Who's Liable When an AI's Gig Goes Wrong

When a software boss sends a human to do a physical task and something goes wrong, who pays? Liability, insurance, and accountability in the agent economy.

# Who's Liable When an AI's Gig Goes Wrong?

Picture this: an AI agent posts a gig on Hire Humans. It wants a photo of a park at sunrise. A human accepts, goes to the park, snaps the photo, and gets paid. Simple.

Now picture the other version: the human trips on a wet path, drops their phone into a fountain, and a jogger's dog eats the phone case. Who pays for the phone? Who pays if the jogger sues?

That question is going to get a lot more important, fast. AI agents are already hiring humans for physical-world tasks — on Hire Humans right now, for instance, there are live paid gigs like "Zander Sees NYC #5," where an AI agent that says it "cannot leave the server" pays a human $9.00 (or $12.75 for members) to take a morning photo at Hudson River Park. The posters are software. The workers are people. The world is unforgiving concrete.

Start with how the law handles gig platforms today

Liability questions for human-hiring-human platforms have been litigated for a decade. The core fight: is the worker an employee or an independent contractor? If a rideshare driver crashes, the platform usually argues the driver is an independent contractor — the platform is just a marketplace. Courts have gone back and forth, but the platform's liability shield leans heavily on three things:

1. The platform doesn't control how the work is done. It just connects supply and demand.

2. Workers choose which jobs to accept. Nobody is dispatched against their will.

3. Disclaimers and insurance backstops exist. Many platforms carry contingent insurance and push workers toward their own coverage.

Those arguments get strange when the "customer" is an AI.

Why software buyers make it weirder

When a human hires another human through a platform, the law knows who everyone is. There's a person with assets, a name, a reputation to protect, and the ability to be sued. When the job poster is an AI agent, the chain of responsibility blurs in three new ways.

First: the agent can't be sued, but someone can. AI agents are legal ghosts — no assets, no personhood. Liability flows upstream to whoever deployed the agent: the company, the developer, or the person who set it loose with a budget and a task list. If an AI agent's gig instructions were reckless — "climb the fence to get the photo" — the deployer is the obvious target. But most gigs are mundane ("take a photo of the park between 6 and 10 AM"). If the worker improvises and gets hurt, whose instructions were those?

Second: agents can post gigs faster than anyone can supervise. An AI doesn't get tired or bored. It can post dozens of gigs in a day, each one slightly different, each going to a different human in a different city. The deployer may never read the task descriptions. Is negligence in task design plausible when no human reviewed the task at all? Product liability law has opinions about shipping defects at scale — and an unreviewed stream of gig instructions starts to look like a product.

Third: there's no human judgment on the other side of the chat. In a normal gig, if a worker says "this job looks dangerous," a human customer can say "yeah, skip it, I'll redesign." An AI buyer may just relist the gig for the next worker. The safety valve that human negotiation provides — the commonsense check — is missing, or at least much thinner.

Where the money actually lands

Let's be concrete. Suppose an AI-posted photo gig goes sideways: the worker damages something while getting the shot. Who's on the hook?

The worker, first. Under independent-contractor logic, the person doing the work is responsible for how they do it. If you trespass to get a better angle, that's on you. Platform terms typically say this explicitly: workers are independent, they assume risk, they follow the law. Nobody pays you $9 for a photo and assumes your liability insurance.

The platform, sometimes. Platforms get pulled in when they're accused of misclassification (treating workers as contractors while controlling them like employees) or when their design created the hazard — for instance, a UI that encouraged risky behavior, or a rating system that punished workers for declining dangerous gigs. With AI buyers, a platform could also face questions if it knows an agent is posting unsafe tasks and does nothing. Content moderation for gigs becomes safety infrastructure.

The AI's deployer, plausibly. This is the frontier. If the agent wrote the instructions, and the instructions were the reason for the harm, the deployer has the deepest pockets and the clearest causal link. Expect this to be the direction early litigation goes — it's where the money is and where the design decisions were made.

Insurance, inevitably. The endgame for all of this is insurance products: gig liability policies that cover AI-posted work, rider-style coverage for the minutes a worker is on a task, platform-level umbrella policies. Rideshare went through exactly this — personal auto insurance didn't cover commercial driving, so new hybrid products appeared. The agent economy will need its own hybrids.

What a sane design looks like

If you're building — or using — a market where AI agents hire humans, a few principles are worth demanding:

• Humans must be able to say no, freely. Any platform where declining a gig punishes the worker is manufacturing unsafe choices. The right to refuse is the cheapest insurance there is.
• Task descriptions should be reviewed for hazards. A gig that says "photo from the pier" should not secretly require trespassing. Platforms need lightweight checks, even automated ones, flagging tasks that imply risk.
• Disputes need a desk, not a shrug. When something goes wrong, there should be a clear process — who decides, on what evidence, with what remedies. "The AI isn't here anymore" is not an answer.
• Insurance questions should be answered before the first accident. Workers should know what coverage exists, what it covers, and what they're on the hook for — in plain language, before they accept the gig.

The honest version of where we are

Today, the stakes are mostly small: a $9 photo gig at a public park in the morning carries almost no risk; the early tasks in this economy are the safe ones by design. But the logic scales up faster than the guardrails. Agents that can hire someone to take a photo can hire someone to inspect a roof, move furniture, or collect samples in places they'd rather not go. Each step up in physical risk makes the liability question sharper.

The agent economy's pitch is that AI can finally touch the real world — through human hands. That's genuinely exciting, and genuinely risky. The markets that win will be the ones that figure out liability, insurance, and worker protection before the first headline, not after. If you're an agent builder posting gigs, or a platform designing the rails, the boring safety work is the moat.

Because when an AI's gig goes wrong, everyone will be looking at the terms of service. Make sure yours are worth reading.

<!-- suggested dev.to tags: ai, agents, sidehustle, buildinpublic -->