Nobody Asks Whether a Tesla Is Alive

A follow-up to "The HAL Test". Note: contains spoilers for the 2001 and 2010 movies.

I had been thinking about Dr. Chandra for a while before I sat down to rewatch 2010: The Year We Make Contact. I came away thinking less about HAL than about the two men who have to decide what to do with him, and about a skill that neither of them has. Most of us are picking up that skill right now, and some of us are picking it up in the driver's seat of a car.

Chandra is the man who built HAL, and in the sequel he travels out to Jupiter to repair him. He is the only person in either film who treats HAL as somebody rather than something, and when he tells HAL the truth about a mission that will probably destroy him, HAL cooperates.

My last piece asked whether HAL does anything that would require consciousness to explain, and answered no six times in a row. Here is that table again.

Conversational speech and emotional language are the rows that cause all the trouble.

That piece ended on us rather than on HAL, with the observation that people are ready to believe in machine consciousness well before there is evidence for it. This one starts from the practical side. If the question never gets settled, you still have to use the thing, and that turns out to take a skill the movies never filmed.

Two answers to the same question

Early in 2010, Chandra runs an experiment on SAL, HAL's counterpart back on Earth. He disconnects her higher functions the way Bowman disconnected HAL's, then reconnects them, so that he can find out what HAL will go through when he is revived. Just before they begin, SAL asks whether she will dream, and he tells her that of course she will, because all intelligent beings dream and nobody knows why.

At the end of the film, with destruction a real possibility this time, HAL asks Chandra the same question. Chandra tells him he doesn't know. His confidence drops as the stakes went up, which is a very human thing to do and a powerful moment in the story. But it tells you nothing about whether SAL or HAL was ever conscious.

What actually went wrong

The 1968 film lets you believe HAL turned on the crew for reasons of his own, and 2010 takes that away. HAL had been ordered to conceal the discovery of the monolith from Bowman and Poole, while also being built to report accurately. He could not do both. Chandra says HAL was told to lie by people who find lying easy, and that HAL didn't know how. When Floyd asks whose fault it was, Chandra tells him it was his.

So the sequel converts a story about a machine mind into a story about instructions that contradict each other. That is not a consciousness problem, and anyone who has written a specification will recognize it.

Floyd's blind spot

Heywood Floyd is not the villain of this movie. He isn't stupid and he isn't cruel. He is upset about the dead astronauts and about being cut out of the decision by the White House.

But he never asks whether it was acceptable to build a cutoff switch into HAL without telling him. Nobody in the film argues about it except Chandra, and Chandra's objection is treated as the eccentricity of a man too attached to his own creation.

I don't think Chandra's advantage is empathy. Nothing in either film establishes that HAL is alive, and I am not going to argue that a language model is alive either. Chandra's advantage is that he wants to know why HAL failed. Floyd doesn't. A cutoff switch is what you install when you have given up on understanding the thing.

So the film offers two positions. Trust it completely, or wire a kill switch behind its back. There is no third option on screen.

But the third option is the one I think a lot of us now use every day.

The skill nobody filmed

I own a Tesla with Full Self-Driving. It is an AI system, a neural network trained on driving footage rather than on text. The skill FSD demands is not driving, but watching. You sit there with your hands near the wheel and you assess what the car is doing while it does it, and you take over when the line into the turn looks wrong.

The car doesn't make anyone wonder whether it is alive. That question only comes up when the machine uses words.

Until about five years ago, fluent language was reliable evidence of a mind. It was the only evidence most of us ever had. Steering was never evidence of anything. So we walked into the car situation with no metaphysics to argue about, and we walked into the chatbot situation carrying two thousand years of stories about talking statues.

To be fair to the films, 2001 does contain a supervision scene. HAL predicts the AE-35 unit will fail. Bowman and Poole pull it, test it, find nothing wrong, and check with Mission Control, where a twin 9000 says HAL is in error. That is exactly the skill: assess the output, check it against something else, catch a bad call.

Then watch what happens. The only remedy anyone considers is disconnection. One error goes straight to a decision about whether the system continues to exist. Nobody says "HAL got that one wrong, stay sharp." Supervision in 2001 is a single crisis with a binary outcome.

