In 2025, the research firm Gartner predicted that worldwide AI spending would reach US$1.2 trillion. By September of that year, America’s largest companies – Meta, Microsoft, Amazon and Alphabet – had already spent more on AI than the US Government had on education, jobs and social services combined. The message of these tech giants is one of positivity and progress: AI will change the way we work, the way we communicate, the way we live. But for award-winning journalist Karen Hao, the gap between PR spin and reality is as stratospheric as those spending stats. In her 2025 book, Empire of AI, she draws on her intrepid reporting from around the world to tell the real story of what’s going on at OpenAI – the company responsible for ChatGPT and bankrolled by Microsoft – and in the sector more broadly. In this conversation with Griffith Review Editor Carody Culver, which has been lightly edited and condensed for clarity, she reveals the extractive, growth-at-any-cost mindset shaping this transformative technology. 

CARODY CULVER: For many of us, it feels like generative AI came out of nowhere when ChatGPT emerged in late 2022. But as you explain, decades of research and internecine theoretical debates precipitated that moment, and a big part of how we understand AI and its threats and opportunities is down to the term artificial intelligence. How did the discipline come to be given this name, and how has this shaped the way scientists and the general public understand what AI is? 

KAREN HAO: When it was called ‘artificial intelligence’, it was the decision of a single person – an assistant professor at Dartmouth University called John McCarthy – and it was against the advice of his mentor, who thought it would be a very, very confusing name, in part because there’s no scientific consensus around what human intelligence is… But this did really reframe the way researchers were thinking about what they were doing. That was the reason, in part, McCarthy wanted to use the term: previously he was calling it ‘automata studies’, but…he wanted people to be more ambitious and really think about how to create human intelligence in computers. It totally changed the way scientists think about it, and it totally changed the way…people who are investing in this technology think about it. 

And then, of course, the public naturally conflates AI with this idea that somehow scientists are recreating another species, or at the very least recreating something akin to what they see in science fiction portrayals in Hollywood: these Frankenstein-like human variations that go rogue. And all the debates around ‘Is this technology existential, is it going to be smarter than us, is it going to completely do away with our need for brains?’ are tied to the conception that this is somehow akin to us and to our intelligence. 

CC: One of the things that really struck me in the early chapters of the book is how the development of any technology, AI being a prime example, is often presented by those researching it as inevitable. It becomes entwined with these vague notions of human progress and benefiting humanity. But then you cite the work of two MIT researchers and Nobel Laureates, Daron Acemoglu and Simon Johnson, who argue that no technology is inevitable. Why is that the case, and why are we so fixated on the idea that it is inevitable? 

KH: They’ve been doing research for decades around historical technology revolutions and what actually leads technologies to be created in the first place, and there are a couple of conditions that need to happen: there needs to be a rallying ambition around a particular idea, and there needs to be enough capital and political will to make it happen. So, the technologies that end up being created are almost always the conception of the rich and powerful, because they’re the ones who have resources to put behind their ideas. 

Just from that, you can see that technologies are not inevitable, because there are plenty of other ideas in the world that are just not given the resources or the political backing to manifest… And because it’s usually the rich and powerful who end up shaping the conception of technology, [that] technology…often is really self-serving to those rich and powerful people…sometimes intentionally, sometimes unintentionally. 

CC: You make it clear that there are all these shifting forces behind the scenes, these powerful elites, and as the field of artificial intelligence developed, two camps emerged: the symbolists and the connectionists. Each of these two camps – I was so fascinated by this – had a different idea of what ‘intelligence’ means, which led to two different ideas about how to advance AI research. Could you talk about those two different ideas and why the connectionists ended up being the team that won the race? 

KH: The symbolists believe we’re smart because we have knowledge – so if you want to recreate human intelligence in computers, you should create machines that are encoded with databases of knowledge. The connectionists think we’re smart because we can learn…so they thought we should build machines that are analysing data and learning from that data. In the early era of the AI discipline, the symbolists were winning in part because [their ideas] seemed to align better with people’s understanding of how intelligence emerges. 

But in 2012, there was a key breakthrough in the connectionist realm called ‘deep learning’. Three researchers showed that by using software called ‘neural networks’ – deep neural networks that have multiple layers for analysing data – they could essentially effectively recognise images at a higher accuracy than had ever been previously done. This became very commercially interesting to the tech industry, so Google acquired the company that the researchers formed…and started pumping a lot of capital into the connectionist approach. 

