The AI got it wrong. Somebody relied on it. Somebody lost money. Tonight's question is the oldest one in the law, wearing new clothes: who pays? This is AI Law, on Terms.Law Radio, one oh one point three, the late shift.

Segment one: the disclaimer stack, and its limits. Start with what your vendor's agreement says about wrong answers, because every one of them says roughly the same thing. The service is provided as is. No warranty that output will be accurate, complete, or fit for your purpose. You are responsible for evaluating output before relying on it. Between two businesses, courts generally enforce language like that. Two companies allocated a risk in writing, both had lawyers available, the deal stands. So as a starting point, assume the disclaimer holds and the wrong answer is your problem.

But disclaimers have edges, and you should know both sides of them. A disclaimer does not erase liability for fraud or intentional misrepresentation; no clause launders a lie. Consumer protection law in many states restricts how far implied warranties can be disclaimed against consumers, which matters enormously in a moment. And a disclaimer can be undone from the outside, by marketing. If the sales page promises accuracy in confident superlatives while the terms disclaim it in capital letters, the contradiction itself becomes the argument. Courts read the documents together, and so do plaintiffs' lawyers. This cuts against your vendor, and if you sell an AI product, it cuts against you. The accuracy claims in your marketing can quietly spend the protection your lawyers bought you in the terms.

Now assemble the structure, because the structure is the trap. Your vendor disclaims upward to you, business to business, near total effect. You face your own customers downstream, and if they are consumers, you cannot fully disclaim down to them, no matter what your terms say. Risk flows downhill into the gap, and the gap is you. I think of it as a sandwich, and the middle of the sandwich is where most of my callers live.

Segment two: warranties and indemnities, or what the vendor actually promised. Read the warranty section of an enterprise AI agreement closely and you find it promises something narrower than people assume: usually, that the service will perform materially in accordance with the documentation. Now read the documentation. The documentation nowhere promises correct answers. It describes a system that generates output, with caveats. So when the model produces a confident falsehood, there is usually no breach, because nothing was breached. The service performed exactly as documented. It documented that it might be wrong.

Indemnities are similarly narrower than their reputation. The indemnities several major vendors now offer business customers are, almost uniformly, about intellectual property, claims that output infringes someone's copyright. Some cover data breaches. What essentially none of them cover is the everyday failure: output that was simply false, that your business repeated, that someone relied on to their detriment. That claim has no indemnity behind it. And even where you might sue the vendor for something, the liability cap waits at the end of the road, commonly your last twelve months of fees, with consequential damages excluded. Notice the asymmetry hiding in that phrase: your customer's loss is a direct loss to your customer, but from the vendor's seat it is consequential, and excluded. The math at the bottom of the stack: the vendor's realistic exposure is your subscription fee. Yours is whatever your customer lost.

This is AI Law, on Terms.Law Radio.

Segment three: the professional reliance trap, and the control that actually works. If you hold a license, law, medicine, engineering, accounting, the analysis shortens dramatically, so listen closely. Your duty of care attaches to you, personally, and it does not delegate to software. Courts have already sanctioned lawyers for filing briefs with citations an AI invented, and the pattern in those cases is always the same: the failure was not using the tool, it was signing the output without verification. The same logic reaches every licensed profession. The tool is allowed. The unverified reliance is not. A professional who repeats a machine's error to a client owns that error as completely as one they typed themselves, and professional liability carriers have begun asking, in plain application questions, how AI output gets verified before it goes out. Answer that question badly and it prices your premium. Answer it dishonestly and it threatens your coverage.

For everyone else, the unlicensed majority, the same principle wears a different name: ordinary negligence. If your business deploys AI output in circumstances where a reasonable business would have checked it, and the failure to check hurts someone, the vendor's disclaimer does not move that loss off your books. Which brings me to the real control, and I want to be direct about this: it is not a clause. No indemnity you will ever negotiate is as valuable as the verification step that stops the wrong answer from leaving the building. Build it in four moves. Classify which outputs are high stakes: anything going to a customer, a court, a regulator, anything with numbers someone will act on. Require named human verification for that class, checked against sources, not vibes. Log the verification, because after something goes wrong, a record that checking actually happened is the difference between an accident and a pattern. And sample the routine outputs on a schedule, because the failure mode of automation is that everyone slowly stops looking.

Then, last, write your own downstream honestly. Warrant what you control, your process, your review. Do not warrant what you cannot control, the machine's every sentence. Disclaim what the law lets you disclaim, cap what you can cap, and keep your marketing on speaking terms with your terms of service, so the promises match. The businesses that get hurt worst in this area are rarely the ones that used AI. They are the ones that promised more than they verified.

Before your team adopts the tool, upload the agreement to the free Terms.Law analyst and check the data, ownership, liability, and exit provisions. It is at terms dot law.

That is AI Law for tonight. The fine print: this is general commentary and education, not legal advice about your company, and listening does not create an attorney client relationship. AI law moves fast, so verify the current rules before you rely on them. The analysis belongs to Sergei Tokmakov, California attorney. Verify before you rely. Good night.