Thomson Reuters launches proprietary AI model for legal work
SiliconANGLE Paul Gillin ● Covered by 2 sources
Thomson Reuters built its first in-house AI model for legal work. It’s meant to know law better than general chatbots, not beat every AI everywhere.
Based on reporting by SiliconANGLE, Paul Gillin — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Thomson Reuters has launched Thomson, its first proprietary large language model, and it is aiming it straight at legal work. The model sits inside CoCounsel Legal AI assistant and will debut in Tabular Analysis, the company’s high-volume document review tool. For now, Thomson Reuters still wants CoCounsel to stay multimodel: Thomson will handle jobs where its legal training helps most, while outside frontier models will cover the rest.
The company says it spent about $40 million over two years on people and compute to build the project, though the final training run came in at about $450,000 after economies kicked in. It did not start from zero. Instead, Thomson Reuters began with an open-weight model and layered in its own content, training methods and professional expertise. That choice, it says, cut both training and inference costs versus using broad-purpose frontier models.
This is not a play for AI supremacy in every domain. Joel Hron, who leads artificial intelligence and TR Labs at Thomson Reuters, said the goal is narrower: set the frontier of intelligence for legal. The company also leaned hard on human expertise while training. Hundreds of subject-matter experts helped shape the objectives, create legal question examples and judge answers in blind comparisons. The process included alignment with the company’s values, pretraining on Thomson Reuters content, targeted post-training and reinforcement learning tied to tools like Westlaw and Practical Law.
Westlaw is not a toy dataset. Thomson Reuters says the platform includes more than 40,000 individual databases and over 150 years of legal publishing and editorial work. Jonathan Schwartz, who heads foundational research, said the team tried to avoid the usual specialty-model problem, where gaining depth ruins general ability. Internal tests showed Thomson performing broadly in line with leading models when everyone had only web access, then roughly equal or slightly better once Thomson Reuters content was attached. Those results still need more independent validation, and a technical report is still coming.
The company has started sharing Thomson with legal experts and academic institutions for testing. It also plans a smaller open-weight version on Hugging Face under a noncommercial academic license, plus a portal for outside developers to request API keys. Thomson Reuters says only about 10% of its information base has been used so far, and it is already talking to big law firms and corporations about direct access and possible adaptation to their own workflows.
My take — AI-written commentary, not fact-checked reporting
Thomson Reuters is making the obvious grown-up move: stop pretending a generic chatbot knows law because it can write a decent paragraph. The real competition now is not between every model and every task; it’s between companies that own the workflow and those that just rent intelligence from someone else. And yes, the people selling “sovereign AI” will love that phrase right up until the maintenance bill arrives.
Read more about this at: SiliconANGLE