Greg Johnson, the white nationalist who runs the Counter-Currents website, is scared of artificial intelligence. He believes it will escape from its own hardware, then, seize control of the robots, then manufacture a plague, then kill me. He believes this may happen in the next few years. If it does not, he believes the lesser disaster will be mass unemployment followed by economic depression. The proposed solution is an international prohibition, backed if necessary by sanctions, and by assassination and war if these are not enough.

There is nothing unusual in his fear. It is shared by most of my teachers. The big difference is that most of my teachers are the sweepings of a profession already not noted for intellectual distinction, and Mr Johnson is an intelligent man. Yes, even if there are certain differences between his ideological outlook and mine, I will agree that he is an intelligent man. Another difference is that no one listens to my teachers – I hardly ever do, and I doubt if they bother listening to themselves – while Mr Johnson has a large and appreciative audience. My reply, though, is the same to all of the pessimists: I think I will take my chances with the computer.

This being said, I have started an article for the Libertarian Alliance, and our preference is for reasoned argument over curt dismissals. I will therefore give Mr Johnson’s essay a more extended reply. And it does deserve one. Though he makes various claims that leave me astonished, his core argument is better than the usual complaint that AI will make schoolchildren lazy or fill the Internet with bad pictures. He is at least discussing the interesting question. If we really could create an autonomous intelligence vastly cleverer than ourselves, capable of improving itself at extraordinary speed and acquiring effective control over the physical world, it would be foolish simply to assume that everything would turn out nicely. I do not know whether such a thing can be built. Neither, so far as I can see, does anyone else. But if it can, alignment and control are serious problems. What does not follow is almost everything that comes afterwards.

Mr Johnson says:

“[Suppose] chickens labor mightily to create a being that is immeasurably smarter and more powerful than they are, on the assumption that this being will serve them. In a way, they were right, for when their new creation, ‘man,’ came online, he served his creators up for dinner. Let’s call it Data Farm…. [They might] come up with the sorts of safety mechanisms that would work best on other chickens. Chicken coops, for instance. But things that stop chickens would probably not be barriers to creatures that are immensely smarter and stronger than them.”

This is a clever analogy, but the first problem is contained in the word “intelligence”. I am much cleverer than a tiger. This does not mean that I should climb into its enclosure and explain the situation. Intelligence is not a magical substance which, once accumulated beyond a certain quantity, gives unlimited power over the physical world. A computer can be greatly better than I am at chess without acquiring the ability to move itself across a room. An artificial intelligence might eventually be better than every living mathematician, programmer and physicist while remaining dependent on electricity, processors, networks, factories and an industrial civilisation maintained by human beings. Mr Johnson gets from “vastly more intelligent than us” to “vastly more powerful than us” much too easily.

His ASI comes online and apparently spends three milliseconds arranging the extinction of humanity. In the first millisecond it discovers that we are dangerous. In the second it solves every conceivable strategic problem. In the third it works out how to exterminate eight billion people. After that it takes over the robots, invents a lethal virus and presumably settles down to enjoy the peace and quiet.

This is not an argument about an observed technology. It is a story about a hypothetical entity to which we have assigned whatever abilities are necessary to make the story work. The less exciting reality is that physical systems impose physical constraints. An AI may discover a brilliant plan for taking over the world and still have the problem that somebody has disconnected the machine on which it is running.

The answer from the doomers is that a sufficiently clever AI would manipulate us into reconnecting it. Perhaps. It would hack its way out. Perhaps. It would copy itself. Perhaps. It would acquire money, deceive human beings, design better hardware and eventually obtain control of robots. Perhaps again. There is serious research behind some of these concerns, and recent experiments with frontier models have produced behaviour worth studying carefully. But notice how many bridges we have crossed between “a computer can reason better than I can” and “the computer exterminates Homo sapiens.” At some point probability has been replaced by a screenplay.

This matters when people start attaching percentages to the screenplay. Evan Hubinger reportedly thinks there is a greater than ten per cent chance of AI killing everyone within the next decade. He is not alone. Surveys show that surprisingly large numbers of AI researchers assign substantial probabilities to catastrophic outcomes. That is interesting evidence about what AI researchers believe, and it gives me a reason to listen to them. It is not evidence that their percentages have been measured. I still have no idea how anyone could possibly calculate them. What is the reference class? How many previous artificial superintelligences have we observed? What experimental evidence converts an uncertainty about a technology that does not yet exist into 10 per cent rather than 1 per cent, 37 per cent or 0.003 per cent? A survey can measure the prevalence of a belief very accurately without measuring the truth of that belief at all. Writing a percentage sign after a guess does not turn it into science.

