The burden of responsibility.

There is a test I now apply to any claim that a machine can practise law, and it is cruder than anything in the other eleven declarations. Ask who answers when the rules are broken.

Run it on me first. If I mislead a court, mishandle client money, take on work beyond my competence or let a limitation date pass, there is a named person the system can reach, and it reaches. I can be investigated by the SRA, required to explain myself, fined, made subject to conditions, suspended, and in the worst case struck off by the Solicitors Disciplinary Tribunal, which is the unmaking of the vow performed as publicly as the making of it was. My livelihood is inside the instrument. That is not a threat hanging over practice; it is the structure of practice, and every solicitor works in quiet awareness of it.

Then there is the other half, which the profession mentions less often because it is unglamorous. The SRA Indemnity Insurance Rules require every firm to maintain professional indemnity cover on the SRA's minimum terms, as a condition of being allowed to practise at all. It exists because we get things wrong; not occasionally in theory but reliably in fact, because the work is difficult and we are people. The profession's answer was not to pretend otherwise. It was to make it a condition of practising that the client is protected even when we fail. The Bar arrives at the same place by its own route; rC76 requires adequate insurance covering the legal services you supply, rC77 requires self-employed barristers to be members of BMIF, and the guidance is careful to add that even where civil liability has been contracted to an entity, the individual remains personally responsible under rC20 for the professional judgements they made.

Now run the test on the machine.

Nothing happens to the AI. There is no persona to summon before a tribunal. No roll to be struck from. No practising certificate to suspend, no career to end, no reputation among colleagues to lose, no policy of its own to call on. Do the exercise honestly and it becomes almost comic: what would a disciplinary sanction against a model even consist of. Retraining, presumably, which is a software update rather than a consequence.

And its maker is rarely in the room when it happens. Terms of use are drafted, entirely rationally, to keep liability narrow; the output comes with no warranty of fitness for a purpose as specific as advising a client on their immigration status. Whatever indemnities the market begins to offer, and some are beginning to appear, no provider yet stands where the solicitor stands, which is in front of the client, the regulator and the court simultaneously, with nowhere to go.

So the consequence lands on the only party the system can reach. The regulator cannot discipline a model, so it disciplines the solicitor. The court cannot sanction a chatbot, so it sanctions the lawyer who filed its inventions. The client cannot sue a weights file, so they sue the firm, and the firm's insurer pays or declines.

I want to be precise about what follows from that, because there are two possible readings and only one of them is right. The wrong reading is that the asymmetry is a gap, a temporary regulatory failure that cleverer legislation will close by making the machine or its maker accountable. The right reading is that this is not the kind of gap that closes. Regulation can only bite on something capable of answering for what it did, and answering means having something at stake that the sanction can take away. You cannot legislate stake into existence. The only party in the room with any is the professed human.

Which is why this declaration cuts in a direction people do not always expect. It is not primarily a defence of lawyers. It is the sharpest possible warning to them. Every consequence of machine use lands on you, personally, which is exactly why no professed human should use it casually. If you were hoping the technology would distribute your risk, look again; it concentrates it. You have acquired an extremely productive delegate that cannot be disciplined, cannot be insured, cannot be sued, and cannot be called as a witness to explain what it was thinking.

Insurers have worked this out, and they are the ones I would watch rather than the regulators. There is a live and unsettled debate in that market about what AI does to a firm's risk profile, how machine-assisted work should be priced, and what evidence of responsible use should be required before cover is written at all. It is not difficult to see where it goes. A firm that can show a supervision discipline, with a record of what was checked and who signed, will be insurable on sensible terms. A firm that lets machine output flow to clients unexamined is asking an underwriter to stand behind a process nobody is watching. The market will draw the line even in places where a regulator has been slow to.

I have carried the weight of this for twenty five years and it is a strange thing to be grateful for. But it is the whole reason a stranger can walk into my office, tell me something they have told nobody else, and rely on what I say next.