As more and more industry money started going into the connectionist approach, they started reaping more and more benefits for their bottom line by imbuing deep learning into not just image recognition but self-driving cars and automated translation. Google ultimately enhanced its ability to target users with ads through search, and then yet more money went into deep learning. Connectionism became unbelievably well-resourced compared with symbolism, which just died on the vine – all grad students going into AI stopped studying symbolism and started studying connectionism and, more narrowly, studying the specific types of neural network architectures that were useful for commercial applications. So, over time, the industry has completely distorted the landscape of ideas within the AI research field to align with a commercial agenda. 

CC: That leads us to OpenAI, the company at the centre of your book. Years before Elon Musk donned his red MAGA cap, he was very worried about what he saw as the existential threat posed by AI. When the AI research lab DeepMind was acquired by Google in 2014, Musk was convinced this was going to be the beginning of the end because Google would put profit above everything else. Then he met Sam Altman, who at the time was president of a startup accelerator company and a rising Silicon Valley star. How did that lead to the creation of OpenAI, and what made OpenAI initially seem like it was going to be a different kind of tech company? 

KH: Musk was very concerned that Google was developing this monopoly on AI researchers after…the acquisition of DeepMind… Altman was just a very strategic, politician-like character, and as the head of Y Combinator, he was trying to spray his investments in a lot of different domains, [such as] nuclear fusion and different types of biotech – he identified AI as another trend that he wanted to get his hands on. He started to cultivate a relationship with Musk and echoed a lot of his concerns, saying, ‘The best way to counter Google is to just build a lab of our own that stands for something fundamentally different.’ 

And so OpenAI was created as a non-profit. It was meant to be transparent, collaborative, democratic, open source. But the foundations from the very beginning were rotten because it was also a project of ego: both Musk and Altman conceived of the lab in part because they were like, we’re the good guys, we want to be the ones to save humanity from evil Google. And that element of corruption then festered and led to a complete one-eighty on their mission. 

CC: The Sam Altman playbook seems to be ‘I’ll just tell people whatever they want to hear and then go off and do my own thing’, which, as you demonstrate in the book, led to complete chaos behind the scenes at OpenAI. In those early days, it had all kinds of projects on the go but no clear direction. Then, in 2019, these rapid shifts thrust the company into the public eye. What happened during that time, and how did it begin to change the way OpenAI was perceived by the industry and the public? 

KH: In the AI field, Musk and Altman were viewed with scepticism because they don’t actually have AI research backgrounds… On top of that, OpenAI approached AI development with a very particular thesis: take existing techniques in the field and scale. That also gave it a bad reputation because in the research world, taking existing things and throwing spaghetti at the wall to see what happens is not exactly considered innovative. And in those early years, there wasn’t really anything to show for this approach. 

But then a couple things happened: they created GPT-2, which was two generations before ChatGPT and was the first inkling that there might be something interesting about large language models. And they created a for-profit to sit within the non-profit. Altman officially became CEO of that for-profit, and then they had a $1 billion investment from Microsoft. People start paying a little bit more attention, [realising that] large language models seemed like an interesting idea…and that money could potentially be raised by pursuing this approach because it was catching the interest of big, deep-pocketed companies. 

CC: You’d been reporting on the field of AI at MIT Technology Review since 2018, and then in 2019, shortly after all this happened, you embedded with OpenAI for three days to write a company profile. What sparked your interest in OpenAI, and what did your experience reveal to you about the inner workings of the company? 

KH: I noticed that the research system was starting to orient slightly around open AI’s approach of scaling and around large language models specifically. I thought the company was beginning to have influence over what kind of AI was developed as well as potentially what kind of AI would be commercialised. Very, very early on, they were already building bridges with policymakers, so I also thought they would have pretty substantial influence on whether policymakers understood AI development and [whether] the public would understand AI development. 

How I pitched [my profile] to OpenAI was, well, you’ve undergone a lot of changes, so maybe it’s time for you to reintroduce yourself to the public? They really liked the idea initially. What I found was…this is a project of altruism, and this is a project of ego. I was talking with these researchers and [realising that] they’re not interested in this just because it could unlock transformative benefits for people – they’re also interested in being the ones to do it, which then started making me ask more questions about what was driving their decisions. Then I realised they were highly secretive, highly competitive, still saying they were a non-profit even though it was clear by now that they had some kind of commercialisation plan on the horizon. Part of the reason why they maintained lip service to the non-profit was because it continued to accrue a lot of goodwill in the public [eye]. That was a fundamental disconnect, and I felt like it should be pointed out. That’s what I ended up writing about, and they did not speak to me for three years. 

