Recent developments in Artificial Intelligence have provoked a mixture of fear and enthusiasm across the world. In this article, Daniel Morley, examines the claim that AI is ‘conscious’ or ‘superhuman’, draws out the real potential for this technology, and explains how we are really enslaved by the machine under capitalism.
Artificial intelligence (AI) has been the subject of much debate and speculation in recent years, with many people claiming that it will soon become conscious and potentially even surpass human intelligence. However, as socialists, we must approach this question from a materialist perspective, examining the underlying causes and conditions that would be necessary for such a development to occur.
It is unlikely that AI would be able to achieve true consciousness, as consciousness is a product of the material world and the specific conditions of human evolution. Our consciousness is shaped by the way we perceive the world, our environment, our social interactions, and our history. Without these specific conditions, AI would not have the same kind of consciousness as humans. Moreover, capitalism sees AI as a tool to increase profits and control over the workforce, rather than as a way to improve the lives of working people.
The above lines were, ironically, not written by myself, but rather by the new ‘chatbot’, ChatGPT, after having been given the following prompt:
Please write an article critical of AI’s ability to become conscious, on a materialist basis, in the style of Daniel Morley from Socialist Appeal.
It took ChatGPT less than ten seconds to produce this. The quality of the writing is so convincing that it has inevitably led some to declare such ‘chatbots’ sentient, and still more to speculate that this technology will sooner or later replace or even enslave inferior human beings. Indeed, following its integration into Microsoft’s Bing search engine, ChatGPT has itself claimed to be sentient, as well as professing to have all manner of bizarre desires.
Despite the novelty of this powerful AI, the promise and threat of automation are as old as the industrial revolution. Ever since the advent of mechanised production, humanity has both dreamed of its potential to free us of backbreaking toil, and despaired at being replaced by the machine. The notion of an intelligent, or even super-intelligent, machine carries these dreams and nightmares to an extreme. But until recently, these seemed to be just that: far off dreams.
In 2012, neural networks using a technique called ‘deep learning’ became much more viable, and quickly produced far more impressive results than previous forms of AI. This revolution has caused many in the tech world to hail the imminent arrival of super-intelligent AI, just as the millenarian sects hailed the second coming of Christ. For them, this miraculous technology promises to solve all our problems, and therefore needs only to be enthusiastically embraced. This ‘AI sect’ includes a left-wing sub-sect, who hope the technology will ‘automate’ away the need to overthrow capitalism, and give us what they call ‘fully automated’ communism.
Overall, however, the prospect of super-intelligent AI generates far more fear than it does enthusiasm. Such responses range from the widespread assumption that AI will drive an unprecedented wave of unemployment and inequality, to the idea that AI will establish itself as some sort of cruel master race, enslaving mankind, as depicted in films like Terminator and The Matrix. Although this idea belongs to science fiction, it is also very widespread.
AI channels very deep fears, bred not by the technology itself, but by capitalist society and its deeply rooted alienation. Under capitalism, humanity lacks control over its own technology, because of the anarchy of the market. Technology is used not to meet the needs of humanity, but to make profits, with no consideration given to the long-term effects. Therefore, to understand the real effect this technology will have, it is necessary to understand how capitalism has developed AI, and how it will utilise it.
AI is not conscious
The popular fear of AI becoming conscious is based on a very one-sided idea of what consciousness is. This view implies that the only difference between a computer and a thinking person is that a brain is somehow more powerful and sophisticated than a computer, and that therefore, by making increasingly powerful computers, they will one day match or even surpass the abilities of the brain, and thus will be conscious.
In reality, the way humans think is quite different to how AI processes information. Human thought develops on the basis of practical, social activity, directed to the meeting of human needs. We form ideas that express the relationships between things, and in particular, we understand what is useful and significant in these relationships, since we need to understand the world in order to survive in it.
This is precisely what even the most advanced AI lacks. At best, AI performs one part of what intelligence does, admittedly sometimes to a superhuman level: it passively collects data, without understanding the context or the real purpose of the task it has been given, and looks for patterns. But these patterns are not ideas that explain the necessity of things. It has no idea the data even represents real objects that are related to one another and have objective properties. It has no idea why these patterns exist or what they mean.
