Thomson Reuters announced on 24 August that it has built its own large language model, named Thomson, for legal work. The company says it trained the system from an open-source base on decades of its own material — Westlaw, Practical Law, Checkpoint and Reuters content — with a reported investment of around $40 million in compute and talent.
The announcement is a small one by the standards of frontier AI spending. It is interesting for a different reason: it is a test of whether owning the right data beats owning the biggest model.
The strategy behind it
Most companies deploying AI licence a general-purpose model from a developer and layer their own data on top through retrieval. Thomson Reuters has done something else: taken an open-source base model and trained it further on a proprietary corpus that competitors cannot obtain, because the company owns it.
Westlaw is one of the two dominant legal research databases in the United States, containing case law, statutes, and — importantly — decades of editorial headnotes and classification written by lawyers to summarise what cases actually decided. That editorial layer is the asset. It is not simply text; it is text that has already been organised by experts into the structure the profession uses.
Whether that produces a materially better legal assistant than a frontier model with good retrieval is an empirical question that has not yet been answered publicly.
From Plain Sight News
Why law is a hard case for language models
Legal research is the domain where AI failures have been most publicly embarrassing. Courts in several countries have sanctioned lawyers who filed briefs citing cases that did not exist, generated by models that produced plausible citations rather than real ones.
The underlying difficulty is that legal citations are exactly the kind of output language models are structurally prone to inventing: highly formatted, drawn from a vast corpus, and superficially checkable only by someone who looks them up. A model trained on a verified proprietary database has a plausible advantage here, since the cases it draws on are ones that exist in a maintained system.
Plausible is not proven. Training on accurate data reduces certain errors; it does not eliminate the tendency of these systems to generate confident text where they have gaps. Any legal AI product should be judged on published evaluations of citation accuracy, not on the quality of its training corpus.
The wider pattern
If this approach works, the implication reaches well beyond law. There are many industries where one or two firms hold decades of specialised, curated data: medical publishing, financial data, engineering standards, scientific literature. Each is a candidate for the same move.
That would mark a shift in where value sits in the AI industry. The assumption behind enormous frontier-model spending is that general capability is the scarce thing. The alternative view is that capability is becoming a commodity — available from open-source bases for tens of millions rather than billions — and the scarce thing is proprietary domain data.
It also raises a competition question. If the most useful AI in a profession is built on a database controlled by one company, the AI inherits that company’s market position. Legal research has long been criticised as an expensive near-duopoly; a model trained on one side of it does not obviously change that.
From Plain Sight News
What it means for the people doing the work
The tasks these systems are aimed at — searching case law, summarising authorities, drafting standard clauses, first-pass document review — are largely the work that junior lawyers and paralegals have traditionally done.
Two views are held with conviction in the profession. One holds that automating routine research frees junior lawyers for judgement-based work and makes legal services cheaper for clients priced out of them. The other holds that the routine work is how junior lawyers learn judgement, and that removing it produces a profession with no path from novice to expert.
Neither has been settled by evidence, because the tools are too new. It is worth noting that both arguments have been made about every previous wave of legal technology, from keyword search onwards — which is a reason for caution about confident predictions in either direction.
From Plain Sight News
Sources
Plain Sight News has not independently reviewed the model or any evaluation of its performance. The figures given for training data and investment come from the company’s own announcement as reported.
Leave a Reply