Last week Thomson Reuters launched Thomson, its first proprietary large language model. Trained on Westlaw and Practical Law content, built over two years for $40 million, with a final training run costing just $450,000. It's believed to be benchmarking competitively against Claude Opus 4.8 and ahead of GPT-5.5, Claude Sonnet 5 and Gemini 3.1 Pro on legal tasks. Harvey launched a similar proprietary model, Tenet, just days earlier.
That’s the news. The headline reason companies are doing this is cost and control. Frontier labs spend billions building general purpose models. Thomson Reuters took a different route. Start with a strong open source foundation, then specialise it deeply on your own proprietary content. Cheaper, faster, and fully owned.
But I think the more interesting story here isn't the cost saving. It’s the control part, more importantly what ownership actually means once a company controls the model end to end.
Every LLM reflects the values baked into what it was trained on. This is a well documented pattern across a large body of academic research. Studies comparing LLM outputs to real world survey data across dozens of countries consistently find models overemphasise moral concerns common in Western societies and underrepresent values more prominent elsewhere. Some models adapt better to certain cultural contexts than others, and not always in the direction you'd expect. The point is, no LLM is neutral. It carries the imprint of whoever built it and whatever it learned from.
Right now, most companies use models built by someone else. OpenAI, Anthropic, Google. That comes with real dependency, but it also comes with a kind of external check, meaning there is information and scrutiny happening outside of any one company’s proprietary data. Obviously there are values and behaviours baked into frontier models too but they weren't decided by a single company for its own commercial purposes (with the exception of the labs building these models and shaping them directly).
Once more companies own their model outright, that check disappears. They control what it's trained on, how it's fine tuned, what it emphasises and what it leaves out. Thomson Reuters framed this as a win for accuracy and depth in legal work, and this may very well be true.
But the same capability that lets a company build a more accurate, more specialised model also lets it build a model that reflects exactly the view of the world it wants that model to have. Nothing about owning your own AI requires that outcome. But nothing about it prevents it either.
A fair counterpoint here is that this isn't new. Every company website is already biased. Nobody reads a "Why choose us" page expecting neutrality. So why would a company's own AI be any different?
The difference is in how people actually treat the two. Nobody mistakes a marketing page for objective analysis, and they read it with the scepticism it deserves. AI output doesn't get read the same way. Research on what's called automation bias shows people consistently over-rely on AI generated answers and are less likely to question or correct them, often treating the system's output as more credible than their own judgement. One study following nearly 2,800 participants found people were markedly less likely to catch an error when it was labelled as coming from an AI system. Another found that simply using a chatbot to help write something shifted not just the words people used, but the opinions they actually held.
Same underlying bias, completely different level of trust. That's what changes once a company owns the model end to end. Not that the bias is new, but that it now arrives dressed as reasoning instead of persuasion.
So my question is, as more companies follow Thomson Reuters and Harvey down this path, and proprietary models become the norm rather than the exception, who is actually checking what these models have been shaped to emphasise, and what they've quietly been trained to leave out?
Key sources referenced:
Thomson Reuters, "Thomson Reuters Leverages its World-Class Data Assets to Launch Its Own Frontier Model", 24 August 2026
LawSites, "Thomson Reuters Launches Thomson, Its Own Proprietary LLM Trained on Westlaw and Practical Law Content", August 2026
Thomson Reuters Institute, "Thomson Reuters Built Its Own AI Model That Now Ranks Among the World's Best"
PNAS, "Moral stereotyping in large language models", March 2026
International AI Safety Report 2026, "Automation bias persists with new AI tools"