Every founder or lawyer who's used ChatGPT or Claude to sketch a contract knows the feeling: the draft looks solid, reads smoothly, and covers the obvious ground. Until a real lawyer reads it and says, 'Did you see this indemnity clause? It exposes you to everything.' LLMs are pattern-matching machines. They've absorbed thousands of contract templates, precedents, and legal writing samples. But contracts aren't about statistical likelihood—they're about boundary-setting under stress. A mediocre clause that works 95% of the time is worthless on the day you need it. Here are seven systematic gaps that appear in nearly every AI-drafted contract, why they happen, and what a safe rewrite looks like. Use this as a checklist before your lawyer sees the draft—and before you sign. 1. Indemnity scope creeps into uncontrollable territory The problem: AI drafts indemnity clauses that are grammatically sound but legally one-sided. The pattern is always the same—broad trigger, vague scope, no carve-outs. What ChatGPT typically produces: "Each party shall indemnify the other against all claims, damages, liabilities, and expenses arising from any breach of this Agreement." This sounds protective. It's actually a trap. 'All claims' includes things the indemnifying party can't control—third-party lawsuits over quality, performance failures caused by the other side's input, even patent disputes that predate the contract. Why it happens: LLMs see 'indemnity' and 'protection' as synonymous, so they build maximum scope. They don't model the negotiation or the day you're actually paying the defense bill. Safe rewrite: "Each party shall indemnify the other against third-party claims arising from (i) that party's material breach of this Agreement, or (ii) that party's gross negligence or willful misconduct, provided the indemnified party: (a) gives prompt notice; (b) grants sole control of defense to the indemnifying party; and (c) does not settle without the indemnifying party's consent. The indemnifying party has no obligation for claims arising from the indemnified party's breach, misuse, or modification of materials provided." This narrows scope to actual fault, requires notice and cooperation, and carves out contributory breaches. It still protects both sides—but only against what they actually caused. 2. Limitation of liability has no teeth or is unilaterally protective The problem: AI either omits liability caps entirely, or drafts them as unidirectional protections that a court will void as unconscionable. What ChatGPT typically produces: "Neither party shall be liable for indirect, incidental, special, or consequential damages. Each party's total liability shall not exceed the fees paid in the preceding 12 months." The second sentence sounds balanced. But read closely: if there are no fees (free trial, setup phase), there's no cap. If one party pays $10K and the other pays $500K, they have vastly different exposure. And 'fees paid' is ambiguous—does it include setup fees, support, overage charges? Why it happens: LLMs default to symmetrical-sounding language without modeling asymmetric commercial relationships. They also don't distinguish between absolute caps (the same for both parties) and proportional caps (tied to what each paid). Safe rewrite (for a SaaS vendor and customer): "Except for (a) breaches of confidentiality, (b) infringement claims, or (c) gross negligence, neither party's liability arising from this Agreement shall exceed the lesser of (i) the actual damages awarded by a court, or (ii) the total fees paid by Customer in the 12 months preceding the claim. In no event shall either party be liable for lost profits, lost data, lost revenue, or consequential damages." This carves out the stuff you actually can't cap (confidentiality, IP), sets a dollar ceiling tied to economic reality, and explicitly eliminates consequential loss. It's mutual—both sides get the same protection—but it doesn't force a $10M vendor to cap liability at $1K fees. 3. IP ownership clauses ignore jurisdiction and work-for-hire nuance The problem: AI assumes IP ownership can be assigned globally with a single sentence. Reality: Indonesia, Singapore, Malaysia, and the UK have different work-for-hire rules, moral rights, and assignment requirements. What ChatGPT typically produces: "All intellectual property created under this Agreement shall be owned by [Client]. The [Contractor] hereby assigns all rights, title, and interest in any work product to [Client]." This reads like a complete IP transfer. In Indonesia, it's half-baked. Indonesian law (Law 28/2014 on Copyright) recognizes moral rights that can't be assigned—only licensed. In Singapore, work-for-hire doesn't exist in statute the way it does in the US; you must assign rights explicitly, and even then, moral rights may survive. An AI-drafted clause that omits these nuances leaves both parties uncertain about what's actually owned. Why it happens: LLMs train on US and UK templates,