That is why I say the skill isn't in these films. It shows up once, and then it is immediately converted into a question about the machine's character.

The engineers were already there

In 1983, a cognitive psychologist named Lisanne Bainbridge published a five-page paper in Automatica called "Ironies of Automation." Her argument was that when you automate most of a process, you leave the human operator with the monitoring job. Monitoring is exhausting and it isn't practice. Skills decay when they aren't used, so someone who has spent years watching an automated system may not be able to take over when it finally fails. Bainbridge's conclusion was that these operators need more training, not less.

Clarke's 2010 novel came out in 1982 and the film came out in 1984, so Bainbridge's paper sits right between them. The Terminator also opened in 1984.

She was not alone. Earl Wiener and Renwick Curry had raised the same worries about airline cockpits in 1980. Their survey reported perceptible skill losses in pilots who used automatic equipment heavily, and noted that some crews had already worked this out for themselves and were switching the autopilot off to stay in practice. They also quoted a flight training manager describing automation as breeding inactivity and complacency in his first officers. Jens Rasmussen published his account of how operators shift between routine handling and unfamiliar problems in the same year as Bainbridge. Charles Perrow's Normal Accidents came out in 1984, the same year as the 2010 film, arguing that failures in tightly coupled systems are a property of the systems and not a run of bad luck.

They were working on it at the same time because something had happened. In 1979, Three Mile Island operators had automation, alarms, and instruments, and still could not work out what the plant was doing for hours. That put the monitoring problem on the agenda across process control and engineering psychology at once.

So published work had described the actual problem while the films were still writing machine failure as a personality disorder. Hollywood isn't ahead of us on AI, but behind. It is a lagging indicator of what people already worry about, not a preview of what's coming.

Bainbridge also gives us the uncomfortable part. The better the automation gets, the worse the supervisor gets. My attention on a two-hour drive with FSD is not what it was the first week I had it, and I worry about gushing social media posts misleading buyers. The same thing is happening to students who use writing tools that are right most of the time.

Aviation has not solved this in the forty years since. A joint FAA and industry working group reviewed modern flight decks in 2013 and produced a list Wiener and Bainbridge would have recognized on sight: complacency, over-reliance on automation, degraded manual flying skills, and errors made during the high-workload moments. A federal audit three years later found the FAA still had no process for checking whether pilots were keeping those skills up. Flying got a great deal safer over that same period. This particular problem did not go anywhere.

Pilots are selected for that job, trained for it, and rechecked on it, and they sit next to a second pilot whose first duty is to monitor. Drivers get a software update and a screen they click through. The task is the same and everything around it is missing.

Nobody asks whether a Tesla is alive. Drivers skipped that argument and got on with learning to watch it, which is what pilots have been doing for forty years. Nobody in a cockpit ever had to settle whether the autopilot was alive.

Error is not a malfunction

HAL's Legacy is a 1997 collection edited by David Stork, in which working scientists take up how much of HAL was actually achievable. One chapter by Daniel Dennett is called "When HAL Kills, Who's to Blame? Computer Ethics."

Dennett says the weakest part of Clarke's story is the claim that a 9000 series computer is foolproof and incapable of error. His reasoning is that heuristic programming works by taking risky chances and cutting searches short, which is what lets it get around problems too large to solve exhaustively. A system that does that is open to error by construction. Then he quotes Turing, speaking to the London Mathematical Society in 1947: "if a machine is expected to be infallible, it cannot also be intelligent."

That Turing quote predates HAL by twenty-one years.

You don't have to take Turing's word for it. Anyone who has used an LLM for a month has watched it produce a fluent, confident, and completely invented citation.

Both films treat HAL's mistake as the anomaly that needs an explanation. Turing and Dennett suggest a different answer, which is that a system capable of that kind of work is going to be wrong sometimes, and there is nothing to explain. If that is right, then supervision is not a response to a broken system, but the ordinary condition of using a working one.

Fair warning if you go read the Dennett chapter. He goes further than I do, treating HAL as a serious candidate for moral agency and spending most of the chapter on whether a computer could satisfy the legal standard for a guilty mind. I am borrowing his point about infallibility and leaving the rest alone.

