You ask an LLM to draft a service contract. It returns something that looks complete: boilerplate language, reasonable structure, even a signature block. You skim it, your client signs, and six months later your accountant flags a $40,000 liability cap that doesn't exist. Your termination clause is silent on early exit fees. Your IP indemnity promises to defend against claims that were never your responsibility to defend. The LLM didn't malfunction. It did what it does: it hallucinated. Large language models are confident confabulators. They generate plausible text without understanding law, jurisdiction, or consequence. They invent pricing terms that contradict your rate sheet. They skip entire sections because they don't appear in their training data often enough. They blend standard language from incompatible contract types. And they almost never, ever flag what they don't know. This is not a reason to abandon AI contract drafting. It's a reason to understand exactly where LLMs fail, and to build a human review checklist that catches those seven gaps before a signature hits the document. Gap One: Indemnity Language Gets Overwritten or Invented Indemnity clauses are where LLMs most consistently hallucinate. They will: Invent mutual indemnity that contradicts your commercial position (especially if the prompt is vague) Promise you'll defend the other party against their own negligence—uninsurable exposure Omit carve-outs for their sole negligence, creating unlimited liability Borrow IP indemnity language from software contracts and paste it into a services agreement, where it makes no sense Why this happens: Indemnity clauses are dense, context-dependent, and vary wildly by jurisdiction and deal type. LLMs have seen thousands of indemnity templates in their training data, but they have no idea which one applies to your situation. They pick the longest or most "complete" looking one and adjust it just enough to seem responsive. Your checklist: For every indemnity clause, verify: Who is defending whom? (You should defend them for their claims, they should defend you for yours.) What triggers indemnity? (Their claim, your breach, a third-party lawsuit? Be explicit.) Are carve-outs clear? (They don't get indemnity for their own negligence.) Is there a cap on indemnity? (Usually mirrors your liability cap, unless IP is at stake.) Does the indemnified party cooperate in defense? (They should. If not, they're gaming you.) Gap Two: Termination Rights Are Silent or Contradictory Watch for these: No early termination right at all, or one so narrow it's useless (e.g., "only if there's a material breach and the other party fails to cure in 60 days") Asymmetric termination: you can terminate at will, they need cause. Or vice versa. Termination triggers that contradict your commercial terms (contract says "one-month notice," but invoice cycle is quarterly) Missing language on what happens to fees, deliverables, or confidential data after termination No termination for convenience, so you're locked in even if priorities shift Why this happens: LLMs often copy termination language from short-term SaaS contracts into long-term service agreements, or vice versa. They default to "mutual" termination rights (which sound fair but often aren't) because that language is common. They rarely include the operational details—what happens to in-progress work, who pays for wind-down, how data is returned. Your checklist: Verify: Can you terminate at will, for convenience, or only for cause? What notice period applies? (Match it to your business cycle.) Do termination fees apply? (Proportional to work completed, or all-or-nothing?) What happens to confidential data, deliverables, and work-in-progress? Is there a survival period? (IP, confidentiality, indemnity usually survive termination.) Are both parties bound equally, or is there an asymmetry you need? Gap Three: Liability Caps Are Invented, Missing, or Miscalculated LLMs will: Insert a liability cap (e.g., "limited to fees paid in the last 12 months") that contradicts your insurance policy or actual exposure Omit caps entirely, leaving you liable for unlimited consequential damages Cap your liability at a dollar amount but leave theirs unlimited Include a cap that's technically present but mathematically impossible (e.g., "capped at 100% of fees" when fees are $0) Forget that certain clauses (indemnity, confidentiality, IP) often sit outside the main liability cap Why this happens: Liability caps require math and judgment. LLMs generate plausible numbers without understanding your risk or your insurable limits. They often default to mutual, symmetrical caps (you're both limited to the same amount) because that language is common, even when asymmetry makes sense for your deal. Your checklist: What is the cap? (Fees paid, ARR, a fixed dollar amount, or unlimited?) Does it match your insurance limits? What sits outside the cap? (Indemnity, IP, confidentiality, or data breaches often d