'Empire of AI' - artificial intelligence as an imperialist project

📚 Finished reading Empire of AI by Karen Hao.

This is a lengthy, very insightful book about the modern AI scene, by which the author means the big generative LLMs such as ChatGPT, Claude and Gemini. It is rather negative in tenor. Which is no bad thing given the hype the companies involved get in other quarters.

It’s important that we, as potential users, understand the companies and technologies we give our data and perhaps money to. And if you haven’t been following the more critical press already, this might truly open your eyes.

Firstly, why “Empire of AI”? Why is the modern practice of “AI” imperialistic in nature?

Whilst not claiming that these companies are instantiations of the horrific and shameful British and other empires seen in the past, the author draws many parallels:

1. The seizing of resources: Traditional empires stole land, minerals, crops and other resources from the countries they ransacked. LLM companies thieve the data and creative work of humanity without consent, as well as in most cases the personal and behavioural data of their users.
2. The exploiting of human labour: Traditional empires took humans as slaves; empires were often predicated on these worst of abuses. AI companies of course don’t quite go that far, but a surprising amount of it relies on very poorly paid annotators and moderators from countries poverty-stricken enough that people are forced to take any job going to survive. Sure, US developers of this stuff often earn astronomical salaries. But the people who have to go through the sometimes psychologically damaging inputs and outputs of the system are secreted away and paid exploitative poverty wages by contracted employers who have no liking for any kind of employee rights. Meanwhile the vast wealth generated by these companies is situated in their home base - invariably rich, very “developed” countries.
3. Physical extraction: Famously, data centres are very unpopular and damaging to the environment - generating pollution from the vast amount of energy they need, health-affecting noise pollution and damage to essential water sources. And guess what, they’re often placed in cheap places - which means poorer communities in regions and countries where the local residents have no power or means to stop their development, despite the many movements who try to fight back. The systems also require the environmentally damaging extraction of rare earth minerals and the like to work. The end result is a further wealth and health transfer from the poor to the unfathomably rich.
4. The claim of a civilising mission: Old empires spread the moral platitudes and religion preferred by the conquerors - supposedly on a mission to save local “savages”. They pretended to help out with infrastructure, technological development and so on; of course what came in that domain was mostly to faciliate the wealth transfer back to the invaders' land. New AI companies claim that their products will help humanity as a whole, that they’re developing AGI that will make everyone’s life utopian - whereas to the extent that these models have changed the world, one could hardly call them utopian. Nonetheless, the media often buys into the as-yet-entirely-without-evidence vision of the bright new future these companies claim they’ll bring, despite the higher-ups in the AI world often appearing to have quite a dark view of humanity.
5. The monopolisation of knowledge: These companies pay high wages to top researchers and scientists, effectively shaping what science is done, what discoveries can be made. The resulting knowledge is treated as proprietary information that can only benefit the company paying, not humanity as a whole.
6. Justifying behaviour by reference to their rivals: Old empires used to warn that if they - the virtuous conquerors, the ethical exploiters - didn’t conquer and exploit some other countries then some worse empire would come along. Famously any argument about slowing down the development of potentially very dangerous AI in America is swept away with the claim that if they don’t do it, the menacing state-driven companies in China will do far worse.
7. Economic power as political power: Olden-days imperialist profiteering saw the East India Company. Modern day big AI leverages its wealth and secretive knowledge to exercise power over the political domains of societies. It writes its own legislation. It effectively bribes our political masters into supporting it. It induces state reliance in its technology, giving it the leverage to drain public coffers whilst getting into the position of being able to claim it’s too big to fail.

This somewhat philosophical take on modern generative LLMs and the pitfalls of their models and human creators is largely told via the story of the founding and development of OpenAI, the creators of ChatGPT. This is a company rather driven by a surprising amount of incompetence and psychodrama, as revealed by the author interviewing hundreds of employees and adjacent folk.

The focus is on the CEO, Sam Altman. In the author’s view he’s a charismatic guy, able to convince a lot of people a lot of stuff without any kind of evidence of follow-through. He’ll say what people want to hear, and then when someone with a different opinion pops up, he’ll tell them exactly the opposite.

This is a particular problem in a company such as this which has employees with extremely different views of what they’re developing and how it should be done; the doomers who demand a focus on control and safety to the extent that some feel if done wrongly it will cause the end of humanity, and the boomers who want to develop (and sell) more powerful models faster, without regulation, as soon as possible. This dynamic is bound to cause all sorts of crazy psychodrama if supervised by a CEO who seems more willing to tell people what they want to hear than tell the truth.

He has incredible amounts of personal ambition and will stop at very little to enable it. He’ll mislead, lie and manipulate those around him to get what he wants.

You may remember the business drama where the board of OpenAI found him so duplicitous that they sacked him. He was back running OpenAI in just a few days. This book gives the most detailed telling of that debacle I’ve seen.

OpenAI (and the other big US players) have also become obsessed with scale. They build ever bigger models requiring more data and an astonishing amount of computation. Size is everything. More computers, more stolen data is seemingly their sole intent for future development. Not for nothing was Sam Altman asking for trillions of dollars of investment. By acquiring many of the top AI researchers through their sheer mass of wealth, they shape what avenues are explored, meaning that a lot of potentially more ethical, more sustainable potential methods around the development and use of this technology are left dormant.

This is a shame because the author doesn’t think AI is inherently bad, it’s not destructive by necessity. She cites an example in New Zealand of the development of a smaller, more targeted model by the Māori population that she finds to be an example of how this technology could be built in a more ethical, more empowering community-driven way - a way that does actually benefit a part of humanity who need help, in this case the Māori fearing the demise of their language - and set up so that it cannot be taken over by a megalomaniac in future. Concerns like consent, data sovereignty, transparency and low impact on the environment were paramount. This is of course unlike the big AI players such as OpenAI today. Despite its seemingly noble original mission, there is no longer anything open about OpenAI of course, beyond its CEO’s wanton ambition.