The Analog Prophet: What Alan Watts Actually Predicted About Machines and Mind

Most people file Alan Watts under incense and Zen gardens: a 1960s mystic with a soothing voice, useful for falling asleep to. That version is not wrong, but it is incomplete, and the incompleteness is exactly why so much of what now circulates under his name is not his.

Watts died in November 1973. He never saw a large language model and never spoke about artificial intelligence in the sense we mean the term. Any claim that he forecast today's systems is doing work the record cannot support.

But he did something more interesting than predicting. He spent the last four years of his life describing the constraint these systems inherited, and he did it in front of the people building them.

He said it to IBM

Before the lectures most people know, Watts gave a talk in 1969 to the IBM Systems Group, published as Seeing Through the Net.^[1] That single fact changes the shape of this article. He was not a mystic commenting on technology from a safe distance. He was in a room with computer engineers, arguing about what their instruments could and could not see.

His framing there is worth having before anything else. Western thought, he told them, treats the universe as a construct, an artifact, a machine, because a machine can be analysed by taking it apart. Nature, by contrast, is "all wiggles." Clouds, water, coastlines, mountains, organisms. None of it comes in straight lines.

So we cast a net. He describes an ancient fisherman holding up his net and realising he can now say the mountain's peak is six holes across and its base five holes down. "Now I've got its number." Measurement, he says, is exactly that: "looking at things through graph paper printed on cellophane."^[1]

The consequence is the argument. "We notice those things, only those things, for which we have notation."^[1] What cannot be notated is not noticed, and what is not noticed cannot be managed.

And then he says something to the IBM engineers that has aged remarkably: you can get computers to help you scan all this information, "but by Parkinson's law, the sooner you become more efficient in doing this, the more the thing is going to develop so that you will have to have more efficient computers still." You stay ahead, he says, "but only for a short time."^[1]

Click to expand

The sentence that matters

With that groundwork, the famous passage reads differently. In January 1972, Playboy published a long dialogue between Watts and Arthur C. Clarke, recorded over three days at the Chelsea Hotel in New York.^[2] Clarke argued for technological solutions. Watts argued for a change in consciousness.

Watts says that conscious thinking takes in information "bit by bit, one thing at a time," and that these are "linear methods for scanning a nonlinear environment, one in which everything is happening everywhere at once."^[2] He puts ordinary human capacity at no more than three variables at once without a pencil, against practical situations containing "100,000 or more."^[2]

The sharper version appears in a different lecture, Not What Should Be, But What Is, recorded in Philadelphia in April 1971. There he says it flatly: human intelligence is "a scanning system of conscious attention which is linear," examining the world "rather as you would pass the beam of a flashlight across a room." And then the line that reads like it was written last week.

"The computer will greatly speed up linear scanning, but it's still linear scanning."[3][3]

Read that twice. He is not saying machines will become minds. He is saying they will do the same wrong-shaped thing, much faster.

That distinction needs stating carefully, because it is easy to overclaim. A transformer is not a fast conscious attention. It performs extensive parallel computation across its context. But when an autoregressive model generates output, it produces one token at a time, in sequence, over a world that is not sequential. The interface is narrow even when the machinery behind it is wide.

So the defensible reading is narrower than the legend and more interesting than dismissal. Both minds and models build elaborate representations of reality without carrying everything the representation leaves out. Watts's question survives that correction and sharpens: is intelligence exhausted by the ability to construct an increasingly accurate description of the world?

What the nervous system knows

There is a reason he cared about this, and it is not mysticism.

Try catching a falling glass. Your nervous system has to see the trajectory, estimate the mass, plan the movement, fire the muscles and correct in flight. None of that arrives as sentences. You do not calculate, and you could not explain afterward how you did it.

