Kevin Crane cuts through the hype to assess the nature of AI and its prospects, and the response needed from socialists

AI has been much in the news throughout 2023, with both ostentatious promises and threats being made about it. From the point of view of workers, the predictions are usually very negative, and this was reflected at this year’s TUC with a breathlessly worded motion calling for legislation to be created to ‘build a boat’ to help workers survive the ‘tidal wave’ of AI threatening to take away jobs, ranging from journalism to customer services. The TUC has formed an ‘AI taskforce’ aimed at developing what they are calling a ‘new legal framework’, involving representatives from multiple unions and professional bodies to advocate to government, a strategy that – as is usual with the TUC – basically avoids engaging with ordinary workers in any real way.

The first problem any such task force will instantly have, is the question: what are we even discussing? Volumes have been written about AI recently, much of it very confusing and unhelpful (often purposely designed to confuse). The Wikipedia definition of AI is a bit of a case-in-point:

‘AI applications include advanced web search engines (e.g., Google Search), recommendation systems (used by YouTube, Amazon, and Netflix), understanding human speech (such as Siri and Alexa), self-driving cars (e.g., Waymo), generative or creative tools (ChatGPT and AI art), and competing at the highest level in strategic games (such as chess and Go).’

That is an absolute jumble of things that have very little to do with each other. Indeed, one could go as far as to say that the only thing that several of those technologies have in common is that they are examples of software that get marketed as AI. This is because AI is, fundamentally, a marketing term. There is no specific piece of technology that is the essential AI.

For the purpose of the topic of jobs and employment, it’s probably best to focus on one area: which is what Wikipedia is calling generative or creative tools. While some of the other technologies listed might be linked to fears of job losses, it is generative software that is unexpectedly automating processes previously thought to be immune. They are the most novel out of the slightly random collection of ‘AIs’ out there and they have a particularly huge hype storm surrounding them, which includes some major controversies. There is also a major workers’ strike happening right now in which they play a key role, something at which it might be nice for the TUC to take a look.

The confusing promises of generative software

ChatGPT kind of exploded into popular consciousness this year, which may have been a triumph for engineering, but was definitely a triumph for marketing. It is very cleverly designed as a promotional tool, right down to the way its interface works. You enter your text, and it does a whole theatrical demonstration of processing what you’ve said, apparently typing it out in real time. This is, of course, not necessary at all. Much like the fake counting sound a cash machine makes, or the silly shutter noise on your smartphone when you take a picture, it enhances an illusion of your interaction with the software.

If you haven’t used ChatGPT, a basic upshot of what it does is that you submit to it a query for a piece of writing, and it generates something that more-or-less meets the description of what you asked for. People often try to stretch it by making absurd requests, like motor-mechanics procedures in the style of a Shakespeare sonnet, but these actually don’t trouble the platform much more than prosaic requests like ‘tell a children’s story about an elephant’. The reason for this is that what ChatGPT does for either process is utterly unlike what a human being would do. A person would have to do research or use their imagination. The machine simply crunches numbers.

ChatGPT is technically called a ‘large language model’, if you’re being polite, or a glorified spellchecker if you’re not. It has a vast repository of statistics attached to words, and the statistics provide information about the associations between words. So, when you enter your request into ChatGPT, what it’s doing is using the statistical information from the words you entered and then performing a series of mathematical calculations. These are used to produce a set of text that has a decent chance of being suitable output, which is judged on the correlation of associational data from words in its database. Any visual generative system you’ve seen works in a similar way; it just combines retained images on a similar statistical basis. There are also music generators which do this with sounds, and I have to say I have a particular personal distaste for that as a concept.

It is crucial to emphasise one thing about this stuff: the machine never at any time has the faintest grasp of the meaning of the words, or images, or alleged music, that they are passing in and out. All these assets are all just cyphers for statistical information, and that statistical information is all derived from material that is found out there on the internet.

A fantasy of value without workers

As already alluded to, one of the highest profile industrial disputes in the world today is, at least in part, about ‘generative AI’. Hollywood is currently paralysed by strikes by both the actors’ and writers’ unions. There’s quite a fun meme of this on social media which has two photo portraits: the one on the left is Sarah Connor from Terminator 2 and it says ‘the woman we thought would save us from the machines in the 90s’, and on the right it has Fran Fine from the sitcom The Nanny captioned as ‘the woman from the 90s who’s actually saving us from the machines’, because the actress Fran Drescher, who played Fran Fine, now happens to be the leader of the actors’ union. Crucial to the whole dispute is that the actors are concerned about their images and voices being used over and over again by greedy, lazy studios – to produce what we depressingly now call ‘content’ – without paying further wages. Similarly, the writers are concerned that increasing quantities of scripting will be churned out of LLMs without them being paid.

