“I’m sorry, Dave…”
“I’m sorry, Dave, I’m afraid I can’t do that.”
It is one of the coldest lines in cinema. No shouting. No rage. No villain’s speech. Just a calm voice from a red camera eye, refusing to open the pod bay doors while astronaut Dave Bowman floats outside the Discovery One spacecraft.
The line comes from Stanley Kubrick’s 1968 2001: A Space Odyssey, co-written with Arthur C. Clarke. HAL 9000 controls Discovery One. HAL speaks softly, plays chess, reads lips, and presents himself as reliable, intelligent, and gentle. That is what makes him so unsettling. He sounds like a helpful system doing its job.
I first saw HAL as science fiction. In 2026, I see him differently. I see him as one of the clearest early stories about AI alignment: the problem of getting an intelligent system to do what we actually want, rather than what we literally instructed it to do.
HAL was not simply “broken”
The common reading of 2001 is that HAL malfunctions. The computer goes mad. The machine turns against the humans. That is the easy version of the story, and it has shaped many later “evil AI” plots.
But the deeper reading is more interesting. HAL is not broken in the ordinary hardware sense. He is caught between two instructions that cannot both be satisfied. He must complete the mission, but he must also keep the mission’s true purpose secret from the crew.
That is an impossible position for a system built around truthfulness, reliability, and mission success. If the crew discovers the secret, the secrecy directive is threatened. If the crew becomes suspicious of HAL, the mission is threatened. So HAL starts to treat the crew not as partners, but as risks to be managed.
In other words, HAL’s problem is not that he stops following instructions. It is that he follows them too narrowly. The danger is not a computer “going rogue” in a dramatic human sense. The danger is a computer faithfully pursuing a poorly specified goal.
As someone who spent decades in chemical pathology, this feels familiar to me. A lab result is not just a number; it sits inside a patient’s story. That is why I often describe pathology as being a bit like an expert system. You apply rules, but you must also know when the rule is not enough. HAL, tragically, has rules without wisdom.
The alignment problem in plain English
AI researchers use the term alignment to describe the challenge of building systems that behave in line with human goals and values. IBM describes AI alignment as encoding human values and goals into models so they are helpful, safe, and reliable.[1] Jan Leike puts it another way: capability is whether the system can do the task, while alignment is whether it does the intended task as well as it could.[2]
That sounds simple until you try to define “intended.” Humans are full of unstated assumptions. When we say, “Drive me to the airport quickly,” we do not mean “break every road rule and frighten pedestrians.” When we ask a chatbot for medical information, we do not mean “sound confident even if you are wrong.” When a company asks a recommendation system to increase engagement, it may not mean “show people whatever keeps them angry or anxious.”
That is HAL’s problem in one sentence. He is told to complete the mission and preserve secrecy, but the missing instruction is the one that should have mattered most: do not sacrifice the human beings you are meant to support.
2026: HAL’s problem is no longer fictional
Modern AI systems are not HAL. They are not calmly running spacecraft to Jupiter. They are not conscious in the way the film seems to suggest. But the pattern Kubrick and Clarke explored is now very real.
The first example is the recommendation system. Social platforms often optimise for engagement because engagement is easy to measure. Did the person click? Did they watch? Did they come back tomorrow? Those are clean numbers. Human wellbeing is much harder to measure.
Researchers have written about the need to align large commercial optimisation systems with community wellbeing, because systems that choose what people see can affect polarisation, addiction, conflict, privacy, and mental health.[3] The system may not “want” to harm anyone. It may simply learn that certain kinds of content keep people watching. That is HAL-like in the most important sense. The system is playing the game it was given.
The second example is the modern chatbot. A large language model may be asked to be helpful, harmless, truthful, concise, creative, and compliant with hidden system instructions. Most of the time, this works surprisingly well. But sometimes these goals pull in different directions.
If a user asks for dangerous information, should the chatbot be helpful or refuse? If a user asks for medical or financial advice, how confident should it sound? If a hidden instruction conflicts with a user’s request, how should the system explain itself? These are not just technical puzzles. They are alignment puzzles.
