Second Chair
600 curated sources — scholarship, industry evidence, primary legal materials and 73 practitioner interviews — read as one record, straight from the source database. Corpus-level figures are computed live; figures attributed to a source are that source's own.
The corpus sorted itself into five families. Each is a doorway into what its topics argue.
AI compresses the production of legal text; judgment, accountability and client context stay with lawyers. Trials find speed gains that are large and consistent, quality gains that depend on grounding, task type and training — and grounded commercial tools that still err in the tens of percent, subtly.
What makes work automatable is not how routine it is but how cheaply the output can be checked — which puts legal drafting in the risk zone where execution is cheap and verification is not. Whether AI levels performance or creates a new divide is genuinely unsettled here.
Benchmark scores, prices, disclosures and star ratings all measure something adjacent to legal quality. Experimentally, liability for outcomes fixes what disclosure and auditability do not — which is awkward, because almost every instrument proposed for legal AI is the second kind.
The instruments people cite when arguing about automatable legal work measure different things, so their conflict is largely definitional. Alongside them sits one large employer survey whose sharpest claim is distributional: 11 workers in 100 are unlikely to get the training they need.
The gains in who enters the profession sit at associate and summer-associate level — exactly the document, research and drafting work most exposed. And the national average describes almost nowhere: partner diversity varies several-fold by firm size and city.
Read all five in full → — every claim cited to its cluster and its source, both clickable into the record.