionnaire response, the MSA, and the sales deck. Same words in all three. That single act removes the inconsistency that turns an honest position into a misrepresentation, and it costs an afternoon.

Related reading