AI Ethics: Feeding the Machine

As is often the case, this blog post calls your attention to a new book we think is worth a peek—Feeding the Machine: The Hidden Human Labour Powering AI (2024) by James Muldoon, Mark Graham, and Callum Cant (whom we will collectively refer to as “MGC”).

As you can tell from the spelling of “labour” (and we will revert to American spellings hereafter unless quoting the authors), this book was published in the UK. The authors are two professors at Oxford and one at the University of Essex who have studied the impact of artificial intelligence (AI) for more than a decade. They are deeply knowledgeable about their topic and, it appears, passionate about it as well.

MGC’s key point is that the grand potential of AI to improve the world (a subject about which they are at least mildly skeptical) has a dark side that should not be overlooked. That promising new technologies often cause their own problems is not new news, of course. In previous blog posts, we have pointed out that the environment-saving promise of electric vehicles (EVs) has its own issues. (See our blog post “AI and the Energy Issue”). Philosopher Travis Rieder recently noted that he had bought an EV for moral/environmental reasons, but added:

There are real problems, though. The batteries that power electric cars require rare-earth elements like lithium and cobalt that must be mined, which is obviously not carbon neutral. And the mining itself can also pose a moral challenge, as rich deposits may be in countries where lax regulations allow for terrible working conditions. The largest cobalt mines, for instance, are found in the Democratic Republic of Congo, where companies have been accused of exploiting so-called artisanal miners, essentially utilizing a form of modern-day slavery (Rieder, p. 267)

Although others have also written about the harms of AI, MGC’s level of detail is impressive. Their approach is generally to use each chapter to put a human face on a different type of employee in the AI production chain and then to add substantial additional factual detail about what they term the “extraction machine»—the system that produces AI but at a heavy cost to many workers.

Among the workers featured are:

· Mercy (a content moderator in Kenya) and Anita (a content annotator in Uganda), both working for a BPO (business process outsourcing company) to which big AI firms are outsourcing work

· Li, a machine learning engineer working for a tech company in London

· Einar, a technician installing a huge data center in Iceland

· Laura, an Irish voice artist whose intellectual property has been used to train new AI models without her consent

· Alex, an operator working in a huge Amazon warehouse in Coventry, England

· Tyler, an investor with a venture capital fund in Silicon Valley

· Paul, a labor organizer in Africa trying to improve working conditions for content moderators and annotators

The big picture message of the book is that excepting Tyler, the venture fund investor, all the other workers are exploited by the extraction machine. These workers’ output is crucial to the development of AI, but they have little control over (and sometimes little understanding about) their particular contributions to AI development. The leading AI producers are huge companies worth billions—Alphabet, Amazon, Anthropic, Microsoft, Meta, Open AI, Tesla, etc. MGC note: “Big AI benefits from what we call ‘infrastructural power’: ownership of AI infrastructure—the computational power and storage needed to train large foundation models. This occurs through their control of large data centers, undersea fibre-optic cables, and AI chips used to train their models.” (p. 12)

For example, Meta needs content moderators like Mercy to review Facebook postings and take down those that violate Facebook’s policies and might cause it to be penalized. Meta contracts with BPOs around the world that, in turn, hire folks like Mercy. The BPOs compete for this business by charging low prices to Meta and its competitors and, in turn, requiring Mercy to spend 8-10 hours a day watching 500-1,000 videos each day that might include suicides, torture, beheadings, rapes, and the like. One day Mercy saw a video of a fatal car accident in which, she realized with horror, her own grandfather was the victim. Mercy has virtually no job security, enjoys few benefits and little mental health counseling, and is paid a pittance.

Anita is a content annotator, identifying objects in endless numbers of pictures so that large language models of AI can learn by processing them to enable a company’s driverless cars to avoid running into things. Her BPO treats her much as Mercy’s treats her. No job security. Long hours. Few breaks. Little support. Aggressive production targets. Pay of slightly over a dollar an hour.

The big AI firms, and, in turn, the BPOs, fight unionization, require workers to sign non-disclosure agreements, and undertake a variety of other efforts to limit worker power to improve their pay, working conditions, and job security.

This book makes many interesting points. In the chapter about Laura the voice artist, for example, it discusses in detail how AI companies use others’ intellectual property without paying for it in order to train their AI models. In the chapter on Einar the technician who is setting up data centers, we learn that it is predicted that by 2028, 27% of all the electrical power in Ireland will be consumed by data centers and that data centers around the world (MGC give examples in Oregon, Uruguay and elsewhere) are using massive amounts of water to cool these data centers, often where there is already a water shortage.

But the primary focus of Feeding the Machine is on abuse of workers and how those workers might gain power to improve their position. As MGC summarize:

If this book could be distilled down to a single message it is that we human beings, are the often-hidden force that powers AI – both physically with our labour but also intellectually through AI ingesting and synthesising our collective intelligence. … Without us, AI ceases to function. It is only through the constant supply of human labour – annotating datasets, coding software, repairing servers, creating new paintings and literature and keeping supply chains functioning – that AI continues to exist. The extraction machine uses human beings like raw material, churning through ever-larger datasets and quantities of knowledge to power its algorithms. As we have shown the machine makes use of workers in different ways depending on their positions in global capitalism. But they are united in all being directed by the logic of extraction. The machine has a purpose to enrich tech company shareholders and concentrate power in the hands of a narrow elite. (p. 220)

We suspect many readers will find this book to be a bit of an overwrought attack on the excesses of capitalism. But it pays to study all the evidence—good, bad, and indifferent—about the future that artificial intelligence is bringing us and to evaluate all the sides in the debate.

Madhumita Murgia’s book Code Dependent: How AI Is Changing Our Lives (2024) provides substantial tells much the same tale about Big AI’s exploitation of workers around the world.