Most people sign contracts they haven’t fully read. Not because they’re reckless — because the language is deliberately hard to parse, and asking a lawyer to explain a freelance agreement or apartment lease feels like overkill. So you skim it, look for the numbers, and hope nothing weird is buried in section 7(b).
AI is genuinely useful here. Not as a legal advisor — that boundary is real — but as a first-pass translator that turns dense boilerplate into something you can actually reason about.
What You’re Actually Trying to Do
Before you paste anything into ChatGPT or Claude, be clear about what you want. There are roughly three different tasks people conflate:
- Understanding what the contract actually says — translation, not analysis
- Flagging unusual or risky clauses — comparison against a baseline of what’s normal
- Deciding whether to sign — judgment that depends on your specific situation, your leverage, and what else is on the table
AI is strong at the first two. The third is yours to make. The failure mode is asking a model to collapse all three into a single “is this contract okay?” prompt and then treating the answer like advice.
How to Actually Do This
Paste the contract text directly — don’t describe it. Models work better with the actual language than with your paraphrase of it. If the document is a PDF, most current tools (ChatGPT with file upload, Claude with document pasting) can handle it, though for longer documents you’ll want to confirm the full text came through. A 30-page commercial lease can push the limits of what gets processed cleanly in a single context window.
Start with a plain-language summary prompt: “Summarize this contract in plain English, section by section. For each section, tell me what I’m agreeing to in one or two sentences.” This sounds almost too simple, but it works — you end up with a version you can read in five minutes, and you immediately notice the sections where the summary feels vague or hedged. That vagueness is usually a signal the underlying clause is doing something complicated.
Then ask specifically about risk: “What clauses in this contract are unusual, one-sided, or potentially harmful to me as the [freelancer / tenant / employee / buyer]? Focus on anything that limits my rights, creates open-ended financial obligations, or waives something I might want to keep.” In my experience Claude does this pass more reliably than ChatGPT — it’s more likely to surface something quietly buried rather than report that everything looks standard when it doesn’t.
A third prompt worth running: “Are there any terms in this contract that are vague or undefined in a way that could cause a dispute later?” Ambiguity is often where contract problems actually live. A clause that says “reasonable notice” without defining reasonable, or “expenses as agreed” without a ceiling, is a different category of risk than an outright unfavorable term — and a fast skim won’t catch it.
One Real Limitation Worth Knowing
Models don’t know your jurisdiction’s law, and this matters more than people expect. A clause that’s unenforceable in one state or country might be perfectly standard somewhere else, and the model has no reliable way to tell you which side you’re on without knowing where you are and what governing law the contract specifies. Always check which jurisdiction applies — it’s usually near the end, in a “Governing Law” section — and treat any jurisdiction-specific analysis the model offers as a hypothesis to verify, not a conclusion to act on.
Models can also miss things. A contract deliberately obfuscated through defined terms — where section 1 quietly redefines “business day” to mean something non-standard, and section 14 uses that definition to calculate a penalty — can fool a model the same way it fools a fast reader. The model might summarize section 14 accurately without connecting it back to the redefined term. Running separate focused prompts rather than one big “analyze everything” prompt reduces this, but doesn’t eliminate it.
Where This Actually Saves You Something
The most practically useful case isn’t a major commercial deal — it’s the mid-stakes stuff that doesn’t justify a legal bill but still matters. A freelance services agreement. A SaaS subscription with a binding arbitration clause buried in the terms. A lease addendum your landlord slipped in at the last minute. These are documents where “I’ll just trust it’s standard” has cost people real money or options, and where thirty minutes of careful AI-assisted reading beats a three-second skim.
I’ve also run this process from the other direction — pasting agreements I’d written myself before sending them out, to check whether something I intended as protective read as aggressive. A model’s plain-language summary of your own writing is occasionally humbling in a useful way.
A Note on Privacy
If the contract contains sensitive data — someone else’s compensation, proprietary business terms, confidential IP — pay attention to what you’re pasting into a consumer AI product. Both ChatGPT and Claude have settings that affect whether inputs are used for training, but the defaults vary and have changed over time. For anything genuinely sensitive, use a version with data controls enabled, redact the sensitive specifics before pasting, or run it locally if you have the setup for it.
Two prompts do most of the work: section-by-section summary, then what’s unusual or one-sided. Run those before you do anything else — if something needs a closer look, you’ll know exactly where to point a lawyer rather than paying them to read the whole thing cold.
