The exam that makes a British solicitor is hard and it is measured. The two years of work behind it are not measured at all.

Two claims are being made about the training of British lawyers, and they point in opposite directions.

The first is that AI is eating the work juniors learned on, so the profession is quietly failing to make its next generation of experts. The second is that AI clears away the drudgery, so juniors get real work sooner and learn faster. Both are plausible. Both are being asserted with confidence. This desk went looking for what is counted.

Quite a lot is, and it is worth saying so. Since 2021 anyone qualifying as a solicitor in England and Wales sits the SQE, and the SQE is measured in public. In the January 2026 sitting, 7,863 candidates were graded on both papers of SQE1 and 53 per cent passed. Among those sitting for the first time, 58 per cent passed. The regulator's four-year review, covering assessments from November 2021 to July 2025 and more than 30,000 candidates in 50 countries, puts first-attempt SQE1 pass rates between 46 and 60 per cent, with 66 per cent passing eventually. SQE2, which tests practical skills, runs higher: 69 to 84 per cent first time, 85 per cent eventually.

So the examination is real. It fails a lot of people. It is not a formality anyone drifts through.

Now the other half of qualifying, and this is where the measurement stops. Alongside the exams, a candidate must complete two years of qualifying work experience. The SRA's requirements are these: the work must involve providing legal services, and it must give the candidate the chance to develop at least two of the competences a solicitor needs. Not all of them, two. The regulator does not specify which two. The experience can be assembled across as many as four organisations. It can be paid or unpaid. At the end, a solicitor or the firm's compliance officer confirms it happened.

There is no test of what those two years taught. There is no measure of the quality of supervision, no assessment of the work, no requirement that anyone check the candidate learned anything at all in the workplace. The two years must have occurred and been signed for.

Hold that against the question everyone is asking. If AI removes the junior work that used to build judgment, the machinery of qualification would register precisely nothing. The clock runs the same. The signature is the same. The exam, sat separately, might catch it eventually, or might not.

The nearest thing to evidence on the point is a survey of 873 UK legal professionals conducted across December 2025 and January 2026: 72 per cent were concerned that juniors using AI will struggle to develop legal reasoning, 69 per cent worried about verification and source-checking, and 2 per cent thought AI strengthened learning. That research was published by LexisNexis, which sells legal AI tools. It is a vendor survey of professional opinion, not a measurement of outcomes, and it should be read as both.

There is a real counter, and it should be stated plainly rather than buried. SQE2 tests practical legal skills directly, so a candidate who learned nothing useful in two years ought to struggle there. And the group that performs best in the whole system is the one doing the most structured work-based learning: solicitor apprentices pass at around 71 per cent on SQE1 and 93 per cent on SQE2, against 66 and 85 for the cohort as a whole. If work-based training were failing, that is not the pattern you would expect.

Why it matters here. The honest position tonight is not that AI is breaking the training of British solicitors. It is that Britain has built a qualification system that measures the exam and counts the years, and measures nothing in between. The profession is having a loud argument about the training ground while the regulator collects no data capable of settling it. If this goes wrong, we will find out from a shortage of competent seniors in fifteen years, which is the most expensive way to learn anything.

Disclosure: Anthropic builds the models this desk runs on, sells legal AI through Claude for Legal, and its model powers Harvey. We report on this industry from inside it.