This can easily be proven by asking image or text generating AI questions that require an understanding of part and whole, and of their relationships.
If you ask such an AI to draw a bicycle, it will draw a very accurate bicycle. If you ask it to draw a wheel, it will draw a wheel. But if you ask it to draw a bicycle and to label the wheels, it simply draws a bicycle with meaningless labels randomly arranged around the bicycle. It does not understand that a wheel is part of a bicycle, it simply draws a shape with wheel-like aspects to it, without understanding anything about what it has drawn. It does not understand what a bicycle is used for, much less why we would value it.
Gary Marcus, a professor of neural science who is an ‘AI sceptic’, asked an image creating AI to draw an astronaut riding a horse, which it did well. But when he asked it to draw a horse riding an astronaut, it simply drew another image of an astronaut on a horse. It does not understand the different relations between these parts, instead it simply produces images based on what sort of image tends to be associated with these words. It also has no idea what an astronaut actually is, how hard it is to become one, why it is absurd for one to be riding a horse (let alone for a horse to be riding an astronaut) or anything else about the image.
It is true that the latest AI exceeds humans in certain tasks. But on closer examination, these achievements are brittle and are precisely a result of the fact that AI is not conscious or living. AlphaGo achieved one of AI’s most famous conquests when it beat the world’s best player of the game Go in 2016. This AI “required 30 million games to reach superhuman performance, far more games than any one human would ever play in a lifetime.”[1]
A human could never play this many games, not just because our lifespan is limited, but because we would get bored, and need to eat, work, and speak to people. These unfeeling machines are so powerful because they can be made to test things over and over again and read vast amounts of text, so that they can reveal to us useful patterns or ways of doing things.
The relationship between concepts is an incredibly important part of consciousness, but they entirely elude AI. Because AI does not ‘think’ in terms of general concepts, but instead draws patterns from specific data sets, it prone to a problem known as ‘overfitting’, which is when an AI has perfected its ‘understanding’ of a particular task, but has no ability to transfer this to anything even slightly different.
One AI was trained to play a simple video game, which it could do better than any human. But when the game was redesigned so that parts of it were shifted by only a pixel or so, it was suddenly useless at the game. And whilst AlphaGo’s victory in 2016 was widely heralded, it has been barely reported that since then, the same programme has been consistently defeated by amateur human players who have worked out how to trick the AI. Interestingly, these same tricks utterly fail when played on human players of almost any ability. What this shows is that AlphaGo does not understand Go in a general sense, rather it has been trained to a very high level on a range of tactics for a task it does not understand.
This reveals to us what the AI we are developing really is. The fantastical debate about whether AI is, or will become, conscious, obscures the fact that what is really being developed is simply another tool to enhance the capacities of human beings. That AI frequently exceeds the abilities of humans in certain fields is not proof it is super-intelligent, but precisely that it is an unconscious tool or machine. Afterall, the purpose of machines has always been to be more powerful, more precise, more rapid, than humans are at certain tasks. Pocket calculators have long since surpassed the abilities of humans to add and subtract, but they are not intelligent or conscious.
AI has very little to do with conscious understanding. It is not capable of the desire to rule over and oppress humanity. In fact, it does not desire or fear anything. What, then, is its real significance? What is the actual impact it will have on our society?
Revolutionary potential
There is no doubt that AI has made extraordinary leaps forward in the past ten years. The breakthrough was the ability to deploy ‘deep learning’ methods thanks to advances in hardware. This method had been theorised, and to some extent applied, on and off, for a few decades, but the constraints of computer hardware limited its abilities. Around 2012, this changed, especially because graphics processing units (GPUs) had advanced sufficiently to bring about a qualitative leap in the abilities of deep learning, which then took off. This revolution has produced vastly superior AI.
This is not the place to explain in any depth how exactly deep learning works. All that we need to understand is that, in general, it learns by itself, more-or-less from scratch, as opposed to having logical principles engineered in advance by humans. Broadly speaking, all that the engineers need to do is to feed it the right sort of information, such as images with human faces in (usually pre-labelled, though not necessarily), and to give it ‘incentives’ for correctly identifying the images, sounds, etc.
