How to do your own legal work with AI

ChatGPT is too agreeable for a lawyer. It fails to challenge and probe. It's too nice. Inhouse has handled over 7 million legal chats, many of which gets pulled in from ChatGPT. What do our lawyers most often find wrong with generic AI?

It's not a hallucination. Wrong like: here's your employment agreement when you needed an offer letter. Here's your phantom stock plan when you needed a simple bonus structure. Here's your cease and desist letter when you have no case. Users, untrained in the law, can't tell what's wrong.

A 2024 Stanford study tested GPT-4 on 800,000 verifiable legal questions and found it was wrong 58% of the time. A 2025 Cambridge study found that people trust ChatGPT's legal advice more than a real lawyer's — when they don't know which is which. High error rate. High confidence. High trust. That combination is dangerous in any domain. In law, it can cost you your business.

Generic AI doesn't think like a lawyer. It can't.

Here's the thing most people don't realize: a lawyer's value is not primarily in drafting documents. It's in everything that happens before drafting. What is your actual goal? Is this the right approach? What facts are you missing? What does the other side look like? What are you about to do that you'll regret?

Generic AI skips all of that. It receives a request and executes it. It will not tell you the document you asked for is the wrong document. It will not tell you your case is weak when you're angry and want validation. It will not ask what state you're in, whether you're at-will, whether your non-compete is enforceable, or whether what you're about to send will make your situation worse.

This is not a bug they're fixing. It's structural. ChatGPT is a general-purpose language model optimized for engagement. It is not trained on how lawyers actually think through problems. It has no mental model for the questions that need to come before the answer.

The consequences are showing up in court. In March 2026, Nippon Life sued OpenAI in federal court — Nippon Life Ins. Co. of Am. v. OpenAI Fdn., 1:26-cv-02448 (N.D. Ill.) — after a woman asked ChatGPT if her lawyer was gaslighting her about a settled case. ChatGPT said yes. She fired her lawyer. Used ChatGPT to draft a motion to reopen. Filed 44 subsequent motions and petitions — one citing a case that does not exist. Lost. Filed a new lawsuit. The insurer eventually sued OpenAI for unlicensed practice of law, seeking $10 million in punitive damages. OpenAI quietly updated its terms of service in October 2025 to prohibit using ChatGPT for legal advice without a licensed professional's involvement. The model didn't change. The liability just shifted to you.

The one thing that makes generic AI dramatically less dangerous for legal work

Stop asking for what you want. Describe everything about your situation and let the AI figure out what you need.

This sounds small. It isn't.

When you say "draft me an employment agreement," you get an employment agreement. When you say "I'm hiring my first full-time employee in Texas, she's going to run my front desk, I'm paying her $52,000, I want to make sure I'm protected and that this is done right — what do I actually need and what should I be thinking about?" — you get a completely different conversation. One that might tell you an offer letter is the right document, not an employment agreement. One that might surface the at-will issue you didn't know to ask about.

The reason this works is simple: a lawyer's job is to figure out what's missing from your picture. Generic AI won't do that unless you force it to by describing so much context that the gaps become obvious. Give it your goal, your situation, your state, your relationship to the other party, what you've already done, what you're afraid of, and what outcome you want. The more you give, the less it can assume. The less it assumes, the less it invents.

A few examples of how this plays out in practice:

Instead of "draft a cease and desist letter to my former employee" — tell it everything: who the employee is, what they did, what your non-compete says, what state you're in, how long ago they left, what harm you've actually suffered. Then ask: do I have a case, what are my options, and what are the risks of sending a letter versus doing nothing versus going to court?

Instead of "create a phantom stock plan" — describe the whole picture: how many employees, what stage your company is at, what you're trying to accomplish with the equity, whether you've raised money, what your cap table looks like. Then ask: is this the right structure for what I'm trying to do?

The goal is not to give the AI a task. The goal is to give it your situation and make it think — knowing it will still miss things a lawyer would catch, but reducing the damage from what it doesn't know to ask.

What Inhouse does differently

The Playbooks on this site are not just articles. They are the legal mental models that our AI is trained on — the decision frameworks that a lawyer applies before picking up a pen. What document do you actually need? When is a TRO better than a cease and desist? When should you not send that letter? When does an offer letter protect you more than an employment agreement?

Generic AI doesn't have these models. Inhouse does — and gets sharper every time a lawyer reviews an output and corrects it. That feedback loop is how legal judgment improves over time, not just text generation.

When you describe your situation to Inhouse, it asks the questions a lawyer would ask before producing anything. It doesn't wait for you to know what to ask. It already knows what's usually missing — because it's been trained on thousands of legal situations and the attorneys who worked through them.

What to tell Inhouse: Don't ask for a document. Describe your situation completely — your business, your state, what happened, what you're trying to accomplish, and what you're worried about. Let Inhouse figure out what you need.

This article reflects the author's experience as a practicing attorney and founder of Inhouse AI. General information only, not legal advice. AI accuracy data reflects conditions as of June 2026. The Nippon Life v. OpenAI litigation is in early stages; consult the court record for current status.