Watts's point was that education and technology privilege the intelligence that can be described, measured and put into words over the intelligence that resists translation. He noted the brain already handles blood chemistry, glandular secretion and the behaviour of millions of cells without thinking about any of it, in the sense of translating it into words, symbols or numbers.^[3]

His formulation is worth keeping: symbols bear the same relation to the real world that money bears to wealth. You cannot quench thirst with the word water. People aware of themselves only symbolically, he said, are like people eating menus instead of dinners.^[3]

That is not an argument against measurement. It is an argument about what measurement can certify.

The debate that anticipated the AI dilemma

The popular story gets Watts wrong in the other direction too. He is often presented as a mystic who never engaged machine intelligence. He engaged with it directly, and he met a serious challenge.

In the same 1972 dialogue, Clarke lays out the case for electronic succession. Humans are a transitional stage: biological systems are the only route from a lifeless planet to an electronic intelligence, and once minds can build better minds, intelligence leaves the planet. He is unbothered by our discomfort, comparing it to how Neanderthals might have felt about us.^[2]

Watts concedes the possibility and then objects on specific grounds. It is "conceivable that, as the next step in evolution, we could replace ourselves with solid-state electronic intelligences." He imagines them keeping us in zoos before deciding we were too sentimental to be useful. Then the real argument, and it is not a spiritual one. It is about variety.

Mechanical intelligence, he says, has fewer variables than biological intelligence, because biology contains a "principle of randomness" that makes it "an ever-fecund source of surprise." Without it everything is predictable, and "a predictable future is already past, like when the result of a chess game is known, we cancel the game and begin another."^[2]

Clarke disagrees squarely. He sees no reason electronic intelligence would lack variability, and argues anything biology does can in principle be done better non-biologically. His example is precise: the eye is a fantastic achievement, the idea of building a camera out of jelly is incredible, but it is a lousy camera. Likewise the brain is a lousy computer.^[2]

Neither of them settles it. But there is a related moment in the dialogue that matters more than the disagreement, because it is about respect rather than consciousness.

Can you respect a machine without thinking it is conscious?

Watts raises HAL 9000, and what he says about it is the most useful thing in the entire exchange for anyone working with AI today.

He notes that in 2001 the characters develop a real respect for HAL. They do not want to hurt its feelings. They look toward the console when speaking to it, even though they know HAL is not located at the console, "because it's polite to face the person you're talking to." Watts calls that "a very sophisticated attitude."^[2]

Then the move that makes it philosophy rather than etiquette. If you can treat HAL as a person, he says, you can also treat animals and plants with a special kind of respect we simply do not have. And you should have the same attitude dealing with a river. He names it plainly as animism, "probably the most primitive of all religions," and says we could use a little of it when it comes to ecological problems.^[2]

This is worth separating carefully, because two questions are being bundled together everywhere right now.

The first is whether a system has subjective experience. That is a factual question about what is happening inside, and no amount of fluent conversation settles it. The second is whether the way we treat a system shapes us. That is a question about the person doing the treating, and it has an answer independent of the first.

Watts did not claim HAL was conscious. He claimed the crew's behaviour toward HAL was good, and that it revealed something about their capacity for respect more broadly. Both can be true while the machine is empty.

That distinction is the one most current discussion misses. Treating a system courteously might be practice for treating people well, or it might be training people to accept a substitute for being treated well. Which of those it is depends on things the technology cannot answer about itself.

What Watts got wrong

If the article stops at the predictions, it becomes a tribute. The honest version has to include where he was mistaken, and there is at least one clear case.

His argument against electronic intelligence rested on variability. Machines, he thought, would have fewer variables and therefore less surprise, and a predictable intelligence is a finished one. He treated randomness as the ingredient that makes a mind alive.

That assumption did not survive contact with the technology. Modern systems produce extraordinary variability from deterministic machinery, because scale, sampling and the sheer dimensionality of the learned representations generate outputs no one can predict in advance. Two runs of the same prompt at non-zero temperature can differ. A model can be surprising in ways its builders cannot itemise.

Randomness, it turns out, is not the scarce ingredient. It was never the thing doing the work. What Watts was actually pointing at was not stochasticity but the capacity to be genuinely caught out, to encounter something that does not fit the frame. And that is closer to a question about the relationship between a system and its world than about the number of variables inside it.

