# The Most Important Feature of a Legal AI Might Be Knowing When Not to Answer > Source: > Published: 2026-08-17 20:20:14+00:00 One of the things that gets overlooked when building AI for legal use is uncertainty. Most AI products are designed to answer questions. Legal questions are different. A person might type: *“I got hurt at work. Can I sue my employer?”* That sounds simple, but it isn't. What happened? Where did it happen? Was the person an employee or contractor? Was another company involved? Was the injury caused by equipment? When did it happen? Has a workers' compensation claim already been filed? A few missing facts can completely change the legal analysis. This is why I think legal AI should not be judged only by how many questions it can answer. It should also be judged by how well it handles questions that it cannot safely answer yet. ** ** Large language models are very good at producing something that sounds complete. That is useful in many applications. It can also be a problem in legal applications. If the system does not have enough information, generating a polished paragraph can create a false sense of certainty. The better approach is sometimes to stop and ask another question. Not: “Here is what you should do.” But: “Before I can give you a useful answer, I need to know whether another party was involved.” That small difference matters. For Jimmy Knows, the idea is to start with what actually happened. A user can describe their situation in their own words instead of trying to figure out which legal category their problem belongs to. The system can then connect the facts to Massachusetts statutes and case law and explain the relevant rule in plain language. That creates a different design philosophy: **Don't make the user speak like a lawyer. Make the system understand the user's story.** This also means the AI needs to recognize when the story is incomplete. For example: “I slipped outside a store and got hurt.” There isn't enough information yet. Where did it happen? Was there snow or ice? Was there a visible hazard? Did the store know about it? Those questions are not just conversational filler. They can determine which legal rules are relevant. ** ** The hard part isn't necessarily generating the answer. It is deciding whether an answer should be generated at all. That means a legal AI system needs something beyond retrieval and generation. It needs a concept of **sufficient facts**. You can think of the workflow like this: ``` User tells their story ↓ Extract relevant facts ↓ Identify missing facts ↓ Are the facts sufficient? ↙ ↘ No Yes ↓ ↓ Ask questions Retrieve authority ↓ Explain rule ↓ Explain next step ``` This is a much more interesting problem than simply connecting an LLM to a legal database. The quality of the system depends heavily on what happens before the final answer is written. ** ** There is a tendency to think a good AI assistant should always sound confident. I think the opposite is often true for legal applications. A trustworthy system should be comfortable saying: “I don't have enough information yet.” That doesn't make the system less useful. It makes the next question more useful. And eventually, that can lead to a much better answer. That's one of the ideas behind Jimmy Knows: make Massachusetts legal information easier to access while keeping the distinction between legal information and advice clear. The platform provides information based on Massachusetts authority and connects people with an attorney when they need one. The future of legal AI probably isn't about making AI sound more like a lawyer. It may be about making AI better at knowing when the facts are not there yet. And honestly, that might be one of the hardest parts to get right.