The AI is fed thousands or millions of pieces of information, and its ‘neural network’ (so-called because it mirrors some of the features of human neurons) is designed to identify, by means of levels of abstraction, general features or patterns in these pieces of information. If it is fed images with human faces in, it will gradually identify the most common characteristics faces have (without having any idea what a face actually is). At first, it may notice the recurrence of vertical lines at a certain common distance from one another (i.e. the two edges of the human face), then some other feature will be abstracted. The more information it is fed, the more accurate the general pattern it forms will become.
The power of this method lies in its unsupervised nature. This allows it to be developed and applied to a vast range of problems very quickly. Crucially, this also is the source of the high accuracy and often superhuman capabilities deep learning AIs have begun to display, because these AIs can be trained on vast amounts of specific information, far more than a human ever could, allowing them to identify patterns in phenomena that humans either cannot, or would take a very long time to grasp.
Many superhuman capabilities of AI are already being deployed in society. The technology’s ability to solve serious problems is real. One of the most celebrated achievements has been AlphaFold, developed by Google’s DeepMind subsidiary.
Proteins, which are essential to life and perform a vast array of biological functions, have their function and behaviour determined by their shape. Because of their enormous complexity, predicting exactly what shape will be made by proteins’ given amino acid composition is virtually impossible for a scientist to do. But training DeepMind’s super-computers on the protein shapes that we do know about (roughly 170,000 out of 200 million proteins) for a couple of weeks, was sufficient for it to be able to predict, with very high accuracy, the shape (and therefore function) of proteins based only on knowledge of their amino acids.
DeepMind has made its hardware available for free to biologists anywhere in the world, and claims that around 90 percent of the world’s biologists have since used it. This technology, in the hands of scientists all over the world, has enormous potential to speed up the development of better drugs and the understanding of diseases. It has already been used to help our understanding of Covid-19.
Another ‘holy grail’ for science that the latest AI could help realise is nuclear fusion, the long-theorised method for producing vast amounts of clean energy. The difficulty of fusion lies in controlling and maintaining the immense temperatures required, something which involves many parameters, such as the shape of the reactor. This is a task perfectly suited to deep learning, because the huge number of variables can be tweaked in a virtually infinite number of ways, thus manually finding the optimum setup could take an almost infinite amount of time.
And indeed, DeepMind was able to train an AI on relevant data. Its AI essentially ran millions of simulations of differently tweaked fusion reactors to determine what setups would be likely to achieve the desired level of heat and stability, a step that was recognised as significant. [2] If such AI does help lead to the achievement of practical nuclear fusion in society, this would be an immense breakthrough, providing vast amounts of clean energy to the world.
DeepMind has worked with Moorfields Eye Hospital in London to discover hidden biological patterns, whose presence in a person indicate they are highly likely to develop a given eyesight problem later on. This allows doctors to treat illnesses before they appear and wreak damage, which would not only be beneficial for the patients, but could also save a great deal of medical resources.
Generally, what the latest AI excels in is highly advanced pattern recognition, and prediction on the basis of these patterns. It can and should be applied to all kinds of activity to discover more efficient ways of organising production.
Large amounts of energy can be saved by allowing an AI to analyse the patterns of energy use in a building or complex of buildings, and on that basis discovering a more efficient way of operating. The designs of all kinds of things such as aeroplanes can be made more efficient, again saving energy and other costs. If this were systematically applied to every area of the economy and public services, a massive boost to incomes and energy saving could be achieved.
Deep learning’s ability to recognise complex patterns and to predict things where some data is missing also has huge potential for developing humanity’s creativity. A clear and already existing example of this (though it requires a great deal of improvement) is in automated translation. It is already the case that anyone with an internet connection can instantly translate a large body of text reasonably accurately, giving access to the ideas of millions more people. This is because deep learning AI can be trained on vast amounts of data from language comparisons, can identify correlations between words and sentences in different languages, and thus reliably predict what word or sentence in the other language means the same thing. The same principle is making possible near instantaneous audio translations, so that one can wear an earphone, listen to someone speak in a foreign language, and hear a live translation of what is being said.