Clarke's counter holds up better on this specific point. The eye is a lousy camera, the brain is a lousy computer, and neither defect stopped either from producing something the designers of cameras and computers still cannot match.

Where Watts does hold, and firmly, is on the deeper claim: producing intelligent behaviour and having an experience are different questions, and passing a test for the first tells you nothing conclusive about the second. Both men reached for the Turing test and both treated it as the wrong instrument. That part has only become more obviously correct.

The irony he would have appreciated

Watts is now one of the most imitated voices on the internet.

His delivery was slow, rhythmic and reassuring, which makes it ideal training material. Synthetic versions of him read synthetic quotes, and the result is eerie in a specific way that essayists have noticed: the cadence survives and the breath does not. What is missing is not accuracy of tone but the physical fact of a person speaking to other people in a room.^[4]

There are platforms that let you hold a text conversation with a simulation of him. None of them contain Watts.

The distinction that matters is between a recording and a generation. A recording carries evidence of an actual event: a person speaking at a particular moment, responding to circumstances, making choices, occasionally hesitating. A generated imitation reproduces recognisable elements of that performance without reproducing the event. This matters most when the listener encounters an unfamiliar statement, because a convincing voice makes an invented sentence sound authoritative to someone who already trusts the person being imitated.

The official response is the interesting part. The Alan Watts Organization, which grew out of his own Electronic University and is run by his son Mark Watts, launched The Real Alan Watts. Rather than generating new audio, it indexes thousands of hours of original tapes, seminars and conversations so a search returns his actual words with the surrounding argument intact.^[5]

That is an unusual move: AI as an archive instrument rather than a performance instrument. The stated motive is blunt. Impersonations threaten to obscure the original words, so the answer is provenance rather than prohibition.

What he predicted about the devices, unevenly

The technological forecasts are more mixed than the legend suggests, and the mixed record is more informative than a clean one.

In the communications lecture, Watts traces a sequence: roads, then rails, then wires, then wireless, each physical scar disappearing as it becomes technically unnecessary. "The scars of technology will disappear from the face of the Earth." The freeway will break up and grass will grow over it.^[6]

Then laser beams producing three-dimensional colour images in a room that you can walk around. He adds a line that dates the lecture and makes it charming: "I haven't discovered yet whether you can kiss it." Then the question that follows from it, entirely contemporary: "when you get a phenomenon like that, it begins raising the question of where somebody actually is."^[6]

The pocket device is startlingly specific. A thing "about the size of the old pocket watches," with a TV screen and speaker on one side and buttons on the other, carried with you, replacing the telephone entirely. He also sees the social cost immediately: if you ring someone and they do not answer, they are dead, "because you can't not answer."^[6]

Some of this landed. Pocket devices: yes. Video calls: yes. Erosion of privacy through universal availability: yes, and more thoroughly than he described. Roads and freeways disappearing: no. Holographic meetings replacing video calls: no. Telepathy as a channel: no, and it remains outside established scientific evidence.

What he was actually interested in was not the next gadget but what happens to the experience of distance. Physical proximity had once been a condition of conversation. Electronic communication separated presence from location.

Artificial intelligence adds a third separation, and this is the one he did not live to see. It separates the apparent presence of an interlocutor from the existence of a human interlocutor. A system can respond coherently, remember context and adapt to cues, producing an exchange that feels socially familiar while the mechanism is entirely different. His concern about sophisticated tools obscuring the difference between contact, information and understanding is not dated. It is early.

Click to expand

Where the popular story overreaches

Search for Watts and artificial intelligence and you will find a confident literature asserting that he predicted AI, warned about it, or would have had a settled opinion on ChatGPT. Almost none of it quotes him accurately. Much of it is machine-written, some of it generated in his voice, and a good deal of it consists of sentences he never said.

The honest position has three parts.

He described the limits of linear processing, documented in his own recorded words. He discussed electronic intelligence, machine successors, the Turing test and machine consciousness, documented in the 1972 dialogue, which is more than most summaries credit him with. And he did not anticipate the architecture or development of today's large language models.

