I have written terms and conditions. In earlier professional projects, as a consultant and a co-responsible party, I sat in the room where a company decides what to put in writing and what to leave out. I know how these documents get set up.

A lawyer drafts the text. The client is the company, so the lawyer protects the company. The language stays neutral on the surface, often layered in subordinate clauses. The document is built as a clean record: proof, if a dispute reaches court, that the customer was informed and clicked confirm. No one reads the whole document; it holds up in court regardless.

Ethics sections in AI projects run on the same logic. Open a handful of them and the pattern repeats. Each names a real problem: bias in an algorithm, discrimination through a deepfake, a lack of transparency in how a model decides. Then comes language that sounds like responsibility: regular audits, ethical guidelines, responsible handling of data. The tone lasts until the next feature ships, and the cycle starts again. Users have already handed over their personal data at registration, often covert and far-reaching in scope, and nobody resists at that point either.

I started reading the ethics sections in these kinds of documents more closely, because they weren’t verifiable anyway. So I tried to trace whether concrete actions could be derived from the words. Take the standard promise of regular audits against bias, usually made in the context of serious discrimination against gender or minority groups. Which audits. At what interval. Against which criteria. Who runs them. Who pays for them. And if an audit finds a gap or an outright violation, does the system get shut down, adjusted or left running. None of these questions gets answered specifically enough to amount to protection or consequence.

That gap is deliberate. Answering those questions honestly would mean naming a real problem before it is solved: we learn more than we are allowed to store, or the chatbot does not behave reliably. That would be honest, and it would be uncomfortable, because it would have to slow the product down while it is still shipping, and it would show that the problems are real and that there is no easy fix. The honest sentence puts a company’s competitive position at risk, while the phrase “we conduct regular audits” reads identical to a genuine commitment and carries no such risk.

My own view is that this is what makes the pattern worse than ordinary corporate caution. A fire extinguisher on the wall that has never been inspected still looks like safety equipment until the moment it is needed. A medicine advert’s closing disclaimer, listing risks and side effects, discharges the advertiser’s duty to warn without changing a single dose anyone takes. Ethics sections in AI products work the same way. They exist, nobody reads them, and they still do their job: they document that the user was told. Documentation was the whole job.

The engineers building these systems are not the ones deciding this. In my experience, technical teams tend to rank ethics low among requirements, and an AI application with real ethical safeguards is harder and more expensive to build than one without them. The vague language downstream of that decision is deliberate, the intended output: legally unenforceable, protective of the company and silent on what happens when something goes wrong.

The honest version would sound different. It would say: this tool can discriminate, and we do not yet know how to prevent it. That sentence would build trust because it is true, regardless of whether the problem is solved. It would also draw less attention and slow a launch down, which is exactly why almost nobody writes it. Attention goes instead to whoever performs ethics as though it were a configuration setting.

The distinction that matters is between mentioned and addressed. A section that names bias and follows it with the phrase “regular audits” has mentioned the problem. It has not addressed it, because addressing it would mean specifying who audits what, how often and what happens when the audit fails. Most readers close the document certain that ethics was taken into account. The section did its job: it documented that they were told.