AI contract drafting tools have real value: they compress a blank page into a working first draft in minutes, not hours. But I've watched them fail in the same seven ways consistently enough that I treat every AI contract as a liability audit, not a finished agreement. The problem isn't that AI hallucinates legal precedent—it's that it doesn't hallucinate selectively. It invents case citations, overreaches on liability caps, and leaves gaps wide enough for a lawsuit. Here are the seven clauses where AI contracts consistently miss the mark, and a checklist to catch them before your lawyer—or worse, a judge—does. 1. Indemnification scope: who actually covers what This is the most common failure. AI contracts write indemnification clauses that are either so broad they're uninsurable or so narrow they don't protect anyone. The error usually looks like: One party indemnifies the other for "all losses arising from breaches of this agreement"—which sounds good until you realize it includes your own negligence. Indemnification only covers direct damages, missing the "arising out of" language that courts use to expand liability. No carve-out for the indemnifying party's gross negligence or willful misconduct. The audit: Cross-check the indemnification clause against your insurance policy. If your insurer won't cover the scope you've agreed to indemnify, you've either overpromised or underfunded the risk. Most AI drafts create this gap because the model has no context for what your company's insurance actually covers. 2. Liability caps that don't survive a real claim AI loves a round number. "Liability shall not exceed the fees paid in the preceding 12 months" is elegant and symmetrical. It's also often wrong, for one simple reason: AI doesn't model the conversation between the cap and the carve-outs. Typical failure patterns: A liability cap that applies to all losses, including consequential and indirect damages—which means neither party has any protection when a real incident happens. No carve-out for data breaches, IP infringement, or breach of confidentiality—the exact scenarios where damages exceed the cap by orders of magnitude. Caps that conflict. You'll see "liability capped at annual fees" in one clause and "unlimited liability for IP infringement" in another, with no clear hierarchy. The audit: List your material risk buckets: data loss, third-party IP claims, service downtime, confidentiality breach. For each one, check whether the cap applies and whether carve-outs exist. If your liability cap is $50K but a data breach could cost your customer $500K, you've just written an unenforceable clause. 3. IP ownership that inverts who controls the work This is where AI systematically overreaches. It tends to award too much IP to the vendor because the model was trained on vendor-friendly boilerplate, then suddenly swings to award everything to the client because it saw that pattern too. The result is contradictory language that a court will interpret against you. Typical failures: "All work product shall be owned by Client" (standard for agencies) combined with "Vendor retains all rights to underlying tools and methodologies" (standard for SaaS)—leaving it unclear who can reuse templates, code, or frameworks. Pre-existing IP is "owned by the party that created it before the engagement," but the clause doesn't define what counts as pre-existing. If you use your own software to build something custom, is the custom output pre-existing or new work product? No distinction between copyrightable work (code, documents, design) and patentable inventions, leaving AI to default to "all IP." The audit: Separately define: (a) pre-existing vendor IP and tools, (b) custom work product created during the engagement, (c) who can reuse each category, and (d) what happens if the engagement ends. AI contracts often skip (c) and (d) entirely. 4. Governing law and jurisdiction that don't match your actual exposure AI picks a governing law almost randomly—often defaulting to California or New York because that's what's overrepresented in training data. If you're signing with a customer in Malaysia, an AI contract might specify California law, which means any dispute goes to California court at California legal cost, even if the contract was signed by parties with zero presence there. Specific failure modes: Governing law specified, but no dispute resolution mechanism. Do you litigate? Arbitrate? The contract says nothing. Arbitration clause that names a venue or arbitrator selection rule that doesn't exist or isn't enforced in your jurisdiction. Forum selection that contradicts governing law. "This shall be governed by Singapore law and disputes resolved in California courts." Singapore courts will reject this. California courts will too, depending on where the defendant lives. The audit: Match the governing law to where both parties have material assets or operations. If you're a services firm in the US contracting with a client in