You prompt Claude with "draft a service agreement" and get 47 lines back in ninety seconds. It looks coherent, uses legalese, and reads like something a junior associate might have written. Then your accountant flags a missing limitation-of-liability cap. Your IP lawyer points out that ownership of work product—the thing you're actually delivering—is vague enough to start a lawsuit. And buried in the payment terms is a clause that lets the client withhold invoices indefinitely if they dispute quality. AI contract drafting tools are remarkable at speed and structure. They fail catastrophically at the details that actually protect you. We tested Claude 3.5 Sonnet , ChatGPT-4 , and Harvey AI (a specialized legal model) against seven non-negotiable contract clauses. The results were clear: AI hallucinates risk away. The problem: speed masquerades as competence Contract generation feels like a solved problem because the output looks professional. But "looks professional" is the opposite of protective. A well-formatted contract with a missing cap on liability exposure can cost you six figures. A payment-terms clause that's ambiguous about late fees kills cash flow. A non-compete that's too narrow or too broad is either worthless or unenforceable. The deeper issue: AI language models optimize for plausibility and completeness of the output , not correctness of the legal substance. They don't "know" what a good liability cap looks like in your jurisdiction. They don't flag red flags. They smooth over vagueness because vagueness is linguistically coherent. AI drafts the words. It does not draft the risk. Clause 1: Limitation of liability—AI softens the floor Every service contract needs a cap on what either party can owe. Without it, a small service failure becomes an existential bill. What we tested: We asked each tool to draft a B2B SaaS agreement and specify a liability cap tied to 12 months of fees paid. Claude 3.5 Sonnet: Generated a cap, but added a carve-out for "indirect or consequential damages"—standard language—and then failed to define what "consequential" means. In practice, courts interpret this broadly. The cap became porous. ChatGPT-4: Capped liability at "the fees paid in the preceding 12 months." Then, three clauses later, added an exception for "breach of confidentiality or data protection obligations." It had created two liability regimes without explaining which applied when. A lawyer has to rewrite this. Harvey AI: Generated a clean cap tied to fees, but did not address whether the cap applies to both parties equally or asymmetrically (common in contracts where one party bears more risk). The clause was technically sound but incomplete for real-world negotiation. The risk: If your contract doesn't explicitly state that liability caps apply to both parties, and don't exclude categories of loss, you've bought a liability shield that only works one way. Clause 2: IP ownership and work product—AI hedges When you deliver a deliverable—a report, custom code, design, strategy—who owns it? AI drafts this badly because ownership questions have no single right answer. The model averages across thousands of contracts and produces language that offends everyone. What we tested: "Draft a statement of work for a digital marketing agency. Who owns the campaign strategy and creative assets?" Claude: "The Client shall own all work product. The Provider retains ownership of pre-existing materials and methodologies." Sounds fair. But "methodologies" is undefined. Does it mean the framework? The templates? The actual campaign strategy for this client ? If the client hires a competitor next year and uses the same framework, can you claim infringement? Probably not—the clause is too vague to hold. ChatGPT-4: Went the other direction: "The Provider retains ownership of all deliverables. The Client receives a non-exclusive, royalty-free license to use the deliverables for their business." This is terrible for a client paying for custom work. But it's also bad for the agency because "license to use" doesn't specify duration or scope. If the client goes bankrupt, do they keep the license? Forever? The clause invites litigation. Harvey: Split the difference: "Ownership of custom deliverables shall vest in the Client. The Provider retains ownership of pre-existing tools, templates, and IP." This is clearer, but doesn't address whether the client can resell the deliverable or use it for client work with a third party. A real IP clause needs to say. The risk: Vague IP clauses become disputes. You think you own your methodologies; the client thinks they own everything because they paid. Lawyers charge ₹2–5 lakhs to clarify what should have been two sentences. Clause 3: Payment terms and dispute withholding—AI ignores cash flow Net-30, Net-60, due on delivery—all have teeth only if you define what happens when the client disputes an invoice. AI drafts payment terms as if disputes are rare edge cases. What we tested: "Draft pa