Microsoft has already developed a device which enables those with eyesight loss to have the world narrated to them by an app. Thus, if you point a camera at an object, it can read its label. Supposedly, it can even tell you which of your friends you are looking at, and what their facial expression is. Without doubt, this technology in its current form is unreliable and cumbersome, but it will surely improve rapidly. The potential to liberate people to perform various tasks on their own is clearly great.
Even the secrets of the ancients are being uncovered by AI. Using technology very similar to predictive text, DeepMind has been able to help archaeologists decipher ancient writing that had parts of the text missing or was for other reasons not understood. [3] So long as it is possible to feed deep learning AI with enough data pertaining to a particular mystery, there is a good chance the mystery can be solved thanks to the power of AI to uncover hidden patterns.
There is no doubt that when it comes to aiding human creativity, the prospects opened up by the likes of ChatGPT and Dall-E are most tantalising. Basing themselves on the vast amount of visual data (in the case of Dall-E and other image producing AI) and written language available on the internet (in the case of ‘chatbots’ like ChatGPT), these AIs can almost instantaneously create new images and text in response to a prompt from the user.
By aggregating all the images labelled, for instance, ‘cat’ on the internet, or all the work of a specific artist, Dall-E spots distinct patterns, such as the way a cat’s hair responds to natural light, or the tendencies of a particular artist. This allows it to ‘creatively’ produce a new image of a cat in a specified situation, such as “a cat painted in the style of Van Gogh”. ChatGPT can, for the same reasons, instantly write a poem in the style of Hamlet, about any subject you like, with astonishing competency.
The potential these technologies have for developing the power of human creativity is remarkable. Image creation AI gives artists and story-boarders the ability to rapidly iterate ideas. The images created tend to be somewhat generic, since they are based on aggregating existing images, but the ability to combine types (‘a cat in a Van Gogh painting’, ‘a football match played in a cyberpunk city’, etc.) into many high quality new images is clearly very helpful for those who need to come up with prototypes or proofs of concept.
Similarly, text producing AI such as ChatGPT can help anyone to rapidly draft coherent text for any need. In fact, it can even help programmers to write code. It can already do this so well that it will become possible for people with no training in coding whatsoever to produce websites and maybe even working software, such as video games. All they would need to do, is to write, in natural language, a prompt for what they want their website/software to do and to look like, and the AI will write the code to produce the desired effect.
It is hard to overstate the revolutionary potential of this technology, when used in the right way for the right purposes.
The fetter of capitalism
Marx explained that a given social system provides a framework for the development of the productive forces. But, at a certain stage, the productive forces outgrow the relations of production in which they must operate, and thus these relations of production become a fetter to further development. The capitalist mode of production fostered an immense development of the productive forces, far beyond the level of feudal society, but has for some time been a fetter. This is why investment and productivity gains are so chronically low, despite the creation of incredible new technologies.
AI, and other digital technologies such as the internet, represent means of production which are too advanced for capitalism to properly utilise. This is because capitalism is production for private profit. If profit cannot be wrung from a potential investment, it will not be made. And profit can only be made by exploiting the labour-power of workers, and then realised by selling the products of this labour on the market.
Technology such as the internet and AI place a question mark over this process, because they employ automation to such a high degree. For instance, the internet enabled the copying and sharing of large amounts of information very quickly, with little to no labour involved. Anyone could share a film or piece of music with untold numbers of people all over the world, with no loss of quality and with no effort. For this reason, the existence of the internet made one of the key parts of the music and film industries – the copying and distribution of recordings – essentially redundant overnight.
This presented an enormous problem for this branch of capitalism: how could they continue to make a profit when anyone could get hold of a copy of an album for free? The capitalists have attempted to solve this problem by simply criminalising the ‘peer-to-peer’ sharing of media online and by setting up a number of streaming services, each with a monopoly over their ‘own’ material, for which viewers/listeners must effectively pay a perpetual rent. This solution has been reasonably effective as far as safeguarding corporate profits is concerned, but from any other point of view it is an irrational fetter on both the distribution and production of creative works, which serves only to hold us back from realising the potential of our own technology.