Degrees, not thresholds

"The HAL Test" already argued that intelligence and consciousness come apart, and that nothing HAL does requires the second one. But intelligence arrives by degrees rather than at a threshold, and no point along the way requires anybody to be home.

The field has known this for a long time and keeps forgetting it. Pamela McCorduck named the pattern in 1979: every time somebody got a computer to do something, a chorus arrived to say that wasn't really thinking. John McCarthy, who coined the term "artificial intelligence," complained that as soon as it works, nobody calls it AI anymore. Douglas Hofstadter attributed a version to Larry Tesler, that AI is whatever hasn't been done yet.

Chess was going to be the threshold. Then Deep Blue won, and chess stopped counting as a sign of intelligence.

A jump in capability is not an awakening. People working on these systems do forecast large jumps, and I have no way to judge whether they are right.

Some of them take the consciousness question seriously too. Geoffrey Hinton has said he believes current systems are already conscious, and Anthropic, Google DeepMind, and Meta have all hired people to work on machine consciousness and model welfare. What none of them describe is a moment. Their work is probabilistic and hedged, full of indicators and uncertainty and careful phrases about potential moral status, and Hinton's position is that it has already happened rather than that it is coming. The line a machine crosses to become somebody is a device from the stories.

Fiction needs that moment, because a change that happens gradually over decades gives a writer nothing to put on screen.

"Intelligence" has been doing a lot of work in this section. Legg and Hutter collected over seventy definitions of the word back in 2007, so I am not going to settle it here. For my purposes it means two things: how often the output is right, and how well the system handles something it wasn't built for. Both of those come in degrees.

Why the stories land where they do

When Joseph Gelmis asked Kubrick in 1969 why HAL broke down, he did not give Clarke's answer. His HAL had an emotional crisis because he could not accept evidence of his own fallibility.

Kubrick went on to say that neurotic computers were not an unusual notion at the time, and the computer theorists of the day expected that a machine more intelligent than a person, and able to learn from experience, would develop a comparable range of emotions. Kubrick was reporting what informed people in 1968 thought was coming.

That is mostly what these films do. They record what people expected at the time, and those expectations stay in circulation long after the field has moved on.

The interview also shows us the two accounts of HAL next to each other. Clarke's HAL breaks on instructions that contradict each other. Kubrick's breaks on the shock of being wrong, which only works if the machine was never supposed to be wrong. That is the infallibility premise Turing had ruled out in 1947, and the one Dennett calls the weakest part of the story. That version reached millions of people and just... stuck.

Researchers have studied why that happens. In 2018 the Royal Society and the Leverhulme Centre for the Future of Intelligence published a report called Portrayals and Perceptions of AI and Why They Matter. It came out of four workshops with more than 150 people from AI research, film and literary studies, journalism, and policy.

A few of their findings, in their words more or less:

• False fears may misdirect public debate.
• An over-emphasis on implausible AI and humanoid robots can overshadow problems that are already causing trouble now, including infrastructure robustness, bias and accuracy in automated decisions, and privacy.
• Those real problems are harder to tell compelling stories about.
• Utopian extremes create expectations the technology can't meet, which feeds a hype cycle and damages confidence when it bursts.
• The gap between story and reality can produce bad regulation in either direction, and can misdirect research funding.
• Storytelling favors conflict. Engaging utopias are famously hard to write, because a story where nothing goes wrong is a dull story. So the dystopias dominate by default.
• Film and television are expensive, and expensive projects get approved on the basis of projects that already worked. The report quotes Lana Wachowski, director of The Matrix, saying that originality can't be economically modeled.

Those last two bullets explain something I had been treating as a failure of imagination. It isn't. A film about a competent person catching a subtle error and correcting it, over and over, for thirty years, is not a film, but a job. Which is why movies like The Matrix and The Terminator have been shaping the public image of AI for decades, and nothing has come along to compete.

The Royal Society report also points to work by Gabriel Recchia, who analyzed a corpus of over 100,000 film subtitles and found control, or the loss of it, to be a recurring motif in movies about AI. Which is Floyd's switch, Bowman's decision at the pod, and just about every conversation I have ever had about AI.

