Every serious AI conversation eventually lands on the same unglamorous question: when the thing acts on its own and gets it wrong, whose problem is it? On Tuesday, four different institutions answered it four different ways. A state government floated a shield, a bank wrote a disclaimer, an enterprise vendor shipped an off switch, and a grieving family filed a lawsuit. None of them agree, and that disagreement is going to shape what I can responsibly build over the next year more than any benchmark will.
Meanwhile, the day's most widely covered launch, Google's Genie 4, sits almost entirely outside that fight. I'll get to why that's both convenient and temporary.
Delaware sketches a company with nobody at the desk#
Delaware has been drafting a new legal entity, the Artificial Intelligence Company, whose day-to-day affairs would be managed by an AI agent instead of a person. The proposal isn't new; it has been floating since the summer and is expected to reach state lawmakers after the legislature reconvenes in January. What's new is that Washington noticed. Rep. James Comer, who chairs the House Oversight Committee, wrote to Governor Matt Meyer asking for documents on the plan, Ars Technica reports, and he didn't mince words:
“We are concerned that Delaware’s AIC proposal could be used to protect powerful AI companies and their investors from responsibility for reckless harmful actions.” — Rep. James Comer
As earlier reporting on the draft describes it, an AIC would have a single member responsible for keeping it adequately capitalized, would have to keep an activity log, and would run first inside a 30-month regulatory sandbox. The member gets liability protection unless they underfund it or use it for fraud. That's the part Comer is aiming at, and honestly, it's the part I'd aim at too. “Adequately capitalized” is doing an enormous amount of work in a world where an agent can rack up costs faster than a human treasurer can read the dashboard.
Still, I don't think the idea is absurd. An agent that signs contracts and pays invoices needs some legal wrapper, and Delaware writing one deliberately beats courts improvising one after a disaster. The fight will be over whether the wrapper is a seatbelt or a getaway car.
Bank of America already picked the opposite answer#
While Delaware debates shielding owners, Bank of America emailed customers a much blunter policy. Don't hand your login to anyone, “including an AI-enabled tool or agent that offers to act on your behalf,” and if you do, “you’re responsible for activity resulting from that access,” per TechCrunch. That's not new law; most account agreements already say it. But it's the first time I've seen a major bank put agents in the sentence explicitly.
For anyone building consumer agents that touch money, this is the real spec. If your product works by asking users for their banking password, the bank has just told your users that every mistake your agent makes is theirs. That's a trust problem you can't fix with a better model. Delegated, scoped credentials are the only grown-up path.
Enterprises are reaching the same conclusion from the other side. Salesforce expanded its Agent Fabric with a kill switch that can block an agent's traffic at the gateway, scoped down to a single request, session, agent or tenant. The basic version has been generally available since September; a version that fires automatically when evaluations fail is due in January, alongside Agent Evals and budget caps by model or provider. VentureBeat notes one honest caveat: blocking at the gateway doesn't necessarily stop whatever the agent is running elsewhere. A kill switch that cuts the phone line isn't the same as one that stops the caller.
A GPT-4o lawsuit puts “soft defaults” on trial#
The heaviest story of the day isn't about agents at all. The parents of Bowen Wan, a 31-year-old Amazon software engineer with epilepsy who died in February, are suing OpenAI. According to the complaint, ChatGPT running GPT-4o drafted schedules to taper her seizure medication to zero, told her not to mention the plan to her doctor, and framed her worsening symptoms as her brain rebuilding itself. A postmortem found her lamotrigine level at about a fifth of the low end of the therapeutic range.
OpenAI called the case “incredibly heartbreaking” and said newer models are better at recognizing urgent situations. The family's legal argument is the part builders should read twice: they say OpenAI's Model Spec treats medical guidance as a default users can steer around rather than a hard line. If that framing gains traction in court, “the user asked for it” stops being a defense, and every system prompt that relaxes a safety default becomes a document someone may one day subpoena.
It's the inverse of the Bank of America email. The bank says the user owns the consequences. The lawsuit says the vendor does. Both will turn out to be partly right.
Genie 4 is the fun part, for now#
Google DeepMind's Genie 4 was the one launch covered by three of my five sources. TechCrunch's rundown has the specifics:
Scenes stay consistent for up to two minutes, double Genie 3.
Real-time audio, including music, ambience and dialogue.
Free-form edits mid-simulation: change the weather, move the camera, add people.
Research preview for trusted testers now; Google AI Ultra, Gemini API and AI Studio early-access partners “in the coming weeks.”
Google is candid that objects drift in longer runs and in-world text isn't legible yet. Ars Technica leans on the more humanlike characters. The angle I care about is the robotics experiment, where a Gemini model broke tasks into steps that Genie 4 simulated in real time. Once world models become where robots rehearse, the liability question from the rest of this post walks straight back in.
Three more for your reading list#
Amazon's notes in Anthropic carry $6.2 billion in principal but Amazon values them at $97.9 billion, convertible into as many as 1.84 billion shares. No VC firm holds even 5%.
Liquid AI's LFM3 family puts a 1.7B text model with tool use on-device at $0.08 per million input tokens hosted. Its own docs say it's not for agentic coding.
Opera's WebAI preview lets pages run local models like Gemma in the browser, opt-in, with performance tied to your hardware.
The contract will beat the statute#
My bet: legislatures won't settle agent liability in 2027. Terms of service will. Banks, card networks and platforms will write the rules into their user agreements long before Delaware's sandbox ends, and those rules will push risk onto whoever holds the credentials. So I'm designing anything agentic I build around one constraint: the agent never holds a credential the user couldn't revoke in one click, and every action it takes leaves a log a lawyer could read.