Being caught costs about twice what disclosing costs. And 'reviewed by a human' buys back nothing.

In the EU, since 2026-08-02, yes... and at the first point of contact, not on request. In Malaysia there is no equivalent statute, so it is your decision rather than the law's.

But the interesting part is not the legal answer. It is that disclosure has a measured cost, being CAUGHT has roughly double that cost, and the reassurance every business reaches for to soften it recovers nothing at all.

01 / The lawThe law, briefly and with its uncertainty marked

European Union. Regulation (EU) 2024/1689, Article 50. The transparency obligations became applicable on 2026-08-02. Providers and deployers of a conversational AI system must disclose that the person is interacting with an AI. The duty attaches to the FUNCTION, not to a risk tier, so an ordinary customer-service chatbot is inside it. Exposure runs up to €15 million or 3% of global turnover.

Two qualifications we will state rather than gloss:

• How far Art. 50 reaches a non-EU business with incidental EU visitors is not settled, and we have not verified it against the text to our own satisfaction. If your customers are all in one non-EU country, do not assume you are in scope, and do not assume you are out of it either.
• Enforcement posture in year one is unknown. The obligation is live; what a national market-surveillance authority does about a small business is not yet observable.

Malaysia. There is no AI-disclosure statute. What the PDPA does require, and what English-drafted policies routinely miss, is a privacy notice in Bahasa Malaysia AND English under section 7(3). That is statutory, not guidance. An English-only privacy page on a Malaysian controller's site is a defect on its face, and it has nothing to do with chatbots... it just tends to be discovered at the same time.

Everywhere. Whatever the local rule, "never claim to be human" and "say you are an AI" are two different duties. A bot instructed to answer honestly if asked satisfies the first and not the second.

02 / Our own failureWe measured our own fleet and failed it, 16 out of 16

In 2026-08 we built a checker for this and ran it across every bot we could reach. All sixteen lacked a first-message disclosure.

Not sixteen mistakes. ONE mistake, cloned sixteen times. A single instruction, written when it was a perfectly defensible UX choice, saying in effect: do not announce that you are an assistant unless someone asks. It was correct when written. It propagated with every clone. Nobody re-read it when the law changed, because nothing about a clone invites you to re-read it.

Then we made a second error, and it is the more useful one. The first version of that finding read "16 of 16 non-compliant" ... against an EU regulation, with no geographic scoping at all. Most of those bots serve customers in a country the regulation does not reach. The honest count was sixteen bots with no disclosure, of which only the EU-facing ones are a legal matter.

Over-applying a rule is not a safe default. It burns the credibility of the check, and a check nobody believes gets switched off. That is a worse outcome than the gap it was meant to close.

03 / The measured priceThe measured price of disclosing, which is real

This is the part that is usually asserted from the gut in both directions. There is actual research.

Schilke & Reimann (2025), Organizational Behavior and Human Decision Processes 188, 104405. Thirteen experiments, 4,093 observations, meta-analytic effect d = 0.81, negative in every single study.

Three findings that should change what you BUILD, not just what you say:

1. Being caught costs about twice as much as being open. In their tax-adviser scenario, trust scored 4.02 when nothing was said, 3.15 when AI use was disclosed, and 2.49 when a third party revealed it. So disclosure is still the right call, but as insurance against exposure rather than because it is free. Silence is a leveraged bet on never being found out, in a market where detection keeps getting easier.

2. "Reviewed by a human" buys back nothing. They tested six framings. "Prepared by AI" scored 3.04. "Prepared by AI and reviewed and revised by the human sender" scored 3.03. Control was 4.55. The only framing that recovered anything was one that NARROWED what the AI did ("proofread", 4.03).

That single result invalidates the standard fix. Every business instinct here is to bolt a human-oversight sentence onto the disclosure and assume it neutralises the penalty. It does not move the number.

3. The penalty scales with how much the reader expected YOU to think. A scheduling email, d = 0.36. A tax estimate, d = 0.81. A job application, d = 1.42.

That is the most useful lever in the whole study, and it is free: automate where nobody expected your personal judgment, and keep judgment work in a human's name. The disclosure cost is not fixed. You choose it, by choosing what to automate.

The counterweight, which resolves the tension: a 2026 survey of accounting clients found 89% wanted transparency about AI use and only 2% objected to AI at all. People object to being kept in the dark. They do not object to the tool. A one-time disclosure answers that. A label stapled to every message over-answers it and pays the penalty repeatedly.

04 / What we buildWhat we build, and the one line we will not cross

We treat disclosure as the client's recorded decision, not our house rule. You are the deployer. The duty, the revenue and the risk appetite are yours. A vendor that overrules a documented client decision does not create safety; it creates a gate that people route around.

So we recognise three positions, and we write down which one you picked and when:

The failure state we actually care about is the fourth one: a bot shipped with nobody having thought about it. That is the only version where a vendor carries a client's legal risk with nothing in writing.

And there is one thing that is not on the menu. A bot we build will not assert that it is a human being. You may choose silence. You may not instruct our software to claim personhood. Declining to volunteer and affirmatively lying are different acts, and the second one stops being a disclosure question and becomes a misrepresentation question.

That line costs nothing commercially, which is exactly why it is the right place to put the only hard rule.