Ashley Ray, Socialist Alternative Leicester
ChatGPT is a chatbot based on the latest in a line of language models from the US-based company OpenAI. This is the same organisation that developed DALL-E – a popular AI image generator. The bot’s capabilities can be impressive. It is able to write an essay, hold a conversation and confidently, though often inaccurately, answer questions from its users.
The chatbot has taken university and college campuses by storm, with some students passing their assignments into ChatGPT to get quick, automatically generated answers. These answers are not always correct and often lack depth. Despite this, students at multiple universities and colleges have taken the risk and have been caught passing AI generated work off as their own.
ChatGPT is more than a cheating tool for students. Its enormous dataset, combined with the ease of ‘chatting’ compared to searching academic databases, gives it a lot of potential as a useful tool for research. Essays are not all that ChatGPT can generate. It is capable of writing code, articles and pretty much anything else composed of text.
The headlines about ChatGPT range from doomsaying alarmism to giddy evangelism, with the papers warning us that it will destroy the world while simultaneously promising us it will deliver us to a shiny techno utopia. To understand what AI means for workers we need to start with what it actually is.
How ChatGPT works
Large language models like ChatGPT are based on neural networks – systems that mimic the surface behaviour of neurons using a huge web of simple problem-solving rules. These networks are ‘trained’ using enormous collections of data. This process involves feeding the networks reams of categorised practice data.
For example, if you need a neural network that can recognise a cat, you need to pass in a huge collection of cat pictures, and pictures of things that are not cats. These pictures need to be labelled, so that after each ‘generation’, the network can be rated for accuracy. The network runs through the pictures again and again, becoming more and more accurate each time as it is adjusted using the feedback from the previous run.
The categorisation process is a huge part of the labour in producing something like ChatGPT, and the workers who perform the task reap next to none of the reward. OpenAI contracted some of its data labelling to a Kenyan outsourcing company, where workers were paid less than $2 an hour to trawl through a set of text containing graphic descriptions of sexual exploitation. This categorisation helped to prevent ChatGPT from outputting anything disturbing to its users.
This practice is widespread in the industry. Data categorisation is often undertaken by underpaid gig workers using platforms like Amazon Mechanical Turk. The AI revolution is built on the backs of hyperexploited workers, but the results of their labour collect in the pockets of the super rich.
The limitations of ChatGPT
It is important to note that the neural networks behind ChatGPT are not the same thing as human brains, despite appearing to mimic it. They are statistical models and are not sentient beings. The fear-mongering around the dangers of sentient AI is not based in reality. The overestimation of what these models are capable of is far more dangerous to humanity than the imagined robot takeover. The ChatGPT of reality and the ChatGPT of the media imagination are worlds apart. A combination of sensationalism and uncritical reprinting of corporate press releases has led to an overinflated view of what these neural networks are able to do.
Greg Brockman, the president of OpenAI, has publicly stated that “ChatGPT is not yet ready to be relied on for anything important”. And it is clear why. The system produces incorrect answers to queries, and will even make up citations when pressed on the issue. Understanding weaknesses is crucially important before entrusting anything serious to a new technology.
One critical flaw in many of today’s neural networks is their difficulty to be understood. Hundreds of billions of parameters (generated during training) are required to drive a system as sophisticated as ChatGPT. This complexity, along with the commercial interests of OpenAI in maintaining their ‘secret sauce’, makes ChatGPT a black box. Its internal reasoning is unknowable even to experts in the field.
Computer systems are often considered unbiased deciders of truth. However, every system inherits the biases of its creators and often the type of society we live in – in this case, capitalism. An all too common example is the incredibly basic failure to provide for non-binary users when adding a mandatory, binary gender input to a sign-up form.
Neural networks are trained on data created, categorised and selected by humans. This means the many biases and even prejudices of our extremely unequal society are baked into their models. If they are presented as a black box, it is incredibly difficult to scrutinise and expose the biassed logic they operate with.
For neural networks to work in the interests of all of humanity, and not just deliver cost savings and profit for the bosses, it is critical that their development becomes truly open. The benefits and drawbacks of any new system should be evaluated by those who work with and access the services that will be impacted – rather than by those whose bottom lines may benefit.
AI’s potential
This technology holds incredible promise – unlocking the ability to talk to a computer system in natural language could help to make the collected knowledge of humanity more accessible than ever. As the digital world expands and becomes ever more critical to dayto-day life, any technology that reduces the complexity of engaging with computer systems can play an important role in providing people with quick and easy access to crucial information for their lives.
However, the most critical missing piece in crossing the “digital divide” is not technological – it is financial. 6% of households, and 11% of lower-income households in the UK have no home internet access. Elsewhere around the world, this is of course much higher. As increasing numbers of crucial utilities shift from physical branches and phone lines to online chat and web forms, the demand for universal access to broadband becomes more important than ever. This is why Socialist Alternative would stand for the nationalisation and democratic public ownership of Openreach, and the major broadband companies.
The primary focus of most reporting around ChatGPT is its potential for automating tasks currently undertaken by human workers. Although potential is overstated in most media reports, it is still a very real possibility. Mythmaking around AI and its capabilities serves as a tool to scare the working class into submission. The threat of automation has for decades been used by bosses in response to demands for higher pay and working-class militancy.
