Will human judgement protect lawyers from AI?

Not as a job description, and on that Richard Susskind is right. In the thirtieth anniversary edition of The Future of Law, reported by Neil Rose at Legal Futures on 28 September 2026, he argues that by the early 2030s generative AI “will be reliable for most practical purposes”, that most systems “will no longer require human supervision”, and that a human in the loop “will more likely diminish than improve performance”. The evidence on this estate cannot yet test that forecast. What it shows is different: the courts have decided who answers when the reliable machine is wrong, the work is leaving the firms that trained juniors, and a Crown Court judge this week wrote eight conditions before letting a tool into a trial. Judgement will not keep a lawyer’s job. It is what a client still buys, if it is still exercised.

What Susskind said, as reported#

Rose’s report describes a book that treats artificial general intelligence, in Susskind’s words “systems that are as apparently capable as humans at undertaking all tasks that require our full range of cognitive skills”, as the frame for legal work. His central move is to refuse the profession’s usual refuge: “the big question here is not whether AI systems are able to make judgements” but “whether machines can handle uncertainty”, and his answer is “manifestly and absolutely” yes, “on the strength of the oceans of data on which they were trained, whereas humans rely on their modest mini-databases of personal experience”. From there he expects non-lawyers to handle their own legal affairs, conventional lawyering to shrink to critical situations, business models to move from selling time to licensing systems, and the firms that build tools to fare better than those that compete with them. It is an argument, made by the person who has made it longest, and it deserves to be met on its own ground rather than waved away with the word judgement.

The reliability record so far#

“Reliable for most practical purposes” is a forecast about 2032. The record to 2026 is on how AI will change law: Damien Charlotin’s database of decisions in which a court found a party had relied on hallucinated material stood at 1,963 cases at its 27 August 2026 update, the largest group responsible was self-represented litigants rather than lawyers, and in Ayinde in June 2025 the Divisional Court settled that lawyers using such tools must check the research against authoritative sources before using it. The commercial research tools handed to juniors were measured by Magesh and colleagues at Stanford getting the law wrong between a sixth and a third of the time, examined on what happens to medical and legal training. None of that refutes Susskind, who is describing systems that do not yet exist. It does show what “most practical purposes” means in a courtroom: a system that is right in the large majority of cases produces, in the residue, a citation that never existed, a submission struck out, and a name in a judgment. A machine can be reliable and a profession can still be built on the cases where it was not.

Where the work is going, measured#

The economics arrived the same week, from Luis Garicano at the London School of Economics in a Brookings paper presented on 25 September. His argument is one sentence: specialists “can spread the fixed cost of acquiring knowledge across many clients. Artificial intelligence reduces the fixed cost and hence benefits those who use this knowledge less often.” So the client’s in-house team does the work the outside firm used to sell, and the outside firm keeps the rare, hard problems. He examined six occupations employing 5.9 million workers in the United States, and found that in four of them, lawyers, accountants, software developers and management analysts, employment in specialised outside firms fell faster between 2022 and 2025 than the pre-AI trend; he cites corporate legal departments reporting more AI use and less reliance on outside advisers. The consequence he names is a “broken ladder”: the less complex jobs in outside firms that trained graduates go first. That is Susskind’s prediction and the estate’s worry in one dataset. The routine work leaves the firm, and with it the years in which a junior learned, by doing the routine work, to tell when the answer was wrong; the same mechanism on whether AI will replace entry-level jobs.

Handling uncertainty is not the same as answering for it#

Here is where the estate parts from Susskind, and the difference is not about what machines can do. A system trained on oceans of data can handle uncertainty in the statistical sense: it can weigh, hedge and pick the likelier reading. What it cannot do is be the person the court holds responsible, the person a client can sue, the person who can be struck off, or the person who knows this client, this judge and this year in a way no training corpus contains. Those are not cognitive skills and no measure of cognitive performance will make a machine acquire them. The estate’s position, set out on whether a professional could be negligent for not using AI, is that the standard of care, on the UK Jurisdiction Taskforce’s July 2026 statement, now points both ways, at failing to use the tool and at using it carelessly, and still falls on the professional, so the value of a lawyer in 2032 is the value of someone accountable who can tell, in this case, whether the reliable machine is right. That capability is only there if it has been exercised. The deskilling evidence is about what happens to it when it is not, and Garicano’s broken ladder is about the disappearance of the years in which it was built. Susskind is right that judgement is not a moat. He has not shown that it is not the product.

A judge writes the rules before the tool#

The same day, Rose reported that His Honour Judge Nicholas Rimmer at Southwark Crown Court had granted six barristers permission to record proceedings and have them transcribed, calling it “something of a leap into uncharted territory” and saying it “is likely to improve productivity and the accuracy of submissions made about the evidence in trial proceedings”. The conditions are the interesting part: the recording is for trial work and case preparation only; transcription on the device without an internet connection, or through one named external service the court had examined, with encryption, no cloud storage and immediate deletion from that service; no dissemination; deletion as soon as practicable after the trial; and the recording “not available to large language model training”, with two products named. A court that expects automated daily transcripts “for all professional court users” wrote down, before permitting a tool, which task it may do, who holds the file, and what must never happen to it. That is Rules Before Tools in a judge’s hand, and the profession Susskind describes will look like that if it gets the order right.

What a firm can decide this quarter#

A firm cannot decide whether Susskind’s 2032 arrives. It can decide four things now. Which tasks the machine may do unsupervised, which it may only draft, and which a named person must do from the source, with the June 2025 duty as the floor. Who can stop a tool in a matter, and whether a junior may do so without being asked to justify the cost. What every fee-earner must remain able to do without the tool, tested rather than assumed, because the ladder Garicano describes is the firm’s own training scheme and it breaks first at the bottom. And how the firm would know the machine was wrong before a judge did: a sample of AI-assisted work checked by someone who did not use the tool, recorded, and read by a partner. A firm that has done those four has a defence to the negligence claim, a product to sell when the routine work has gone in-house, and juniors who will be able to tell when the reliable machine is wrong. A firm that has not has Susskind’s future without the part he says survives.

What this does not show#

Susskind’s book has not been read for this page; the quotations are Legal Futures’ report of it, and a report of an argument is not the argument. His claims are forecasts and cannot be tested until the decade he names. Garicano’s paper is a conference draft read through Brookings’ summary; the employment declines are United States figures for six occupations, the comparison is against a trend rather than a control, and other causes, from interest rates to offshoring, are not ruled out on the page read. The hallucination count is a running tally compiled by one researcher and understates cases never reported. The Southwark order is one judge’s decision on one application under one rule and sets no precedent. Nothing here measures whether lawyers who use AI heavily lose the ability to tell when it is wrong; that is inferred from the deskilling record in other professions, and a firm that ran the sampled check above would be the first to have the number.