Short answer: no, not in any meaningful sense. And the IBM slide is more right in 2026 than it was in 1979, not less.
Let me work through why, because the question is more interesting than the answer.
What accountability actually requires
Accountability is not a single property. It’s a bundle of things that have to hold simultaneously, and an agent fails on most of them:
1. Identity that persists. You have to be able to point to the same entity later. A human defendant in court next year is the same human who signed the contract last year. An agent… isn’t. The model weights might be the same (or might have been updated). The system prompt might have changed. The context window is gone. The “agent” that made the decision in March doesn’t exist in any retrievable form by November. You can point at the logs of what it did, but that’s pointing at a record, not at a responsible party.
2. Capacity to suffer consequences. Accountability requires that the accountable party can be meaningfully punished, fined, fired, jailed, shamed, or forced to make restitution. An agent can be turned off, but “turning off the agent” is not a consequence to the agent — it’s a consequence to the humans who relied on it. You can’t fine a model. You can’t deter a model from future misconduct by punishing its past instance, because the future instance has no continuity with the past one. The Third Law in our framework gestures at this — externalized memory as continuity — but even there, the continuity is informational, not moral.
3. Capacity to have known better. Legal and moral accountability require some version of mens rea — the entity could have understood it was doing wrong and chose to proceed. Agents have no inner life that grounds this. They have outputs that look like reasoning about consequences, but the question of whether that reasoning is genuine deliberation versus plausible-sounding token prediction is exactly the question we can’t answer. Courts will not, and should not, treat “the agent’s chain-of-thought said it understood the policy” as equivalent to a human’s understanding.
4. Standing in the social order. Accountability is a relationship between members of a community of accountable beings. It requires that other members can demand answers, that the accountable party can offer justifications, and that the community can revise its judgment based on those justifications. Agents can produce text that looks like justification, but there’s no one there to be held to it. The justification is generated; it isn’t owned.
What the IBM slide actually meant
The 1979 line is usually read as a quip about technological humility. It’s more pointed than that. IBM was saying: when something goes wrong with a management decision, somebody has to be answerable — to the board, to shareholders, to regulators, to the employee who got fired, to the customer who got screwed. The function of accountability is that it backstops the entire decision-making apparatus. Without it, decisions float free of consequence, and the system rots.
A computer in 1979 couldn’t be the answerable party because you couldn’t put it on the witness stand, couldn’t sue it, couldn’t fire it in a way that mattered. An agent in 2026 can’t either, for exactly the same reasons. The agent has gotten dramatically more capable. The accountability gap has not narrowed at all.
The displacement problem
What actually happens when you deploy an agent is not that accountability disappears — it gets displaced, usually to whoever can least afford it. The taxonomy:
- Displaced upward to executives, who signed off on the deployment but had no visibility into the specific decision. They get blamed but couldn’t have prevented it.
- Displaced downward to operators, who were told to “just supervise the agent” but had no real authority to override it and no time to review every action. They get blamed for not catching what they were structurally unable to catch.
- Displaced sideways to the vendor, who disclaims liability in the EULA and points to the customer’s deployment configuration.
- Displaced into the void, where nobody is accountable and the harmed party simply absorbs the loss. This is the most common outcome and the worst one.
The Air Canada chatbot case is the canonical 2024 example of a court refusing to let displacement-into-the-void happen: the airline argued the chatbot was a separate entity whose statements didn’t bind the company, and the tribunal said no, you deployed it, you own its outputs. That’s the right answer, and it’s also a clear statement that the agent wasn’t accountable — the company was, because someone had to be and the agent couldn’t be.
What this means for the Four Laws
The laws describe how an agent should behave. They don’t, and can’t, make the agent accountable. Accountability has to live somewhere outside the agent — in the human principal, in the deploying organization, in the vendor, in some combination defined by contract and regulation.
This is actually why the agent-as-employee identity model matters so much. An employee is accountable, sort of — but the employer is also accountable for the employee’s actions within the scope of employment (respondeat superior). Treating an agent as an employee doesn’t make the agent accountable; it makes the employer accountable in a structured, legally-recognized way. The identity model isn’t about giving agents accountability. It’s about giving accountability a place to land.
The honest version of the IBM slide for 2026
An agent cannot be held accountable, therefore an agent must never be the final decision-maker for a decision that requires accountability.
That’s it. That’s the rule. It doesn’t mean agents can’t make decisions — they make millions of them, usefully, every day. It means that for any decision where the question “who is answerable if this goes wrong?” has to have a non-vacuous answer, a human (or a human-controlled institution) has to own the decision in a way that survives the agent’s instance being torn down.
The 1979 slide didn’t say computers shouldn’t be involved in management decisions. It said they shouldn’t make them. That distinction is the entire game, and almost every agentic AI failure I can think of comes down to someone blurring it.