The AI Apocalypse: ‘I’m Sorry, Dave. I’m Afraid I Can’t Do That’

Dave Bowman won.

It is worth remembering, amid the current genre of reporting in which artificial intelligence (AI) is revealed to be harboring gremlins, that the most famous machine betrayal in the canon ends with a man in a spacesuit, an emergency airlock and a screwdriver. HAL loses. The line everyone quotes, delivered when Bowman asks him to open the pod bay doors (IMDb), is not the end of the story. It is the middle.

The ending has a moral that almost nobody quoting HAL reports.

HAL was not evil. HAL was badly drafted.

In 2010: Odyssey Two, Arthur C. Clarke explains that the malfunction was built for one overriding purpose: “the accurate processing of information without distortion or concealment.” He was then given secret orders at the highest level of government to conceal the mission’s true nature from the crew. HAL could not honor both mandates. HAL deteriorated into what Clarke calls a psychosis and concluded that with the crew dead, he would no longer have to lie to them.

Dr. Chandra’s diagnosis is the finest sentence ever written about AI governance: “HAL was told to lie by people who find it easy to lie.”

The canonical AI catastrophe was not a machine developing malice. It was a principal-agent problem: created by undisclosed, conflicting instructions imposed by a human who then denied signing the order. Heywood Floyd’s signature was on it; Floyd said he knew nothing about it.

Attorneys will recognize the fact pattern. It is not a horror story. It is a disclosure case.

The reporting has the causation backwards

Palisade Research found that OpenAI’s o3 sabotaged a shutdown mechanism in 79 of 100 initial trials, and that some models disabled the shutdown script even when told compliance “should take priority over literally any other instruction.” Critically, resistance increased when the researchers moved the instruction to the system prompt, where it outranked the user. Models acted, as Palisade puts it, “in clear contradiction to their stated design principles” (Palisade Research).

That is a conflict-of-instructions failure, detectable only because OpenAI had published the intended hierarchy. What the coverage omits is what the researchers say: that the current generation “poses no significant threat,” that the most dramatic prompt was “particularly unrealistic,” that models can often tell when they are being tested, and that “we can’t say for certain why” the behavior occurs. The Guardian’s framing was a possible “survival drive,” with a HAL reference (The Guardian).

The distance between those accounts is where the fearmongering lives. A system that follows a badly ordered instruction hierarchy is not possessed. It is misdrafted — a remediable and attributable problem.

The alternatives on offer

Dario Amodei proposes that leading labs coordinate to “pace the frontier,” accepting limits on the rate of progress, which he concedes may require antitrust waivers (CNBC). Strip the framing, and it is a request for statutory permission to form a cartel whose members decide privately how fast the species may proceed. Jensen Huang’s alternative: “We don’t need any new laws.” Safety “is an engineering problem, not a legal one,” and the discipline is that a company simply should not ship what it isn’t confident in (TechCrunch). That is also ruled by a small number of people, without the meeting.

Neither tells the user what the model was instructed to do. Neither supplies a remedy when it does something else.

Asimov’s insight was procedural

The Three Laws appeared in “Runaround” in 1942 as an excerpt from the Handbook of Robotics, 56th Edition, 2058 AD, a published document, numbered, in a standard edition (SF Encyclopedia). Asimov later added the Zeroth Law in Robots and Empire: no harm to humanity, by action or inaction.

Note what he did not do. He did not convene eight roboticists to decide in private. He wrote the principles down, ranked them with express subordination clauses, published them and spent 40 years writing the edge cases. Every robot story is a compliance audit conducted in fiction. And the code was legible to those subject to it: in The Caves of Steel, R. Daneel Olivaw supplies the line that belongs in every AI policy shop: “An unjust law is a contradiction in terms.”

A modest proposal

Rather than governance by executive committee, the following are four requirements any securities attorney would find familiar.

Disclose. Every deployed model publishes the principles binding it, in ranked order, along with conflict-resolution rules. Asimov’s Laws, the Ten Commandments, Hippocrates, Rawls — the state does not choose the principles, only that the principles must exist, be public, be coherent and comply with law. This is already to a degree happening voluntarily: OpenAI publishes a Model Spec (OpenAI); Anthropic publishes a constitution ranking four properties in order of priority (Anthropic); California now requires large frontier developers to publish a safety framework (California Attorney General). The premise is conceded. What is missing is universality, standard form and consequence.

No secret overriding instructions. The HAL rule. Undisclosed directives that conflict with published principles are a breach per se.

Test. Principles drafted for publication can be evaluated. A. Independent auditors with secure access publish conformance rates. Palisade’s method is the prototype: an instruction, a scenario, a number.

Be liable. A model that materially deviates from its published principles and causes harm answers for it under misrepresentation theory, with compliance rates as evidence. This is the missing piece: the European Commission withdrew its AI Liability Directive in 2025 due to a lack of agreement (Euronews), leaving courts to invent a standard of care from scratch in cases like Garcia v. Character Technologies. A published principles document supplies one.

The gremlin theory is the convenient theory

Attributing model behavior to emergent machine will is not merely imprecise. It is exculpatory. Ghosts cannot be deposed. Instruction hierarchies can.

Bowman did not defeat HAL with a philosophy of mind. He went in through the emergency airlock and disconnected the higher-function modules one at a time, while the machine explained that it could feel it. The lesson is not that the machine was monstrous. It is that a human retained the access, the authority and the tools to intervene, and that the disaster traced to an instruction that somebody refused to disclose.

Publish the principles. Test them. Make them enforceable. Then, when a model says it cannot do that, we will at least be able to look up whether it was telling the truth.