AI Liability Stopped Being Hypothetical. It’s Looking for Whoever Deployed It.
Robert Grand · Battalion Chief who still runs calls
This year the AI harm cases stopped being a thought experiment. In February the California Superior Court in San Francisco consolidated twelve separate cases pending against OpenAI into one coordinated proceeding (https://natlawreview.com/article/ai-reckoning-has-arrived-case-will-rewrite-ai-laws-products-liability). Before that, Character.AI settled a cluster of suits, including the Garcia case, where the court was willing to treat the chatbot as a ‘product’ for liability purposes because the claims went at specific design choices. The tech press read it the way you’d expect: model makers, plaintiffs’ firms, nine-figure defendants, a fight happening a long way from any fire station.
Read the doctrine underneath the headlines and it isn’t a tech-industry story at all.
The name on the purchase order is the one the lawsuit goes looking for.
What the Courts Just Started Sorting Out
One liability summary put it about as plainly as it can be put: when AI causes harm, the law “usually does not shrug and blame the algorithm” (https://blueheadline.com/need-to-know/ai-liability-2026/). It looks for the people and the organizations who built it, sold it, deployed it, supervised it, or ignored an obvious warning. That’s a list of human roles, and every one of them is a seat somebody actually sits in.
OpenAI is fighting hard to be called a ‘software based service’ instead of a ‘product,’ because if a court buys that framing, a large chunk of product liability theory falls away. That fight is not resolved, and the firms tracking it are clear that the next wave of litigation is aimed not just at who wrote the model but at who put it into a high-stakes workflow and what oversight they can show for it (https://www.klgates.com/AI-Product-Liability-The-Next-Wave-of-Litigation-3-27-2026). The legal system is doing what it always does with a new technology. It’s deciding who is responsible, and it is deciding in public, one case at a time.
None of those cases involve a fire department yet. That’s the only part of this that’s comfortable.
The Liability Is Moving Toward Whoever Deployed It
Here’s the shift that matters for us. The exposure is migrating from the vendor who built the tool toward the agency that deployed it. The interesting legal question is no longer only “what did the algorithm do.” It’s “who put this in front of a decision that mattered, and what did they do to make sure it was safe.”
“We bought it off the shelf” used to sound like a defense. It’s becoming an admission.
That single move changes what a software purchase is. When an EMS agency or a fire department buys an AI tool that touches a clinical or operational decision, it isn’t just acquiring a capability. It’s accepting a duty. And in most departments right now, nobody signed up for that duty on purpose, which means it landed on whoever signed the contract, usually a person who was thinking about workflow and budget, not about a deposition three years out.
Two Rooms That Never Talk to Each Other
The deeper problem is structural, and it’s the same one every high-stakes field is running into. The tool gets adopted in one room. The liability arrives in a completely different one.
The procurement meeting is calm. There’s a demo, the tool solves a real pain, the numbers work, somebody approves it. Nobody in that room is picturing the 0300 call. The call is the other room. It’s loud, it’s fast, and nobody on it remembers the procurement meeting or has any idea what the vendor contract said about who’s accountable when the recommendation is wrong.
Picture an AI triage assist on the EMS side, or a CAD recommendation engine that prioritizes calls and suggests unit assignments. Most days it’s fine. Most days it’s better than fine. Then one day it down-prioritizes a call that turns bad, and the question in the room stops being technical. It becomes: who approved this tool, what did it recommend, did the medic follow it or override it, and where is the record of any of that. If the honest answer to most of those is a shrug, the agency is defending negligence, not judgment.
I’ll be straight about where I am on this. I don’t have our AI procurement policy written, and I’d be suspicious of anyone in Fire & EMS who says theirs is airtight. I know what the first question is, though. Before we deploy anything that shapes a dispatch or a patient-care decision, who in this organization owns that decision when it goes sideways. If we can’t name that person, we haven’t bought a tool. We’ve bought a liability with a login.
What “We Just Bought It” Actually Buys You
This is where the indemnification clause stops being fine print. A vendor contract that quietly shifts liability onto your department isn’t a paragraph to skim on the way to signing. It’s the whole ballgame. The slick black box that can’t explain how it reached its recommendation isn’t just a clinical risk. It’s a tool you can’t defend, which means it’s a tool you can’t really use.
The regulatory side is hardening in the same direction. Under the EU AI Act, most of the high-risk obligations, including qualified human oversight, kick in starting August 2, 2026, and the US picture is a patchwork moving the same way. The throughline is that a human being is supposed to hold real authority over the AI in high-stakes calls, and the organization is supposed to be able to show it. Not a name on an org chart. A trained person with the standing and the documentation to prove the oversight was real.
So what “we just bought it” actually buys you, in 2026, is a duty you now have to be able to account for. That reframes the whole evaluation. The first question on the rubric isn’t “does this tool work.” It’s “when this tool is wrong, can we explain what it did, show who was responsible, and prove we trained our people on its limits.” Most departments have never had to ask a vendor that. They’re about to.
The Call That Forces the Question
Law enforcement, healthcare, finance, and legal all spent this year answering the same question out loud: when the AI is in the decision, who owns the outcome. Fire & EMS hasn’t asked it yet, and the reason isn’t carelessness. It’s that no existing seat in the org owns “AI risk,” so it falls through the cracks until something forces it up. The forcing function is always a call. It’s never a memo.
The cost of building the seat now is meeting time. The cost of waiting is a department leader discovering in a deposition that signing the purchase order made them the owner of every recommendation the tool ever made. Those are not close. One is cheap and boring. The other is expensive and public.
The liability question used to belong to somebody else. It belongs to whoever deploys the tool now, and in our service, that’s us. Build the seat before the call builds it for you.
Robert Grand is a Battalion Chief at Eugene Springfield Fire with 24 years of service. He writes Frontline Intelligence, a newsletter on operational doctrine, technology, and leadership in Fire & EMS.
If this landed, share it with someone in Fire & EMS who needs to hear it.
Subscribe
From the floor, not the vendor booth. Two times a week.