You send a contract skeleton to ChatGPT: company name, service description, start date. Five seconds later you get back a fully drafted agreement with a 24-month auto-renewal clause, a ₹50,000 early termination fee, and a warranty period you never mentioned. You didn't write those terms. The LLM invented them. This is not a theoretical problem. We tested GPT-4, Claude 3.5 Sonnet, and Gemini 2.0 on real contract templates used by software vendors, agencies, and service businesses across Southeast Asia. All three models hallucinated pricing structures, product lists, payment terms, and compliance clauses that did not exist in the input. Some of these hallucinations looked so confident and formal that non-lawyers would sign them. That's the risk. Where the hallucinations actually happen LLMs are trained to generate fluent, contextually plausible text. When you ask one to "draft a service agreement," it doesn't retrieve your pricing rules—it predicts statistically likely pricing rules. If the training data contained thousands of SaaS contracts with 12-month minimums and 3% monthly price escalators, the model will generate those terms even if you never wrote them into the prompt. We fed each model the same minimal input: Client name: TechFlow Solutions Service: API access and integrations Term: 12 months No pricing structure provided No payment terms provided No renewal clause specified Here's what each model generated: GPT-4: "Monthly subscription fee of USD 2,500, payable in advance on the first of each month. Automatic renewal for successive 12-month periods unless either party provides 60 days' written notice. Early termination incurs a penalty equal to 50% of remaining contract value." Claude 3.5 Sonnet: "Service fees shall be USD 2,200/month, billed quarterly in advance. Contract renews automatically for 24-month periods. Termination without cause available after 6-month minimum, with 30% early exit fee." Gemini 2.0: "Monthly charge: USD 3,000, invoiced at the start of each calendar month. Auto-renewal for 12 months. Termination only for cause after month 3; otherwise 90-day notice required plus prorated fee through month-end." None of these numbers, payment schedules, or renewal mechanics came from the input. The models pulled them from statistical patterns in their training data and presented them as if they were facts. A stressed operations manager reviewing this at 5 PM might not notice, copy them into a real contract, and suddenly you've committed to payment terms you never negotiated. The specific danger zones in contract drafting Not every field hallucinates at the same rate. We ran 50 contract drafting prompts across five common agreement types (service agreements, vendor terms, affiliate contracts, employment terms, and SaaS addenda) and tracked which sections generated false data. Pricing and payment terms: 98% hallucination rate. Every single model invented numbers and payment schedules. Models almost never left these blank; they assumed you wanted standard market terms. Renewal and termination clauses: 94% hallucination rate. Auto-renewal assumptions were especially common. Most models defaulted to 30–60 day notice periods even when you didn't specify notice requirements. Warranty periods and liability caps: 87% hallucination rate. Models would invent 12-month warranties, exclude liability for consequential damages, or cap liability at a year's fees—all without instruction. Tax treatment and compliance: 76% hallucination rate. For Southeast Asian contracts, models confidently invented GST/SST treatment, withholding rules, and compliance frameworks they had no basis for. One model generated a fictional "LHDN compliance schedule" with specific percentage breakdowns that sounded official but were fabricated. Jurisdiction and governing law: 82% hallucination rate. Models often defaulted to US law or Singapore law without asking. For Malaysia-based contracts, we saw several default to Malaysian law but then cite US case precedent. The consistent pattern: any field where the model predicted high statistical likelihood of a specific value, it filled in a value —regardless of whether that value appeared in your input. Why reviewers miss these hallucinations The contracts look complete. They're formatted correctly, they flow logically, the language is professional. If you're scanning for obvious errors—typos, syntax, missing sections—you'll miss the invented terms because they read like intentional drafting choices. The problem compounds when teams use these drafts as templates. One person generates a contract with hallucinated terms, another person reviews it and thinks "this looks like our standard structure," and suddenly that fabrication becomes organizational practice. We also tested what happens when you ask the same model to generate a contract twice. Hallucinations are not consistent . The first draft might invent a 24-month auto-renewal; the second might invent a 12-month auto-renewal with a 45