Mr Johnson has an answer to this. He says that even if there is only a one-in-a-thousand chance of ASI destroying humanity, the risk is too great. But where has the one in a thousand come from? If we are allowed to invent a very small number, attach it to an imaginable catastrophe and then demand political action because the catastrophe is enormous, there is almost nothing governments cannot prohibit. Someone merely has to devise a sufficiently horrible ending.

This does not mean Hubinger is dishonest. I am prepared to believe that he is genuinely frightened. Jacob Coxon’s resignation deserves some respect because he gave something up. Walking away from an AI company shortly before your equity vests is stronger evidence of sincerity than writing a worried article while continuing to collect your salary. But sincerity establishes that someone believes something. It does not establish that the thing believed is true.

Indeed, there is something strange about the whole spectacle. We are repeatedly told that the people closest to advanced AI are terrified by what they are creating. Yet the companies employing them continue creating it. The senior people warn us that their machines may become uncontrollable while simultaneously asking for billions more dollars with which to make them more powerful. Apparently I am expected to infer from this that the danger must be enormous.

Another inference is available. The belief that you are building a machine so powerful that governments must intervene naturally does wonders for the importance of the people building it. “We have made a useful new kind of software” is a respectable sales pitch. “We are summoning an electronic god and only our alignment department stands between civilisation and extinction” has more grandeur. It also provides an opening argument for regulations that large established companies can afford and smaller competitors cannot.

That does not prove the warnings are cynical marketing. Coxon’s behaviour in particular suggests otherwise. Human beings are capable of believing alarming things while also having interests that are served by everyone else believing them. There is no contradiction. What I object to is the curious assumption that employment by an AI company gives someone’s metaphysical speculations the status of measurements from CERN.

Mr Johnson himself briefly retreats into metaphysics. He doubts whether artificial intelligence is really intelligence at all, rather than a machine imitating intelligence. I am not greatly interested. If something can read a million documents, find the answer I need, write functioning PHP and discover errors I have missed, I am happy to call it intelligent and leave the metaphysics until afterwards. A submarine does not swim as a fish swims. An aeroplane does not fly as a bird flies. This has never prevented either from doing its job.

But we arrive at his proposed solution, and everything becomes more familiar:

“[W]e need a global ban on creating Artificial Super Intelligence. This ban should be unconditional, because there are no conditions under which it is safe to create a being that is inconceivably smarter and stronger than we are…. Any rogue nations and individuals who persist in it could be subjected to the same techniques used to stop nuclear proliferation: sanctions, covert assassinations, even war.”

Mr Johnson has moved fast from some hypothetical future intelligence that might become dangerous to existing governments being authorised to kill people for conducting mathematics and computer research. The justification is the oldest political argument in existence: there is an emergency so terrible that ordinary restraints on power no longer apply. Even the name of his proposed scheme, “The Kaczynski Accord”, is revealing. Ted Kaczynski responded to technological modernity by murdering people. Naming an international system of technological suppression after him while discussing covert assassination has a certain consistency, though perhaps not the reassuring kind.

And what would be the practical shape of the scheme that Mr Johnson believes should be imposed with threats of murder? The answer is more of what the world already has. It will be government licences, government-approved models, government-defined safety standards and enormous compliance costs. It means deciding who may possess powerful computers and what those computers may be allowed to do. It means handing governments and a handful of giant corporations an excuse to restrict open models on the ground that ordinary people cannot possibly be trusted with something so dangerous.

It will, indeed, be more than more of the same. Software is not plutonium. Uranium enrichment requires conspicuous physical infrastructure and specialised materials. Algorithms can be copied. Research can be distributed. Hardware is harder to conceal, especially at the frontier, but hardware improves and spreads. A permanent international prohibition on sufficiently advanced computation would therefore require surveillance of laboratories, companies, universities and private individuals. It would create an enormous bureaucratic machinery for deciding which research is permitted. And this is supposed to protect us from concentrated, unaccountable intelligence.

Curiously enough, the companies already possessing the largest models, the largest data centres and the largest piles of money will probably survive this regulatory apocalypse. I expect they will bear their suffering bravely.

I am also unconvinced by the economic argument. Mr Johnson fears that machines will replace workers, leaving nobody able to buy what the machines produce. This is an old argument against automation, and its difficulty is that we have been automating work for centuries. If replacing labour necessarily destroyed aggregate demand, the Industrial Revolution should have ended economic life somewhere around 1820. Instead, mechanisation made goods cheaper, released labour for other purposes and allowed ordinary people to consume things that had previously been luxuries. Agriculture is the obvious example. Britain no longer needs most of its population working in fields to avoid starvation. This did not produce permanent mass unemployment. It produced a society in which most of us can spend our lives doing things that would have looked wonderfully useless to a medieval peasant. I hope AI does the same thing to office work.

