What does Dr. Jonah Tebaa on Why Your Customer-Facing AI Should Admit It's an AI mean in practice?
According to Dr. Jonah Tebaa, customer-facing AI must admit it is software because failing to disclose creates an immediate trust liability, even if answers are accurate. In a composite illustration of a Beirut clinic handling 340 messages, an assistant deflected direct inquiries simply because no response was defined. To prevent trust injuries before legal mandates spread, Dr. Jonah Tebaa recommends an eight-point pre-launch checklist that enforces unprompted disclosure in the initial reply, logs human-inquiry sessions, and assigns a single named owner.
Dr. Jonah Tebaa's latest work turns away from a question he has written about before, how an organization should respond after an AI system fails, toward an earlier and quieter one: whether a customer is ever told, before anything goes wrong, that they are talking to a machine at all. His argument is that this decision, made or left unmade, is now a live trust liability for small and mid-market businesses across Lebanon and the Gulf, independent of how accurate the AI's answers happen to be.
To ground the argument, Dr. Tebaa uses a composite illustration, not a real named business, of a Beirut clinic whose WhatsApp assistant handled 340 patient messages in a month. Fourteen of those messages asked, in Arabic, in the French-inflected Arabic common in Beirut, or in English, some version of one question: am I talking to a real person? In the illustration, the assistant answered honestly seven times and deflected the other seven, not out of any deliberate deception, but because nobody had written down what the system should say when asked directly.
A Trust Problem Before It Is a Legal One
Dr. Tebaa is careful to frame this as a trust question first and a compliance question only secondarily. In his view, most owner-operators in the region treat disclosure as a line in the terms of service, while worrying almost entirely about whether the AI's answers are correct. He argues this gets the priority backward: a customer who receives a wrong answer from a system they knew was software tends to forgive it, while a customer who later realizes they spent real time describing a problem to software they believed was a person experiences something closer to a trust injury, regardless of whether the underlying answer was right.
He does note the regulatory backdrop without overstating it for a Lebanese or Gulf audience. Transparency obligations for AI systems that interact directly with people are already part of the EU AI Act, and he expects that regulatory instinct, that people have a right to know when they are speaking to a machine, to reach regional platforms and consumer-protection expectations faster than most SMEs currently assume. His practical point is that none of this requires legal counsel to fix today. It requires a short, written checklist, run before launch.
An Eight-Point Pre-Launch Checklist
The checklist Dr. Tebaa proposes has eight testable conditions. Disclosure must appear, unprompted, in the first substantive reply a customer receives, not buried in a linked document. There must be a plain answer, tested against the exact phrasing customers use, including Arabic and the code-switched forms common across the region, ready for the moment someone asks directly whether they are speaking to a robot. Disclosure should repeat at high-stakes moments, before payment or anything resembling professional advice, not only at first contact. A working path to a human must exist and be tested by someone outside the build team, not assumed from a vendor's specification sheet. A single named person, rather than an undefined "IT" or "the vendor," must sign off on the disclosure wording and remain responsible for it after launch.
The remaining items address ongoing accountability rather than the initial launch decision. Every session in which a customer asks whether they are talking to a human should be logged, Dr. Tebaa argues, so leadership can audit the system's actual behavior rather than trust that a training document was followed. The disclosure trigger should be re-tested whenever a vendor pushes a model or prompt update, since tone changes made for unrelated reasons can quietly make an honest answer sound evasive. And a written remediation step should exist for the case where a customer does proceed without realizing they were speaking to software, so the business has an answer ready rather than a hope the scenario never arises.
What Separates Honest Disclosure From an Accidental Dodge
Dr. Tebaa illustrates the distinction with two short, contrasting replies. An honest disclosure states plainly what the system is, what it can do, and offers a path to a person without being asked, in a single opening message. An accidental dodge, by contrast, often says nothing false at all. It simply mirrors the warmth and phrasing of a helpful human colleague closely enough that a customer has every reasonable basis for assuming they are dealing with staff rather than software. In his analysis, that gap between the two replies is not a matter of tone. It is the entire disclosure decision, settled or unsettled, inside a single line of copy that most teams never stress-test before launch.
His broader conclusion is that the fix is inexpensive relative to the exposure it prevents. Running an eight-point checklist against a customer-facing AI system before it goes live costs an afternoon and a named owner. Discovering the gap after a customer notices costs a business the one thing a checklist cannot restore after the fact, which is the customer's assumption that the business was being straightforward with them from the first message onward.
Related evidence: Article 50 of the EU AI Act requires providers to design AI systems that interact directly with people so that those people are informed they are interacting with an AI system, unless that is already obvious in the circumstances and context of use. (EU AI Act Article 50 transparency obligations)
NIST's AI RMF appendix on human-AI interaction notes that AI systems can autonomously make decisions, defer decision making to a human expert, or be used by a human decision maker as an additional opinion. (NIST AI RMF appendix on human-AI interaction)