It is a bleak vision of the future, to be sure, to think that films and TV might one day just be run off digital sausage machines, with little human involvement other than keeping the machine running. The studios certainly appeal to this prophecy of doom to demoralise the workers. From a socialist point of view, however, it is crucial to argue that what these studio bosses are indulging in is in no way novel. They are in fact doing something capitalists have done, over and over again, for about 300 years. They are trying to live in a fantasy world of value without workers.

Attempting to use generative systems to produce content is much more problematic than enthusiasts for the tech would have you believe. LLMs, in particular, are becoming somewhat notorious for returning output that looks superficially plausible, but that contains serious defects that the machine cannot eliminate because it does not comprehend its own produce. If one were to take the earlier example – the Shakesperean Haynes motor manual – ChatGPT would certainly return you something that looked like a Shakesperean sonnet, but God help you if you tried fixing a car following the directions.

They would almost certainly be completely wrong, just a set of steps that have the appearance of instructions, but that are not based on anything technical or verifiable. Because the LLM can’t appreciate meaning, or veracity, or consistency, it has no way to know if what it has said makes sense. People have been coming a cropper in the last few months trying to use ChatGPT to write their school essays or, even worse, fill out legal forms. You will probably get sentences in standard English, but they probably won’t be correct information.

There are some things that an LLM cannot do at all. The most basic example is jokes: they cannot tell them. A joke is fundamentally a play on anxieties and misapprehensions: setup, mislead, payoff. There are no statistics that help you calculate that. If you want new original jokes, you are going to have to pay someone to write them.

Capitalists are also treading on each other’s toes in fundamental ways with generative systems, because they have faceplanted right into their own frameworks of copyright. Big battles have begun in the American courts over whether or not generative content can be copyrighted, and right now, the tech bros are very much losing. Just last month, a US judge ruled that a generative image was not copyrightable, because no human being created it.

The striking US screenwriters also managed to make a very good point about where the content ultimately derives from. If you ask ChatGPT, ‘should Hollywood agree to the union’s demands?’, it always says yes … because it derives the answer from text that the union’s membership writes, so of course that’s what it returns!

The long-term implications of generative systems reading assets off the internet and then putting output based on those assets back onto the internet should also be obvious. The only thing these systems can do is churn existing information, and it won’t be long before they are to a large extent churning their own output.

Karl Marx would not find any of this stuff surprising: he had identified that labour creates value, and as much as capitalists would like for that not to be true, it remains the case. Generative content is what Marxists call ‘dead labour’, which is an accumulation of pre-existing commodities that, ultimately, undermine profitability and precipitate crisis for capitalists.

The long dark twilight of tech

We are probably only in the early stages of this particular controversy over AI. The Hollywood strike has been a key development, because it has changed headlines from ones that tech entrepreneurs want to a counternarrative about for what generative tools can and should be used.

What’s been much less discussed, because there has not been a fight about it, is that newsrooms are being absolutely gutted at the moment, with declining numbers of journalists being expected to use such tools to cover for their ever-increasing workload. This is also not actually new; the great journalist Nick Davies had identified twenty years ago that over-stretched journalists were doing what he called ‘churnalism’. That is to say, churning over other copy, much of its public-relations propaganda rather than reporting. This unfortunate process has simply been accelerated by technology.

Workers can and should struggle, because sensible analysis of this situation doesn’t favour the tech industry propaganda that the technology is unstoppable. As with so much else about information technology, the displacement of jobs by generative software merely displaces necessary productive labour from one location to another. Even where ‘generative AI’ makes something useable, workers are needed to process its output. Capitalists will always dream of being free of workers, but they never are.

How should the left respond to the rise of this technology? The maybe unglamorous answer (not that Hollywood picket lines don’t tend to be glamorous) is that we simply need to extend trade-union organisation to the white-collar workers that have to manage and process the input and output of these machines. This is really always the answer to any technological change under capitalism.

There are other big issues to talk about with AI that should be the scope of further discussion: the craziness about supposed AI apocalypses, the racism and discrimination in AI surveillance, and the inanity of the ailing tech companies that pretend AI is going to save them from oblivion. Essentially, whenever we are discussing tech at this time, we have to recognise that we are in many ways at the end, not the beginning, of an era in technological development, and the media hype that we are constantly being battered by is a demonstration of that.

Immediately before AI became the big story, we were assured by the media – and even government – that the next big thing was ‘Web 3.0’. This was meant to be an utterly terrible confluence of financialised technologies like cryptocurrency and non-fungible tokens with that oh-so-boring humbug, virtual reality. Billions were spent on this nonsense by the tech sector, to produce quite literally nothing.

The present tech sector is in its own Ragnarök: a long dark twilight, in which the apparent magic of its ability to make money has been utterly exposed by the truth of an economy that was just funnelling cheap credit, energy and resources to gormless ‘entrepreneurs’. AI is simply another straw they are clutching at to delay the inevitable, and terrible, return of REALITY.