Your AI contract generator just drafted a vendor agreement in three minutes. It includes a 2% royalty cap, a $500K indemnity floor, and mandatory Singapore arbitration—none of which you asked for, all of which you'll now negotiate out. This is not a hallucination in the sense of a minor factual mistake. This is the LLM filling structural gaps in your prompt with invented constraints that feel plausible and legally formal. By the time your legal reviewer flags them, the contract has already circulated to the other side, and you look either incompetent or like you're trying to sneak terms in at the last minute. The three liability gaps that open when AI drafts contracts are predictable, costly, and sealable. But they require a different review workflow than traditional contracts do—one that assumes the AI has not hallucinated facts, but has hallucinated terms . Gap 1: Scope creep gets buried in liability caps AI contract drafts almost always include a liability cap clause—usually $X or Y% of contract value—because that's a standard feature in the template it absorbed during training. What it rarely does is tie that cap to what the contract actually covers. Real example: Your AI draft says "The vendor shall indemnify the Client against all third-party IP claims, capped at $250,000." But your actual contract has three separate modules: software license, data processing, and staff augmentation. The indemnity should cover only the software IP—not the data work, which has different risk and should have different cap. The AI didn't know to ask, so it wrote one cap for three scopes. The cost: You pay the vendor's insurance on a $250K cap when you should be paying on $80K for the one scope that matters. Or worse, you inherit risk on two scopes that have no cap at all because the blanket cap was so tight the vendor won't sign without carve-outs. How to seal it: Before you send the draft out, audit the liability section against your actual scope. If the contract has multiple services, deliverables, or risk zones, write separate liability caps for each. If your AI draft lumps them, rewrite that section by hand. This takes 20 minutes and prevents a week of negotiation. List every distinct service or deliverable in the contract (e.g., API integration, data storage, security audit). For each, note the realistic risk exposure (e.g., API failure = downtime cost; data storage = privacy breach cost). Set the liability cap at a defensible multiple of that (e.g., 2× annual spend on that service). If the AI draft uses one cap for all, replace it before sending. Gap 2: Indemnity clauses invent payment paths that don't exist Indemnity is a contract promise to cover the other side's losses from a specified risk. AI drafts commonly invent how payment flows—"the indemnifying party shall reimburse within 30 days of invoice" or "the indemnified party may offset against future fees." These sound reasonable and never existed in your original brief. The problem: Your risk model assumes indemnity is insurance-backed (the vendor's E&O policy pays) or bonded (the vendor has set aside money). You never meant for it to be a direct invoice-and-pay flow. But the AI draft now commits you to that. Or the reverse: You wanted offset rights (if the vendor breaches, you don't pay them), but the AI draft says all indemnity claims go to arbitration and must be cash-settled, which means you pay them first and maybe get money back in six months. Real example: An AI contract for a freelance developer included this: "Developer shall indemnify Client against IP claims, and Client may offset indemnity amounts against Developer's monthly invoices." But this was a contract for $8K of work over two months. If an IP claim came in on month three, there were no future invoices to offset against. The indemnity was now unenforceable because the payment mechanism didn't match the contract timeline. How to seal it: Before the draft leaves your team, trace the indemnity payment path end-to-end. If the contract ends on date X, but indemnity claims can arrive through date X+2 years (standard tail), your payment mechanism must work on date X+2 years. If you want offset, say so explicitly. If you want insurance only, say that. If you want cash payment, specify the trigger (claim admitted, arbitration award, settlement agreement). Identify when indemnity obligations end (contract end date + tail, usually 2–3 years). Identify the only acceptable payment mechanisms for your business (insurance claim, offset, cash, bond draw). If the AI draft uses a mechanism that won't work on that timeline, rewrite it. Add a single sentence: "All indemnity claims must be submitted by [date]; indemnity obligations expire on [date]." Gap 3: Dispute escalation paths lock you into the wrong venue or timeline When your contract calls for arbitration but doesn't say where, or says "good faith negotiation" but doesn't say for how long, the AI draft fills in invented answers. "Disputes shall be resolv