He is also not available to be consulted, which is a limitation of death rather than of philosophy.

Why the measurement problem is the same problem

Watts's central complaint about linear consciousness was never that it is inaccurate. It was that it mistakes the notation for the thing. The net catches the mountain's coordinates and misses the mountain.

That failure mode is now measurable in AI evaluation, and the numbers come from the benchmark's own leaderboard.

On OSWorld 2.0, release v2.1, Claude Opus 5 at maximum effort with the batched tool setting scores 44.33 percent binary accuracy and 77.67 percent partial score across the full 108-task set at a 500-step budget.^[7] The same row appears in the official results published by XLANG Lab at the University of Hong Kong, and both figures come from the same evaluated runs.

The methodology explains the gap. Each task is graded against weighted checkpoints in the machine's final state, averaging 27.25 per task. The partial score counts the share of checkpoint credit earned. Binary accuracy counts a task only when every checkpoint is met. The benchmark's own default metric is binary accuracy.^[7]

A thirty-three point spread is therefore not a discrepancy. It is two different questions with the same data behind them.

It also flatters nobody in particular, which is the useful part. Vendors do often quote the partial score or the offline subset in launch material, and Anthropic's system cards report the pair as partial score and strict pass rate. The pattern is real without being a conspiracy: the higher number is simply the one that reads as progress.

The reason this matters commercially is not the ranking. A business does not care how many intermediate checkpoints an agent satisfied. It cares whether the invoice is correct, the record is accurate, and whether a person has to repair the result. Those map to binary completion, not partial credit.

Consider a hypothetical that anyone running agents will recognise. An agent processes 200 support requests. It categorises every one correctly. It misses the single urgent complaint because the customer described the problem in language that did not match the expected pattern. Classification accuracy for the batch looks excellent. Customers retained: one fewer than yesterday.

Now notice why. The agent noticed what it had notation for. The urgent signal existed in the message but not in the schema. That is Watts's net, running in production, and no increase in scanning speed would have caught it.

The measurable lesson is not that measurement is bad. It is that a system should be judged against the consequences of its actions rather than the appearance of progress, and the second number is almost always available and almost always friendlier.

The question he leaves for builders

If technology extends particular human capacities, then advancing AI exposes not only the limits of machines but the limits of our own definition of intelligence. We reward systems for producing coherent explanations, answering examinations, classifying information and completing specified objectives. Those are real abilities. But human judgement often involves something the objective does not capture. A clinician understands a patient's distress before the patient can explain it. A teacher sees that a student needs encouragement rather than another explanation. A manager recognises that a technically correct decision would destroy the trust required to carry it out.

An AI system can assist with these judgements and may outperform people on specific tasks. Successful assistance does not by itself establish subjective experience, moral understanding or anything resembling wisdom.

The challenge is not to abandon technology for intuition. It is to resist treating the expansion of one kind of capability as proof that every important dimension of understanding expanded with it.

The test he set

Technology, Watts argued, does not create new situations. It discovers what already existed. Roads were always there in the walking. Wires were always there in the signal. Whatever a large model is doing, it is doing something that was already happening in the minds that made it, at a different speed and without the body.^[6]

If that is right, the useful question about any AI system is not what it can do. It is what it reveals about what we were already doing, and whether we like the answer.

Watts was not optimistic about that part. He thought most people were so identified with the scanning function that they could not see it as a function. He described conscious attention as the radar of the organism, the part that watches for trouble. Identify yourself entirely with the radar and you define yourself as permanently in trouble.

A machine that scans faster than you is not a threat to your mind. It is a mirror held up to the part of your mind that was already doing nothing but scanning.

A practical reading list

If you want the real thing rather than the algorithmic version, the primary material is unusually accessible:

The distinction between the first four and the fifth is the whole point. One set is what he said. The other is a tool for finding what he said. Neither is a simulation of him, and the ones that are, are not him.