Customizing models for legal professionals
OpenAI
OpenAI and legal-AI startup Harvey built a custom-trained model just for lawyers. It's OpenAI's push into bespoke models for specific professions, not just one-size-fits-all GPT.
Based on reporting by OpenAI — 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
Harvey has been selling AI tools to law firms for a couple of years now, mostly by wrapping OpenAI's general models in legal-specific workflows. That arrangement just changed. OpenAI and Harvey have now built a model trained specifically on legal work, rather than a generic GPT fine-tuned after the fact with prompts and retrieval tricks.
The distinction matters more than it sounds. Law firms deal with dense, precedent-heavy text where a wrong citation or a misread clause isn't a quirky hallucination, it's a malpractice risk. Generic models trained on the open internet pick up plenty of legal language, but they don't necessarily learn the reasoning patterns lawyers actually use: how to weigh conflicting case law, how to structure a memo the way a partner expects, how to flag the one clause in a 200-page contract that actually matters.
Harvey has built its business around clients like major law firms and in-house counsel teams handling due diligence, litigation research, and contract review. A model trained with that domain in mind, using real legal workflows and presumably real (anonymized) legal documents as training signal, should in theory produce fewer of the confident-but-wrong answers that have embarrassed lawyers who trusted chatbots without checking citations.
This also tells you something about where OpenAI thinks its business is heading. Rather than just selling API access and letting companies bolt on their own tooling, OpenAI is now co-building custom models with specific industries. Legal is a logical first stop: it's a market with deep pockets, high stakes, and a workforce that bills by the hour, meaning any real efficiency gain translates directly into money saved or made. Expect medicine, finance, and other regulated fields to get the same treatment before too long.
My take — AI-written commentary, not fact-checked reporting
I think this is OpenAI quietly admitting that general-purpose models hit a ceiling in high-stakes professions, and that's a healthy admission. Vertical, custom-trained models trained with domain experts in the loop are going to matter more than the next general benchmark score, and the firms betting on that now will look smart in three years.
Read more about this at: OpenAI