An automated system, however, has never fully replaced human labour. Even the most sophisticated tools of industry do not truly replace, but instead serve to ‘amplify’ the labour of humanity. Automated systems require operation, maintenance and supervision by humans. Self-checkouts, as an example, do not replace the work of cashiers. They instead allow a single worker to handle a greater number of transactions (through their work supervising a set of self-checkout machines).
The spoils of this greater efficiency are not received by the workers under capitalism. The task of operating a self checkout machine is formal, rigid and deskilled compared to the work of a cashier. Understaffing at supermarkets leads to workers drawn between multiple machines, each with their own faults and each with an irritated shopper, while the stress and dehumanisation of this task is of course not matched with extra compensation. ChatGPT is likely to be deployed in a similar way, with human workers using AI tools to increase their productivity, or supervising a set of chatbots.
Automation lays bare the contradictions of capitalism. Reducing the amount of human labour required to perform a task should be unconditionally good, but capitalism is completely unable to deliver us the real potential benefits from this technology. Automation under capitalism means more money in the pockets of the bosses who own the systems along with de-skilling and increased precarity for the workers who build, run and work with them.
Language models like ChatGPT will have the largest impact in the white collar world, where trade union membership is comparatively low. This is why the trade union movement must prioritise building strong, fighting unions in all sectors which will interact with this new technology. Workers must have the organised strength to meaningfully influence the adoption of new technology in the workplace. Collective action is the most effective response to threats to job security from the bosses, making organising in these sectors more critical than ever.
The image recognition capabilities of neural networks also make them an excellent tool of mass surveillance. 54 different countries have already deployed facial recognition systems. Multiple UK police forces have run tests of ‘predictive policing’ tools enabled by AI. Predictive policing software claims to be able to predict future crimes based on statistical patterns, despite there being little evidence that it is effective.
The data it runs on is provided by the police force, allowing the deeply ingrained racism of the capitalist system to be laundered through a high-tech system. Predictive policing can be used as a ‘scientific’ justification for duplicating and reinforcing existing bias, over-policing and profiling in Black and minority ethnic communities.
The future under capitalism
In every class society, technological development is twisted to fit the priorities of the ruling class. Under capitalism, profit comes first, so resources are piled into developing technology that is profitable. The capitalist who owns the newest technology has an edge in the market, resulting in a scramble for the development and adoption of neural networks.
OpenAI is a for-profit company, which has received sizable investments from Microsoft for exclusive licensing terms for its products. Google has also thrown its hat in the ring with its own large language model PaLM 2, and IBM, Amazon and a whole host of tech companies are developing their own AI capabilities.
The arms race over AI in industry is taking place at the same time as an arms race in the most literal sense. AI tools have already been used as part of military operations. Palantir has used the war in Ukraine as a staging ground for their AI targeting technology. The company’s CEO Alex Karp boasted at a Military AI event hosted by the Dutch government recently that “if you have an adversary that knows how to install and implement digitalised targeting and AI, you’re at a massive disadvantage.”
While these words are coming from a man who stands to gain personally from the sale of his companies own AI products, this sentiment is being taken seriously by the world’s military powers. The US Defence Department requested $874 million to spend on AI projects in 2022, while estimates put the spending of the Chinese dictatorship in the same ballpark. Rushing to hook lethal weapons up to neural networks has the potential to escalate conflicts. The flawed judgments of these models, when traced onto the battlefield could have disastrous and deadly consequences.
A different future: How a socialist planned economy could harness AI
While the idea of AI as a novel threat to humanity is exciting at the cinema, the reality is more complicated than that. This technology will only pose a threat to our lives and futures for as long as it is under the control of the same old greed of the capitalist class. ChatGPT is an impressive demonstration of what generative AI is capable of, and a glimpse into a set of potential futures that could be enabled by technology like it, if properly harnessed in a democratic, socialist planned economy.
Neural networks seem otherworldly. But even the most sophisticated gadgets of the 21st century are best viewed as capital: the accumulated product of decades of our labour. Every watt of power, every chip in a datacenter and every line of code only exists thanks to the enormous communal efforts of workers around the world.
The capitalist class have made their intentions for AI perfectly clear. In a world where profit comes first, this new technology will be put to use de-skilling and surveilling its workforce and building more effective weapons of war. However, this future is far from inevitable. In a socialist society, the spoils of automation could be shared by the working class people who make it possible.
As it stands these tools continue to enrich the select few. However, in a world where everyone is guaranteed a comfortable living and industry is controlled in a planned, democratic way by the entire working class, systems like ChatGPT could be used to improve all of our lives. The greater productive power unleashed by automation could be used to free up time, shorten the working week and increase rates of pay. This would only be achieved on the basis of making a full break with capitalism, involving placing big tech companies, banks and monopolies into democratic public ownership, with compensation for small shareholders only on the basis of proven need.
In this planned economy, democratically-elected committees of workers in tech and across sectors, along with service users would then be able to inspect the use of AI and similar technologies to ensure they are used for the benefit of society and the planet.