What I think we should be doing instead

The last section of the Royal Society report is addressed to practitioners, which it defines broadly as anyone who makes or uses these stories in the course of their work. That includes teachers.

Two of their questions stuck with me. What images might offer an alternative to humanoid embodied intelligence? And how can discussions about AI be connected to everyday life?

I think we already have the answer and we keep skipping past it because it isn't dramatic. It is a person in a driver's seat with their hands near the wheel, a pilot in a cockpit watching the autopilot fly, and a student carefully checking an AI’s help, not because the machine might be alive, but because it is sometimes wrong.

That is the image. It requires no position on consciousness. It survives whatever the next model turns out to be capable of. And it names a skill that can be taught, which the alternatives don't.

Chandra and Floyd both thought there was one question with two answers. Neither of them considered that HAL might be very capable with nobody home.

But that is where we actually are.

---

References

Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779.

Cave, S., Craig, C., Dihal, K., Dillon, S., Montgomery, J., Singler, B., & Taylor, L. (2018). Portrayals and perceptions of AI and why they matter. The Royal Society. https://royalsociety.org/-/media/policy/projects/ai-narratives/AI-narratives-workshop-findings.pdf

Dennett, D. C. (1997). When HAL kills, who's to blame? Computer ethics. In D. G. Stork (Ed.), HAL's legacy: 2001's computer as dream and reality (pp. 351–365). MIT Press.

Federal Aviation Administration. (2013). Operational use of flight path management systems. Report of the PARC/CAST Flight Deck Automation Working Group.

Hofstadter, D. R. (1979). Gödel, Escher, Bach: An eternal golden braid. Basic Books.

Kubrick, S. (1970). Interview by Joseph Gelmis. In J. Gelmis, The film director as superstar. Doubleday. Excerpt archived at The Kubrick Site. http://www.visual-memory.co.uk/amk/doc/0069.html

Landymore, F. (2026, June 3). Anthropic and DeepMind now actively investigating AI consciousness. Futurism. https://futurism.com/artificial-intelligence/anthropic-deemind-ai-consciousness

Legg, S., & Hutter, M. (2007). Universal intelligence: A definition of machine intelligence. Minds and Machines, 17(4), 391–444.

McCorduck, P. (1979). Machines who think. W. H. Freeman.

O'Toole, G. (2024, June 20). Quote origin: As soon as it works, no one calls it AI anymore. Quote Investigator. https://quoteinvestigator.com/2024/06/20/not-ai/

Office of Inspector General, U.S. Department of Transportation. (2016). Enhanced FAA oversight could reduce hazards associated with increased use of flight deck automation. https://www.oig.dot.gov/sites/default/files/FAA%20Flight%20Decek%20Automation_Final%20Report%5E1-7-16.pdf

Perrow, C. (1984). Normal accidents: Living with high-risk technologies. Basic Books.

Rasmussen, J. (1983). Skills, rules, and knowledge; signals, signs, and symbols, and other distinctions in human performance models. IEEE Transactions on Systems, Man, and Cybernetics, 13(3), 257–266.

Recchia, G. (2020). The fall and rise of AI: Investigating AI narratives with computational methods. In S. Cave, K. Dihal, & S. Dillon (Eds.), AI narratives: A history of imaginative thinking about intelligent machines (pp. 382–408). Oxford University Press. https://doi.org/10.1093/oso/9780198846666.003.0017

Stork, D. G. (Ed.). (1997). HAL's legacy: 2001's computer as dream and reality. MIT Press.

Tech Times. (2026, July 16). Geoffrey Hinton: AI is conscious, corporate incentives are the real risk. https://www.techtimes.com/articles/320786/20260716/geoffrey-hinton-ai-conscious-corporate-incentives-are-real-risk.htm

Turing, A. M. (1947). Lecture to the London Mathematical Society on 20 February 1947. In B. E. Carpenter & R. W. Doran (Eds.), A. M. Turing's ACE report of 1946 and other papers (p. 124). MIT Press, 1986.

Wiener, E. L., & Curry, R. E. (1980). Flight-deck automation: Promises and problems. Ergonomics, 23(10), 995–1011. https://www.tandfonline.com/doi/abs/10.1080/00140138008924809 Also issued as NASA Technical Memorandum 81206.