You hit 'generate contract' on an AI tool and get back a 12-page document. It looks coherent. It has sections. It feels real. Then your lawyer reads clause 7 and says: where did this pricing formula come from? Nowhere. The LLM invented it. AI contract generators are useful for speed, but they are aggressively unreliable on liability, indemnity, tax, and asset transfer terms. They hallucinate numbers, invent obligations, and skip jurisdiction-specific rules. In Indonesia especially—where stamp duty, asset transfer clauses, and notarization create real legal friction—an AI draft can expose you to rejection, renegotiation, or enforcement failure. This playbook identifies seven high-risk clauses where LLMs consistently fail, shows what they typically get wrong, and gives you safe rewrites to use before your lawyer ever touches the document. This is not legal advice; it is an audit checklist. Why LLMs fail on contract terms Large language models are trained to produce plausible text, not accurate legal terms. They see patterns in training data and reproduce them, but they don't validate whether a term is actually legal, fair, or applicable to your deal. Key failure modes: Hallucinated pricing: The model invents discount percentages, volume thresholds, or payment schedules that never appeared in the prompt. Generic indemnity: Standard indemnity clauses are copy-pasted without party-specific obligations or carve-outs, leaving you liable for events you shouldn't cover. Jurisdiction gaps: Tax clauses, asset transfer rules, and notarization requirements are often omitted or stated incorrectly because the model trained on primarily US or UK contracts. Circular liability: Liability caps are set but not tied to actual insurance, leading to unrecoverable loss scenarios. Vague force majeure: 'Acts of God' is broad; 'pandemics' is undefined; 'supply chain disruption' is infinite. LLMs rarely narrow this. Missing termination triggers: The model includes termination clauses but omits survival language, confidentiality tails, or data return obligations. Undefined definitions: Key terms like 'deliverable', 'acceptance', or 'material breach' are used but not defined, creating ambiguity. Clause 1: Payment and pricing—where LLMs invent numbers AI drafts typically produce something like: "Invoices shall be paid within 30 days. Volume discounts apply: 5% at 100 units, 10% at 500 units, 15% at 1,000 units. Late payments incur 2% monthly interest." The problem: You never said 'volume discounts' or 'monthly interest'. The model hallucinated both to sound plausible. If the buyer disputes an invoice, that invented term becomes leverage against you. Safe rewrite: "Pricing is as stated in Schedule A. Invoices are issued upon [delivery / acceptance / completion]. Payment terms are net [30/45/60] days from invoice date. No volume discounts or late-payment interest apply unless explicitly agreed in writing and countersigned." This removes the hallucination, pins pricing to a schedule you control, and requires written amendment for any deviation. Test it: ask the AI to draft only payment terms, not pricing, and supply pricing separately in a schedule. Clause 2: Indemnification—asymmetric risk you didn't ask for LLMs often produce one-way or over-broad indemnity: "Vendor shall indemnify, defend, and hold harmless Client from any and all claims, damages, losses, and liabilities arising from Vendor's performance, non-performance, or any act or omission relating to this Agreement." This is weaponized. It covers Client's own negligence, third-party actions, and market changes—not just Vendor's fault. Vendor is now liable for Client's injury during a site visit, even if Client ignored safety warnings. Safe rewrite: "Each party shall indemnify the other for third-party claims arising solely from that party's breach of this Agreement or its negligent performance. Indemnification does not apply to: (a) claims arising from the other party's use of deliverables outside the scope agreed; (b) Client's modification of deliverables; (c) claims arising from Client's negligence or violation of law." This is mutual, narrowly scoped, and carved out for Client misuse. In Indonesia, include: "Indemnification excludes claims covered by mandatory statutory liability caps under Indonesian contract law." Clause 3: Limitation of liability—disconnected from actual risk Common LLM output: "Neither party's total liability shall exceed the fees paid in the past 12 months, provided that liability for gross negligence or willful misconduct is unlimited." The trap: If you charged ₹5L/year and the contract causes ₹50L in Client loss, you pay ₹5L cap. But the carve-out for 'gross negligence' is undefined—Courts might argue ₹50L mistake qualifies, voiding the cap. Safe rewrite: "Except for indemnification and confidentiality, neither party's cumulative liability exceeds [2x annual fees / fixed ₹XYZ amount]. Liability for breach of confidentiality or willful misconduct is unlimi