Legal AI’s Strongest Protections Go to the People Who Least Need Them

Harvey, the legal AI company, raised $200 million in March at an $11 billion valuation. Part of what investors are paying for is trust. Harvey does not use customer data to train its models. It deletes prompts, outputs, and documents on a schedule each customer sets, down to zero days. It lets firms pick the country where their data lives. A recent review by data protection lawyer Mahdi Assan called Harvey an example of a company getting AI data governance right, and the evidence backs him up.

Look at who gets that protection, though. Harvey works with most of the 100 largest law firms in the country, more than 500 in-house legal teams, and Fortune 500 companies. It publishes no prices, and there is no way in without a sales call. Buyers report seat costs starting around $1,000 a month, with minimum purchases of 25 seats or more. There is no scandal in any of that. Enterprise software works this way. It does mean the strongest data protections in legal AI belong to the clients who already have the best legal protection money can buy.

Now look at who is actually walking into court with AI. Researchers at MIT and USC examined 4.5 million federal civil cases and found that people representing themselves rose from a steady 11% of cases, a share that held for almost twenty years, to 16.8% in fiscal year 2025. Self-represented people filed roughly 23,000 federal cases in 2022 and roughly 41,000 in 2025. The share of complaints showing markers of AI-generated text climbed from near zero to more than 18%. The study is a working paper, its detection method has limits, and it cannot prove on its own that AI caused the surge. The direction is still hard to argue with. People who cannot afford lawyers found a tool that writes legal documents for free, and they are using it.

Here is what those people get in place of Harvey’s protections. What you tell a lawyer is privileged, meaning a court usually cannot force it into evidence. What you type into a chatbot can become a company record. Sam Altman said as much last summer, warning that there is no legal confidentiality when people bring their problems to ChatGPT, and that OpenAI would have to turn conversations over under a subpoena. That risk has already arrived. A federal court ordered OpenAI to preserve user chats as potential evidence in the New York Times copyright case, including chats people believed they had deleted.

When the free tool invents a case citation, the person who filed it takes the hit alone. Courts have dismissed cases, tossed appeals, and ordered self-represented people to pay the other side’s fees over AI-fabricated citations, and defense firms now publish playbooks for attacking AI use by self-represented plaintiffs.

Both tiers are built on the same families of models. Harvey is built on models from OpenAI, Anthropic, and Google. So is the $20 chatbot. What separates them is the wrapping. Harvey proved these protections can be built. The market then priced the fullest version of them for enterprise buyers. Cheaper legal AI tools exist, but the strongest protections sit at the top of the market, attached to clients who were never going to lose a case over a subpoenaed chat log. Meanwhile the Legal Services Corporation found that low-income Americans get no legal help, or not enough, for 92% of their serious civil legal problems, and Stanford researchers count an unrepresented party in roughly three quarters of civil cases. Nobody using a free chatbot to fight an eviction turned down Harvey first. The alternative was nothing.

None of this makes Harvey the villain. The company built for a market and did governance work many AI companies skip. Some judges even say AI-drafted filings are easier to rule on than the handwritten ones they replaced. But when data protection becomes a product feature, it flows toward whoever can pay for it, and the people carrying the most personal risk into court end up with the least protected tools.

Assan’s review called trust one of the few real differentiators left in AI. He’s right, and that is exactly the problem. A differentiator is something most people do not get. The protections exist now. What nobody has worked out is how they reach the person arguing for their home alone, or whether trust in AI stays a luxury good.