Reading your subcontract with ChatGPT: what it catches and what it misses

At almost every event we attend, someone tells us they already run their contracts through ChatGPT. That's a sensible instinct. A free first pass beats signing a 40-page subcontract unread. But it helps to know what a general-purpose chatbot does well, and where it's likely to steer you wrong.

What it's good at

• Summaries. "What are my payment terms, and when can they terminate me?" gets a quick, readable answer.
• Finding clauses. It can point you to the indemnity, the notice provisions and the insurance exhibit in a long document.
• Explaining jargon. "Primary and non-contributory" or "flow-down" in plain English.
• Preparing questions. It can help you turn a vague worry into a specific question for the GC or your attorney.

Where it falls short

• State law. Contracts don't say which of their clauses your state won't enforce. A multi-state GC form may include a pay-if-paid clause that California courts generally won't enforce, or an out-of-state venue clause that California law voids for local subs. A chatbot reading the document alone may treat both as if they'll stick.
• Priorities. It tends to flag everything. A list of 25 "concerns" doesn't tell you which two are worth spending your leverage on, or what the GC will actually accept.
• Your situation. It doesn't know your margin on the job, your history with this GC, what your insurance covers, or what's normal for your trade.
• Confident mistakes. Chatbots can state things that sound authoritative and are wrong, including citations to cases that don't exist. Lawyers have been sanctioned for filing briefs with invented cases a chatbot produced. Treat any legal citation it gives you as unverified.

Mind what you paste

• Confidentiality. Your subcontract may itself be confidential, and it often contains pricing and project details. Check before you upload it anywhere.
• Training settings. On consumer ChatGPT accounts, you can turn off "Improve the model for everyone" in Settings, under Data controls, so new conversations aren't used to train OpenAI's models. Temporary chats aren't used for training either. Even with training off, a conversation you rate with a thumbs up or down may be used, so skip the rating on sensitive chats. OpenAI says it doesn't use content from ChatGPT Business, Enterprise or Edu workspaces for training.
• Redact what you don't need. Names, prices and addresses usually aren't needed to understand a clause.

How to use it well

1. Ask specific questions: "List every deadline in this contract that, if I miss it, I lose a right to money or time."
2. Ask it to quote the clause it's relying on, then read that clause yourself.
3. Tell it the state and your role: "I'm an electrical subcontractor on a private project in California."
4. Use the answers to build a short list, then have someone accountable confirm the two or three changes worth asking for.

A chatbot can get you to the right questions faster. The answers, especially the ones that decide whether you get paid, still need someone who'll stand behind them.