CC: The truth hurts. And Sam Altman feels straight from central casting: he’s this kind of Messianic tech leader, he’s got seemingly unwavering self-belief, he can charm and persuade all the right people at the right time. One of your two epigraphs in the book is from his blog in 2013, where he wrote that the most successful founders don’t set out to create companies but are on a mission to create something closer to a religion. Your analogy in the book for the ways these tech companies operate is empire. Could you explain why this analogy best captures what’s going on? 

KH: First of all, these companies have amassed an extraordinary amount of political and economic power to the point where they’re pretty much more powerful than other nations in the world, except for maybe the US Government. You could also argue that they’re becoming more powerful than the US Government, because [it] has absolutely no interest in being a counter to them and their ambitions. From that perspective, we need to stop thinking of them as just businesses. They are political actors, geopolitical actors, and their actions have consequences for people all around the world in the way that no government would. 

But the way they’ve amassed that power is where I see so many parallels with the way empires amass power. They extract an extraordinary amount of resources, most of which are not their own, and they rewrite the rules to suggest [these resources] were always theirs. They exploit an extraordinary amount of labour, not just in the production of AI… AI itself is being designed right now as a labour-automating technology, so it erodes workers’ rights on the way in and on the way out. They monopolise this knowledge production by snapping up all these AI researchers and distorting the scientific field. They are essentially re-forming that field in their image. Whatever science comes out is good for them and continues to perpetuate their ability to keep doing whatever they’re doing, and [the] kind of research that would undermine their efforts is censored. Last but not least, they engage in this existential arms-race narrative where they have to be an empire, but a good empire, because there are evil empires in the world, and they, as the good empire, are ultimately on a civilising mission to bring progress and modernity to humanity. 

CC: It’s interesting too because Sam Altman is one of the tech elites who’s talked recently about universal basic income – ‘AI took all the jobs away, but it’s okay because UBI [universal basic income] can help everyone survive.’ I wonder what your take on that is. If these people are so disconnected from the reality of normal life that they come out with these grand announcements when what they’re actually doing is destroying the social contract… 

KH: UBI to me is just such a weird – it’s literally just the welfare state, except that in Silicon Valley’s conception of it, they don’t pay taxes to the government to enable a more democratic version of UBI where it’s actually accountable to the electorate; [instead,] they’re the ones that distribute the money. It’s just privatising something that should be public. It’s like when Elon Musk created The Boring Company to reinvent transportation, and he’s talking about subways. And you’re like, no, we already have that – it’s a public good and you’re just making it worse. 

CC: This notion of zealous belief is such a strong throughline in your book – not just belief in charismatic leaders like Altman but belief in the possibilities of technology. Key to your narrative is this long-running disagreement between two different camps of believers in the tech space: the doomers, who are concerned with AI’s existential risks, and the boomers, who believe that facilitating technological progress is a moral imperative. Boomers seem to be winning, but their mission is incredibly expensive, and there’s a lot of talk at the moment about whether we’re in an AI bubble – NVIDIA just got a market capitalisation of US$4.4 trillion. What are your thoughts on this? Is the bubble going to burst? 

KH: I think it’s about to implode – in part because the premise of the business succeeding is people finding that much value in AI that they’d pay that much money for these technologies, and then the companies will turn a profit, or those companies monetising people’s intimate data at enough of a margin that they become profitable. We’re not seeing either of those things happening. Generally speaking, people are not adopting AI at the rate companies need, [nor are they] willing to pay for AI at the rate companies need. And, technically speaking, the technology is pretty limited in terms of which types of tasks it can actually help people with. One of the problems is that it’s not accurate… [It’s] being pushed into law, into finance, into healthcare, into all these areas where accuracy is really important. But thus far, the adoption is being sustained by people’s lack of understanding of the laws of the technology, not by actual evidence that it’ll be beneficial. 

Once the evidence plays out that it’s less worth dealing with the errors of the technology than it is to just constrain where the technology is adopted, the adoption bubble will deflate. It could be really horrible for the global economy – because the money that’s invested into this endeavour is coming from people’s retirement funds and university endowments… Public market investors I’ve spoken to have pointed out that they think the bubble will be even worse than the real-estate bubble because with the real-estate bubble, it was one industry that went boom, but with the AI bubble, eight out of the ten richest companies in the world all have their valuation currently over-indexed on AI. We’ve never had that kind of consolidation in the market on a single thing. 

CC: It’s terrifying. You talk a lot in the book about OpenAI’s obsession with scale, which I guess is one of the things that got us to this point. When they train their large language models, they’re so fixated on scaling up that they just keep pumping in more and more data, so the quality of that data has diminished with time. This has led to them setting up content-moderation factories in developing countries, where grossly underpaid workers sift through absolute junk. 