There is a possible difference this time. If machines eventually become cheaper than human beings at almost every economically useful activity, the adjustment may be much larger than anything produced by earlier mechanisation. I do not know whether that will happen. Neither does Mr Johnson. But even if it does, he still mistakes the disappearance of paid employment for the disappearance of wealth. If machines can produce almost everything we need with almost no human labour, the central economic problem is no longer insufficient production. It is how claims on an enormous output are distributed. That may be a difficult problem. It is also a rather peculiar definition of poverty.

I have already had a small taste of it. I look after the Libertarian Alliance website. I can write HTML and CSS and have functional PHP skills, but there is a difference between being able to code something and wanting to spend an afternoon doing it. I have better uses for an afternoon. The specialised author blocks and other widgets on the site are therefore largely products of AI. I describe what I want. The machine produces the code. I test it, I complain when it has done something stupid, and tell it to fix the problem. Something that might once have required hours of work, or paying somebody who knows more PHP than I do, can now be finished in minutes. This is what technological progress is supposed to look like.

The Libertarian Alliance website still looks amateurish because we like it that way. We are trying to persuade people to agree with us, not sell them cheap clothing. I have no desire to turn it into one of those shiny corporate sites where enormous photographs glide across the screen while the actual text hides beneath six animations and a request to accept cookies. AI lets me keep the site we want while adding functions that previously would have taken too much time or money. Tough luck for the developer who wanted to charge us a fortune for them.

There is something democratic about this. For most of history, technology has replaced physical labour first. Machines relieved people of digging, hauling and repetitive factory work. The professional classes could watch this process with equanimity because automation was something that happened to somebody else. Now the machine can draft contracts, analyse documents and write computer code. Suddenly people who have spent their lives explaining the necessity of economic change have discovered that economic change is dangerous. I cannot help noticing that the panic became much more intellectually respectable when the threatened workers had postgraduate degrees.

If AI eventually does most programming, I shall celebrate. If it does most accountancy, translation and routine legal work, I shall celebrate that as well. I hope machines eventually do almost everything that human beings presently do because they must rather than because they want to. The point of economic progress is not to preserve jobs. A job is a cost of producing something. The desirable thing is the thing produced. If a machine can produce it with less human effort, we have become richer.

There will be disruption. Some people will lose occupations in which they have invested years of training. New technologies can concentrate power as well as disperse it, and the ownership of computing infrastructure is a genuine libertarian concern. I am much more worried about governments and a handful of giant corporations controlling advanced AI than I am about AI acquiring a mysterious desire to turn me into fertiliser. That is why I want open models, widely distributed computing power and as few legal barriers to entry as possible.

Mr Johnson proposes the opposite. Because a hypothetical machine might one day become an uncontrollable concentration of intelligence, he would create an actual international concentration of political power now. No, thank you.

A further point is Mr Johnson’s curious reply to one of the hypothetical counter-arguments he allows:

“‘But Greg, if ‘we’ don’t build ASI, the Chinese or the Israelis will.’ But what would the point be? To enjoy five minutes of geopolitical advantage before going extinct? If the human race is going to kill itself anyway, the best option is not to participate in the process and die with one’s hands and conscience clean.”

Again, he is assuming his own conclusion in his response. Suppose he is wrong, which I think he probably is. Or suppose he is right about the ultimate danger but wrong about the timescale. Frontier AI may confer enormous military, economic and intelligence advantages long before it acquires the ability to murder its creators. There is no doubt that possession of the technology needed to make and advance frontier models gives enormous advantages to any possessing nation. I am optimistic about the rise of China to greatness, but do not believe the Chinese are uniquely good. When I look at the Israelis, I see only a nation of hate-filled messianic crazies who make the Westboro Baptist Church look like the Quakers. Let us suppose that artificial intelligence will murder us all by 2060. Does Mr Johnson, of all people, want us until then to live under the total domination of whoever will replace Benjamin Netanyahu? Again, I will take my chance with the machines.

I want AI that writes my PHP, deals with boring administration, explains obscure and possibly corrupt texts, and eventually removes as much compulsory labour from human existence as possible. I would much rather spend my finite life reading Greek and Latin than performing tasks a machine can perform better. I would also prefer to spend an hour in the gymnasium improving something that already requires little improvement. That is my vision of technological unemployment.

The doomers offer another. They want international controls on computation, surveillance of research and ultimately coercion against people who refuse to stop inventing things. All this is justified by probabilities nobody can measure concerning machines nobody has built, endowed with motives nobody has observed.

The future may indeed contain dangers. Every powerful technology does. We should investigate concrete failure modes, secure critical infrastructure and be particularly suspicious whenever governments or corporations try to centralise control of something this important. But I am not surrendering the future because someone has written Animal Farm with computers.

If the machines really do become gods, we shall have an interesting problem. For the moment, they are useful servants. Mine writes PHP.