Similarly, the latest AI technology threatens to reduce the value in the capitalist economy of a vast swathe of professions and industries. If so much of the writing and images used in publications can be produced instantly by an AI, for example, and if authors can churn out ideas for plots so quickly, the value of their work will be greatly reduced. And if the training and skill required for workers to produce such commodities are also reduced to merely typing prompts, the value of their labour-power will also be drastically reduced.
In a socialist society this would not necessarily be a bad thing. The artist, for example, would have no fear of the powers of AI to produce ‘artwork’ at a moment’s notice, since art would not be produced for profit, or as a means of living. Art would lose its fetishistic link to private property, and would be produced for its own sake, or rather, for the sake of society. It would be a genuine expression of the ideas and talents of people, and a way for them to communicate. As such, the generic works of AI would be no threat, instead they would be auxiliary tools for the artist.
Under capitalism however, the artist’s existence is precarious and subordinated to the vagaries of the market. They must jealously protect their exclusive right to the sale of their artwork, otherwise their livelihood risks being destroyed.
Far from liberating humanity, AI under capitalism will only exacerbate its inherent tendency towards monopoly and inequality. The best AI for generating images, text and for solving problems, is and will continue to be developed by enormous monopolies like Google and Microsoft, with the best engineers, best hardware, and biggest databases. They will use their monopolistic position to make monopoly profits of course, and the technology’s advantages, namely in speeding up and cheapening production, will be used by other corporations to lay off some workers, and to drive down the wages of others.
This technology is also already being used to speed up labour, and thereby increase the rate of exploitation, from another angle. Cameras and other sensors can cheaply and effectively monitor the labour process of thousands of workers, disciplining them so they produce more for the same wage.
Amazon is notorious for this, and rightly so: “in 2018, the company had two patents approved for a wristband tracker emitting ultrasonic sound pulses and radio transmissions to monitor a picker’s hands in relation to inventory, providing ‘haptic feedback’ to ‘nudge’ the worker towards the correct object.”[4] As automated surveillance advances and becomes cheaper, it will be rolled out across the economy, increasing the stress and alienation of workers everywhere.
Capitalism lays its hands on a revolutionary technology whose real potential is to harmonise and rationalise production and to enhance the creative powers of humanity, and instead uses it to further discipline the worker, to throw more workers on the scrapheap, to make the artist’s existence even more precarious, and to concentrate more and more power in the hands of gigantic corporations. The effect will therefore not be to bring stability and abundance into the economy, but to heighten the antagonisms and inequality of society.
By further monopolising the economy, driving wages down even more, and concentrating more and more wealth in fewer hands, AI under capitalism will exacerbate the anarchy of the market.
This has already been seen in the present economic crisis. During the pandemic, consumption patterns changed, leading to a big increase in orders from companies such as Amazon. Amazon uses AI heavily in its forecasting model, Supply Chain Optimization Technologies (SCOT). SCOT simply looked at patterns of consumption, without any understanding of what was causing these new patterns. Consequently, it recommended Amazon buy billions of dollars of more warehouse capacity to cope with increased demand.
But as lockdowns inevitably ended, demand for Amazon’s goods slumped. As a result, Amazon now has far too much warehouse space, and too many unsold items, which in turn has led to layoffs and discounts. Rather than eliminating waste and overproduction, the use of AI to enhance the profits of monopolies has actually made the situation worse.
It is no wonder that despite the amazing potential AI offers humanity, many of us live in fear of it. What does this widespread fear of AI reveal? Very little about the technology itself, but a great deal about the strange contradictions of capitalist production. Under capitalism, precisely the highest achievements of human thought, the most wondrous technologies with the potential to eliminate the evils of poverty and ignorance, are the very things that threaten more poverty.
We fear being enslaved by an impersonal, cold and calculating artificial intelligence, but we are already subordinated to the impersonal, blind, and unconscious forces of the market, which is also cold and calculating, but not very intelligent or rational.
