Through my reporting, I’ve come to understand two things: Artificial intelligence is a technology that takes many forms. It is in fact a multitude of technologies that shape-shift and evolve, not merely based on technical merit but with the ideological drives of the people who create them and the winds of hype and commercialization. While ChatGPT and other so-called large language models or generative AI applications have now taken the limelight, they are but one manifestation of AI, a manifestation that embodies a particular and remarkably narrow view about the way the world is and the way it should be. Nothing about this form of AI coming to the fore or even existing at all was inevitable; it was the culmination of thousands of subjective choices, made by the people who had the power to be in the decision-making room. In the same way, future generations of AI technologies are not predetermined. But the question of governance returns: Who will get to shape them?

Within two days, Bender had sent Gebru an outline. They later came up with a title, adding a cheeky emoji for emphasis: “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?”

Gebru assembled a research team for the paper within Google, including her colead Mitchell. In response to the encouraging words in Dean’s annual review, she flagged the paper as an example of the work she was pursuing. “Definitely not my area of expertise,” Dean said, “but would definitely learn from reading it.”

In total, it presented four key warnings: First, large language models were growing so vast that they were generating an enormous environmental footprint, as found in Strubell’s paper.

Second, the demand for data was growing so vast that companies were scraping whatever they could find on the internet, inadvertently capturing more toxic and abusive language as well as subtler racist and sexist references.

Third, because such vast datasets were difficult to audit and scrutinize, it was extremely challenging to verify what was actually in them, making it harder to eradicate toxicity or more broadly ensure that they reflected evolving social norms and values. Finally, the model outputs were getting so good that people could easily mistake its statistically calculated outputs as language with real meaning and intent. This would make people prone not only to believing the text to be factual information but also to consider the model a competent adviser, a trustworthy confidant, and perhaps even something sentient.

But what seemed to bother Dean the most was how other people had misread Strubell’s research to make Google look significantly worse. The “Stochastic Parrots” paper, Dean argued, risked exacerbating this issue. Because Gebru did have access to Google’s internal numbers and was citing Strubell’s external estimate anyway, it could appear as if Strubell’s calculations were an accurate reflection of the company’s emissions. To Dean, this justified his and other senior executives’ criticisms of Gebru’s paper: If Gebru had wanted to cite Strubell, she should have chosen an estimate that was not Google’s Evolved Transformer; if Gebru had wanted to cite the Evolved Transformer, she should have sought internal Google numbers.

Some researchers found this logic frustratingly inadequate. Google had never made those internal numbers public previously, even in response to Strubell’s original paper; now it was blaming Gebru for its own lack of transparency while also refusing to let her cite publicly available estimates based on legitimate assumptions. Never mind that the company had unceremoniously forced out Gebru before she’d even had a chance to consult internal numbers and revise her paper. The only possible outcome of this catch-22 was censorship of critical accountability research.

Finally, to redistribute power along our third axis, influence, we need broad-based education. The antidote to the mysticism and mirage of AI hype is to teach people about how AI works, about its strengths and shortcomings, about the systems that shape its development, about the worldviews and fallibility of the people and companies developing these technologies. As Joseph Weizenbaum, MIT professor and inventor of the ELIZA chatbot, said in the 1960s, “Once a particular program is unmasked, once its inner workings are explained in language sufficiently plain to induce understanding, its magic crumbles away.” I hope this book is just one offering to help induce understanding. It builds on the work of the many scholars, journalists, activists, and educators before me who have dedicated themselves to public education. May it be a new ground upon which many more after will rise up and build.