AI does not sign: why professional responsibility will remain human
Delegating tasks to the machine is reasonable; delegating judgment is an abdication. The signature at the foot of a pleading is a declaration: I have read this and I answer for it.
There is a scene repeating itself in too many law firms already. A lawyer receives back from the AI tool a reviewed contract, a case-law summary or a draft pleading. The text is fluent, the citations look correct, the formatting is impeccable. And at that moment the most important professional decision of the process takes place: actually reading it or taking it as good. The entire debate about AI in legal practice comes down, in the end, to that instant.
Public discussion has settled into two caricatures. The enthusiasts sell a firm where AI works and the lawyer bills; the catastrophists announce the end of the profession. Both overlook the central legal fact: responsibility cannot be outsourced. When a pleading reaches a court with an invented judgment, the sanctioned party is the signing lawyer, not the model that hallucinated. When a contract contains a clause harming the client, the one answering before the professional liability insurer and before the bar is the lawyer. AI has no licence, no policy and no assets. AI does not sign.
Proof that this is so lies in the tools' own terms of use: no legal AI provider assumes responsibility for generated content. All of them shift it to the professional user. The market, more honest than the debate, has already decided who signs: the party with something to lose. And that someone, today and in 2030, will be a licensed person with a name at the foot of the document.
This is not an argument against technology, but for using it as what it is: an extraordinary production tool under human supervision. The correct comparison is not the replaced lawyer but the brilliant subordinate without experience or licence: they produce much and fast, and everything they produce must pass through the hands of someone who answers for it. Firms that understand it this way are obtaining real advantages, more speed on the first pass, more coverage in research, more time for judgment, without giving up anything essential. Those who delegate judgment accumulate a risk that one day will be called a disciplinary file or a liability claim.
There is also a commercial dimension the technological debate obscures. The client does not pay for text: they pay for guaranteed judgment. If the lawyer's value were producing words, AI would already have won. What the client buys is that a person with a name, a licence and assets answers for the correctness of that advice. The more the machine trivialises draft production, the more value shifts towards the only thing that cannot be downloaded: a decision assumed with responsibility.
Some reply that this is guild fear, that other professions resisted automation and lost. The comparison is misleading. The legal profession does not defend a privilege but an architecture of guarantees: someone identifiable answers, with their assets and their licence, for the advice a citizen receives. That architecture was not invented by lawyers to protect themselves; society built it to protect whoever needs legal help. Introducing AI within it is perfectly compatible. Claiming the machine replaces it means offering the weakest client advice without anyone responsible.
Regulation is moving that way. European professional guidelines, the CGPJ instruction on judicial use of AI and the first sanctioning cases all point to the same principle: human supervision is not a recommendation, it is the condition of use. We develop that supervision in our human-verification protocol for contract review, and the context in the CCBE guide on generative AI and the analysis of AI and professional secrecy. The machine will write ever better. The signature will remain ours.
