Almost everyone spontaneously answers yes: of course we should know. Then you ask whether a spellchecker should flag itself, whether a customer service team using AI to draft its replies should mention it, whether a friend who had a machine proofread their message should warn you. And the certainties dissolve.
The topic resurfaced this week with the episode of the watermark bypassed in four hours. Behind the technical question of detection lies a question of principle that is rarely tackled head-on.
The case for: consent
The reasoning is solid. We do not grant the same trust, nor the same kind of attention, depending on who we are dealing with.
Facing a human, we deploy an extraordinarily fine social skill: gauging their sincerity, competence, interests. We developed this in our article on good faith: this skill does not work with a machine, because there is nothing to gauge. An AI's tone contains no information about its reliability.
Not knowing therefore means using a measuring instrument on an object it was not designed for. You think you are sizing someone up; you are sizing up nothing.
Add to that the manipulation argument. A system designed to seem human exploits a reflex we do not control, as we explained in our article on anthropomorphism. Knowing does not disable the reflex, but at least allows us to correct it consciously.
The case against: where does the obligation stop
The objection is not bad faith from companies wanting to conceal. It is more embarrassing: where do you draw the line?
If an AI writes an entire message, disclosure seems necessary. If it fixes a typo, it seems absurd. In between lies almost all real-world usage: a reworked draft, a suggested idea then reformulated, a proposed outline then modified.
Requiring disclosure for every use would render the information useless through saturation, exactly like cookie consent banners. A notice displayed everywhere informs no one.
And there is a deeper objection: we have never demanded this of any other tool. No one indicates that a text was proofread by an editor, written by a ghostwriter, or produced by a consulting firm.
The strongest line is not drawn by the degree of machine involvement, but by what the interaction entails. When someone believes they are in a relationship with a person, they have a right to know. When they are using a service, what matters is the quality of the service. A chatbot posing as a human advisor deceives about the relationship. A customer service whose replies are drafted with assistance deceives about nothing, as long as the answer is correct and a human stands behind it.
The three cases where the answer is clear
This criterion yields three areas where doubt is not permitted.
Affective relationships. A system designed to create attachment must be identified as such. This is the topic we covered in our article on AI companions, and China has already ruled by banning such devices for minors.
Vulnerable situations. Health, psychological distress, legal or financial advice. Someone addressing what they believe is a professional must know what they are talking to.
Decisions that carry weight. As we saw regarding recruitment: the European framework already provides that a person subject to a significant automated decision must be informed of it and able to contest it.
The enforcement problem
One difficulty remains, and the watermark episode makes it glaring. Even if we agree on the principle, how do we enforce it?
Technical marking can be bypassed in hours. Voluntary disclosure depends on good faith. And a legal obligation only holds if the violation can be observed, which is precisely what we cannot do.
This suggests the solution will not come from detection but from responsibility. Rather than requiring every piece of content to be marked, hold the publisher accountable for what they publish, regardless of the tool used. That is a significant shift: we stop tracking the machine and return to the one using it.
What to take away
Our answer, and it is debatable: the right to know is not absolute; it is proportional to what the interaction entails.
No one needs to know that a professional email was proofread by a machine. Everyone needs to know that the warm voice listening to them every evening is not someone. In between, the useful criterion is that of relationship: as soon as a person might believe there is someone on the other side, you must tell them.
Perhaps that is the simplest formulation: what we owe the other is not transparency about our tools. It is not letting them believe in a presence that does not exist.