The legal industry is a useful test case for the rest of the professional world, because it has a very unforgiving feedback mechanism. Get something wrong, and a judge tells you in public.
LexisNexis CTO Greg Dickason offered a rare honest account of how the company manages AI in a domain where errors carry real consequences. The company has around 4,000 engineers building its AI systems in-house, and it tests outputs by having AI models check each other, then having several hundred practising lawyers assess the results. That is a costly, slow process, which is exactly the point. There are no shortcuts when a wrong answer can get an attorney sanctioned.
The core problem Dickason describes is architectural. General AI models are trained on enormous volumes of data and asked to do everything. When you give one a legal question, it draws on patterns across millions of documents and produces a confident-sounding answer. But legal research requires tracing a specific case to a specific citation, which either exists or it does not. That is a task where general AI fails badly.
Stanford researchers tested this directly. General-purpose AI tools hallucinated on legal queries between 69% and 88% of the time. Purpose-built legal tools from the two dominant providers, LexisNexis and Thomson Reuters, still hallucinated more than 17% of the time. That is one wrong answer for every six queries, even from the best available legal AI products.
Courts have been watching. A research database maintained by a Paris academic now tracks over 1,400 cases globally where AI-fabricated content reached a judge. In 2025 alone, researchers documented new hallucinated court filings appearing at a rate of two to three per day. In the United States, over 128 lawyers have faced formal sanctions. Penalties have included being removed from cases, referrals to state bar associations, fines up to $31,000, and orders to notify clients of the errors.
The pattern Dickason highlights is particularly telling. Some of these lawyers were not reckless. They simply used the AI tool they already had open on their computer, the same one they use to draft emails or look up a recipe, and trusted its output for a legal question. That trust is the problem.
LexisNexis's product, now called Lexis+ with Protégé, is built to limit this. It keeps AI grounded in a verified corpus of legal documents rather than the open internet, uses citation validation tools, and offers multiple AI models rather than one. A Forrester study commissioned by the company estimated a 344% return on investment over three years for large law firms, though that figure comes from a company-paid study and should be read as directional rather than definitive.
The broader point for professionals outside of law is simple. AI makes plausible errors, not obvious ones. A made-up legal case looks exactly like a real one. A hallucinated number in a procurement analysis, a fabricated regulation in a compliance review, a wrong insurance clause in a contract summary: these errors have the same shape. They sound authoritative and they are wrong.
The answer is not to avoid AI. It is to use tools built for your specific domain, insist on source citations you can verify, and never treat AI output as final without a human check. LexisNexis is spending hundreds of engineers and hundreds of lawyer-hours making that check happen inside its product. Most general AI tools spend nothing on it.