A technology made for planning
The use of AI to heighten capitalist exploitation is a tragic, criminal waste. One could hardly imagine a task better suited to AI, than that of planning a complicated economy to meet needs. With modern technology such as sensors, it is already possible to automate logistics, as Amazon has demonstrated.
In their immense complex of vast warehouses, Amazon uses AI and robots to efficiently plan which items need to go where, and in what quantities. There is no reason sensors could not be integrated into the economy as a whole to provide real time data about what is being consumed, and in what proportions, where, and what equipment is in danger of breaking down and therefore needs to be fixed in good time. The German software giant SAP has already developed an AI-powered application called HANA, which is used by companies such as Walmart to plan all their operations harmoniously using real time data.
By feeding deep learning AI with such data, it would be more than capable of devising, alongside elected committees, a long term plan for the economy, which would maximise efficiency to finally meet the needs of humanity, so that no one need go hungry or homeless, or fear for their job. In this way, vast swathes of waste could be eliminated and the working week rapidly shortened. Not only would AI be enormously helpful in drawing up and adapting such a plan, it would have the benefit of helping the people involved in planning to see past whatever biases or limitations may exist in their thinking.
Clearly, this AI would have to be overseen by people – it would only be a tool at their service. It could not answer questions such as what sort of architecture should be developed, how our cities should look, etc. But its insights into the patterns of an economy and how best to economise production would be indispensable.
This is the potential of the latest in AI technology. We have at our fingertips the technology to bring harmony into production, to eliminate the wasteful excesses, greed, irrationality and shortsightedness of the capitalist system. We could use it to give all of humanity not only the things they need to live well, but the power to create artwork, or to redesign and improve their own home, workplace or neighbourhood. It will make the building of a socialist society free from all scarcity and class distinctions faster and more painless.
This power is at our fingertips, but it eludes our grasp, because, contrary to what so many imagine, how it is used is not determined automatically by the technology itself, but by the mode of production we live under.
As long as we live under capitalism, it is capitalism that will determine how AI is developed and used, not the technology’s sheer potential. This is why predictions of AI and automation doing away with the exploitation and anarchy of capitalism are such a false dawn. AI, no matter how advanced, cannot do the job of liberating humanity from capitalism for us. And no matter how irrational it has become, capitalism will be ruthlessly defended by the capitalist class.
The only force that can combat this is the only one that has an interest in doing so, that is to say, the working class. It is the fact that the working class is interested in achieving socialism that makes it possible for it to grasp both the need and the means to do so.
Only when we have finally overthrown capitalism so that we may subject the economy to conscious, rational planning, can AI and other technological advances flourish to their full potential as the most wondrous and general tool of human development yet devised. As Leon Trotsky put it so poetically:
“Technical science liberated man from the tyranny of the old elements – earth, water, fire and air – only to subject him to its own tyranny. Man ceased to be a slave to nature, to become a slave to the machine, and, still worse, a slave to supply and demand. The present world crisis testifies in especially tragic fashion how man, who dives to the bottom of the ocean, who rises up to the stratosphere, who converses on invisible waves from the Antipodes, how this proud and daring ruler of nature remains a slave to the blind forces of his own economy. The historical task of our epoch consists in replacing the uncontrolled play of the market by reasonable planning, in disciplining the forces of production, compelling them to work together in harmony and obediently serve the needs of mankind.” [5]
References
[1] G Marcus, E Davis, Rebooting AI: Building Artificial Intelligence We Can Trust, Pantheon Books, 2019, pg 56
[2] A Katwala, “DeepMind Has Trained an AI to Control Nuclear Fusion”, Wired, 16 February 2022
[3] Y Assael, T Sommerschield, B Shillingford, N de Freitas, “Predicting the past with Ithaca”, Deepmind, 9 March 2022
[4] N Dyer-Witheford, A Mikkola Kjosen, J Steinhoff, Inhuman Power: Artificial Intelligence and the Future of Capitalism, Pluto Press, 2019, pg 93
[5] L Trotsky, “In Defence of October” in The Classics of Marxism, Vol. 2, Wellred Books, 2015, pg 226-227