KH: This is one of the things I think is not effectively understood. Some people don’t even realise there are people in these countries being exploited in the first place, but the other misunderstanding I commonly confront is people who say, well, the internet is kind of like that anyway – social media has reams of people who are poorly paid, but [without them] we wouldn’t be able to have social media. 

I have lots of problems with that argument, but AI is quite different in that you wouldn’t need all these content moderators if you took a different approach to AI development. It specifically becomes a problem when you take the scaling approach and you try to scrape the whole of the English-language internet to train your systems, because then you start scooping up all this awful content, and your datasets are so large that they become incomprehensible… None of these companies have effectively figured out how to filter out the grotesque content from the internet… 

The best way to filter it before it reaches the user is to filter it on its way out of the model rather than on its way in. So, they filter in and out, but they know that the filtering in is imperfect, so the filtering out is like a second catch. We also know that it’s imperfect because now people are having AI psychosis due to problems with ineffective filtering on the way out. Then you end up having all these content moderators who need to train that filter by annotating any possible example of the type of content the filter should be blocking, which then requires them to wade for eight hours a day through awful, awful content and end up psychologically traumatised. 

CC: There are devastating environmental impacts, too. AI models require physical data centres to run, and these centres use colossal amounts of energy and potable water. Again, this comes down to an obsession with scaling and disproportionately affects countries in the Global South, particularly Chile. You describe Chile as ‘ground zero’ for a new scale of extractivism. How have these technology data centres changed in the wake of the AI boom, and why have countries in the Global South become the targets for tech companies to move in, extract and then move on? 

KH: The data-centre movement is unlike anything we’ve ever seen before. The modern internet is obviously also built on data centres, but previously, the pace at which data centres were expanding roughly matched the pace at which they were improving their efficiency. So most developed countries were seeing a flatlining in energy demand or a decline in energy demand, even as more data centres were being built. Now, the pace has dramatically exceeded any efficiency gains, so we’re seeing a historic uptick in the amount of energy that needs to be consumed globally. Almost all that uptick is due to the data centres being built. These companies have run out of land and energy and fresh water for supporting this, especially in places that have those resources, because, at the same time this is happening, climate change is accelerating. 

That’s why they’re pushing into some of these communities: there’s just not enough land in the US now; in a lot of developed countries, communities are pushing back more aggressively, and then it’s more easily escalated to the English-language media, which is really bad for the companies. So, if they move to another country where there’s less press, or the press is writing in languages that are less international, at least they can manage the PR more easily. 

A lot of these developing countries want the data centres because they think it [represents] economic opportunity. They don’t realise when they enter the bargain that they’re also giving up an extraordinary amount of their natural resources. If they realise it, they somehow think it’s still worth it – many of them live with a legacy of colonialism [and have] this mentality of ‘in order to be relevant to Global North powers, we just need to open our resources up for extraction’. So, you end up with this dynamic that repeats the dynamics of the past, where certain countries are just being hollowed out for their resources to continue perpetuating this expansion that ultimately is completely unsustainable for the planet and for people – not just in the environmental and public health senses but also in a pure physics sense. The laws of physics tell us that there is literally not enough stuff to be consumed in the world to support the trajectory these Silicon Valley companies say they’re on. 

CC: You talk a lot about people in the AI industry who have tried to speak out about these issues and have been silenced. But you also tell some really galvanising stories about people who are fighting back in different ways, like activist groups in Chile or an Indigenous couple in New Zealand who created their own LLM to help new generations of speakers learn te reo to try to revitalise the Indigenous language. Do you think we’ll start to see more of this pushback? 

KH: That’s actually been one of the things that’s been most heartening about being on tour: I’ve met so many people who are engaging in this kind of resistance. It’s not just anti-AI; it’s anti–Silicon Valley’s model of tech imperialism in general. I think a lot of people, whether they fully grasp the difference between AI versus their smartphone or social media or whatever, generally feel that they have a crappy relationship with technology these days.

One of my personal missions with the book is to help people connect the dots for what they need to push back against – and how. There’s so much opportunity for the average person to add their voice to shaping how technology is developed in the first place, and often people don’t realise that they have access to that kind of influence. You have the ability to go to your city council and tell them, ‘This data centre project, it’s not being transparent about the amount of energy and fresh water it’s using from our community, so we cannot allow it to proceed.’ That gives them something to grab on to, to operationalise the feeling they already have. People really get the issues, and they have the energy to do something about it. It’s just a matter of directing that energy. 

_{Image: Immo Wegmann from Unsplash}