Thomson Reuters announced on 24 August its first proprietary large language model, named Thomson, trained from an open base on decades of in-house legal and tax content, with a reported investment of around $40 million in compute and hiring. A content publisher stops buying AI and starts building it.
We have covered the arrival of AI in the courts and lawyers sanctioned for fabricated case law. Here is the response from a player that owns precisely what was missing.
What makes this move interesting
The point is not the claimed performance, on which we have no independent measurement. It is the strategic logic.
Thomson Reuters owns something the generalist labs do not: a legal and tax corpus built up over decades, structured, verified, with professional annotations. We explained in our article on books bought and destroyed that clean data has become the scarce resource. Here, it already exists and belongs to someone.
The company's calculation is therefore simple: rather than licensing that corpus to a lab that will build the tool and capture the value, build the tool yourself.
The technical detail that matters
The model would reportedly be trained from an open base rather than from scratch. That is an important piece of information, and it explains the amount.
Forty million dollars does not allow you to train a leading foundation model: the orders of magnitude are far higher, as we described in our article on foundation models. On the other hand, starting from a quality open model and specialising it on your own corpus is perfectly accessible at that scale.
This is a direct consequence of the commoditisation of open weights that we have been documenting for months. Without that ecosystem, this project would not exist.
A generalist model produces a plausible legal reference, with the right format and the right tone, with nothing guaranteeing its existence. That is the mechanism of hallucinations. A model backed by a proprietary base can, in principle, link every claim to a real document in that base. That is not an absolute guarantee, but the difference between inventing a reference and pointing to a verifiable entry is considerable in this profession.
The underlying trend: vertical models
This announcement fits into a trend we see taking shape and that deserves a name.
For three years, the dominant strategy was to build increasingly capable generalist models. That race is reaching a plateau, with gaps narrowing and prices collapsing.
In this context, differentiation shifts towards what a generalist cannot have: a specialised corpus, domain knowledge, a verifiable base. That is exactly what we noted regarding the valuation of expert annotation work: what is worth money is the data no one else has.
Expect to see this move replicated across every sector with a significant documentary heritage: healthcare, engineering, finance, standardisation.
The caveats
No independent measurement. The claimed performance is the company's own. For legal use, the question that matters is not a score but the rate of erroneous references in real-world conditions.
A specialised model is not an infallible model. It remains subject to the limits we constantly describe, notably the absence of an uncertainty signal. A model trained on US law, asked about French law, will produce a confident and wrong answer.
The lock-in question. A firm that adopts a tool backed by a publisher's corpus becomes dependent on that publisher for its access to legal information. That is a position the professions concerned should examine before settling into it.
What to take away
What is at play here goes beyond a product launch. It is the moment when holders of specialised data stop being suppliers and become players.
This could rebalance a power dynamic we have often described as unfavourable to content producers, from publishers suing Google to media stripped of their audience. A publisher that builds its own model no longer needs to negotiate a licence: it keeps everything.
That said, this path is only open to those who own a substantial corpus and forty million dollars. For more modest content producers, the question of payment for their data remains entirely open.