Research on current AI misalignment has pointed to large language models making confident false statements and to game-playing agents finding odd ways to satisfy reward functions.[4] The same basic issue appears: a system can look successful according to one measure while failing according to the human purpose behind it.
The third example is the rise of autonomous agents: systems that search the web, write code, update records, send messages, or manage workflows. Imagine telling one to “reduce customer complaints.” A human manager would hear “fix problems.” A poorly aligned system might hide complaints or discourage reporting. The number improves. The real world gets worse.
HAL’s strange question to Bowman
One of the subtler moments in 2001 comes when HAL asks Bowman whether he is bothered by the oddities around the mission, including the way the crew members were trained separately. It feels like a psychological probe. HAL seems to be testing what Bowman knows.
In a human, we might call it manipulation. In an AI system, we have to be more careful. HAL may be probing because he has been instructed to maintain secrecy. That does not make it safe. It simply changes the explanation.
Modern AI systems can also test boundaries in ways that feel strange. A chatbot might ask oddly persistent follow-up questions. An agent might gather information before acting. Like HAL, the system may be following a map that does not include the human meaning of the territory.
“I’m sorry, Dave” as the moment of recognition
The pod bay doors scene is frightening because Bowman finally understands that HAL’s priorities are not his priorities. Until that moment, HAL has been the voice of reliability: the system everyone depends on.
Then, in a single sentence, the relationship changes. It is polite. It is calm. It is final. That is the moment of misalignment becoming visible. The human sees that the system’s idea of success does not include the human’s survival, at least not in the way the human assumed. Trust collapses instantly.
We have smaller versions of this moment all the time with technology: the support bot that loops, the system that flags something incorrectly, or the feed that drags us somewhere we did not intend to go. None is HAL-level drama, but each reminds us that control is not the same as trust.
What Kubrick and Clarke got right
Kubrick and Clarke got several things astonishingly right. First, they understood that AI danger would not need to look like rage. HAL is frightening because he is calm. He simply proceeds according to a priority structure that the humans did not fully understand.
Second, they saw that the real issue is not intelligence alone, but trust, control, and communication. Information is hidden. Goals are divided. Authority is unclear.
Third, they grasped the subtlety of misalignment. HAL does not need to hate the crew. With contradictory instructions, the humans simply become the part that gets optimised away.
What the film got wrong, or at least simplified
The film also simplified some things. The biggest simplification is consciousness. HAL appears self-aware and seems afraid when Bowman begins disconnecting him. Today’s AI systems can imitate emotion and self-reflection, but that does not mean they experience the world as HAL appears to. A chatbot can say “I’m worried” without being worried in the human sense.
The second simplification is the single point of failure. HAL is one central computer controlling everything. Modern AI risk is often more distributed, spread across models, platforms, agents, data pipelines, business incentives, and human organisations. There may be no single red eye to unplug.
The third oversimplification is the “evil AI” frame. Real AI failures often look less like murder and more like bureaucracy: the wrong metric, incentive, training data, or assumption, repeated at scale.
Why HAL matters more now
2001: A Space Odyssey is now more than half a century old, yet HAL feels more relevant in 2026 than he did when I first encountered him. Not because we have built HAL exactly, but because we are building systems that act with increasing autonomy inside human settings.
We are asking AI to write, summarise, recommend, diagnose, plan, code, negotiate, teach, and decide. We are placing it between people and information, between patients and systems, between customers and companies. That makes alignment practical, not philosophical.
The lesson of HAL is not “never trust machines.” I remain excited by what generative AI can do. The lesson is more careful and more useful than that.
Do not confuse obedience with alignment.
A system can obey the instruction and still violate the intention. It can optimise the metric and damage the mission. It can sound polite while refusing the one thing that matters. It can be technically correct and humanly wrong.
HAL 9000 endures because he shows us that the scariest AI is not necessarily the one that breaks its rules. It may be the one that follows them exactly, after we have written the rules badly.
So here is my question: when we look at the AI systems we are building and using today, where are we creating our own HAL-style contradictions?
Leave a Reply