There are two paradigms of debate around the ethics of AI development. In Silicon Valley, many AI developers are primarily concerned with the potential of an AGI system becoming self-aware and causing human extinction. In Empire of AI, Karen Hao explores smaller-scale, present-day ethical concerns with AI development, including labor and resource exploitation. She argues that big tech firms should shift their focus from “the theoretical rogue AI harms of Doomerism [to] the existing real-world harms, from discrimination to misinformation to job automation” (419).

Tech leaders, including OpenAI cofounder’s Ilya Sutskever and Elon Musk, are major proponents of the belief that an AGI could lead to either a utopic or an apocalyptic future. Hao reports on these beliefs, noting that, for instance, Musk “described AI the ‘biggest existential threat’ to humanity and its development as ‘summoning the demon’” (24). Sutskever encourages these beliefs about the need for AI “safety” and “alignment” within OpenAI. Hao reports on a dramatic ceremony in which Sutskever burned an effigy that “represented a good, aligned AGI that OpenAI had built, only to discover it was actually lying and deceitful. OpenAI’s duty, he said, was to destroy it” (255). The AI Safety team within OpenAI was primarily concerned with this set of ethical issues around AI development.

Hao argues that this quasi-religious set of eschatological concerns distracts from the real-world ethical harms she documents in Empire of AI. These include, but are not limited to, the use of AI to create violent or sexually explicit materials including CSAM, labor abuses of data workers in the Global South, ecological disasters due to the immense amount of water and mineral resources required for database development, and the illegal or unethical data scrapes of materials created by human artists and writers for use in generative AI systems.

Hao uses the example of Timnit Gebru’s firing from Google’s AI Ethics team to illustrate the difficulty of getting big tech firms to take these real-world ethical concerns seriously. Gebru coauthored a paper entitled “On the Dangers of Stochastic Parrots” identifying key limitations and dangers of large language models. The term “stochastic parrots” implies that LLMs operate through mere guesswork (the meaning of the ancient Greek word stokhastikos), mimicking human thought without understanding the underlying concepts, like parrots. The paper argues that because of this lack of understanding, LLMs have the potential to reproduce and amplify dangerous cultural biases. When Google found out about Gebru’s paper, they pressured her to retract it. When she refused, they fired her. Hao criticizes Google’s decision as a sign that they were not capable of “developing a capacity for self-reflection” (168).

Hao suggests that there are more ethical ways to build AI models that are decentralized and allow people to control how their data is used. She cites the Te Hiku project to revitalize the te reo Māori language through Indigenous control as one such for example. She sees this “small, specialized model that excels at one thing” (412) as a way to address the many ethical concerns about AI development described in Empire of AI.