Everyone laughed politely when I asked whether AI might, in the end, kill us all — everyone apart from one researcher from a US-based thinktank who told me that, as a pessimist, he believes there is about a “20 per cent chance that AI poses an existential threat to life.”
“That’s a smaller percentage than you used to say,” an eminent professor from Stanford University quipped, “outlooks have improved?”
“Well, just as long as people are afraid, fear is healthy."
The night before November’s AI Safety Summit was due to take place at Bletchley Park I was invited to a drinks event hosted by Google DeepMind, one of the world’s largest and most influential AI companies. Headquartered in London since its inception as DeepMind in 2010 and founded by two Londoners (Demis Hassabis and Mustafa Suleyman) and a Kiwi (Shane Legg), the start-up was acquired by Google in 2014 and over the past fifteen years has been at the forefront of an AI revolution that — depending on which side you land on — could solve the climate crisis and cure all cancers or achieve God-like super intelligence and kill us all in ways we’ve yet to dream possible.
We're in a long term mission to develop ever more general, more capable AI technology through research and you've got to take a long term view to do that
In the 24 hours leading up to the Summit, world leaders — including US vice president Kamala Harris — and tech titans like Elon Musk, had flown into London for two days of intensive talks about how to keep the world safe in this era of transformational tech. A week before that Rishi Sunak had made a speech in which he warned that, “mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war”. He borrowed the warning from a stark 23 word open letter which circulated back in June, signed by a coterie of AI industry leaders — DeepMind founder Hassabis among them.
No such doomsaying at the party I noticed. “It's no longer a technology of the future, it's here today…” Hassabis said in a speech at the beginning of the evening. “We have projects working all the way from weather prediction and predicting the path of hurricanes, to identifying new materials that could transform our economy and society.” The most doomy it got, in fact, was a nod to the frontier nature of this work: “as with any new transformative technology, we're going to see new challenges and risks that we're also going to have to manage.”
Still, in the 24 hours after the party, when most of the attendees I’d sipped lukewarm Prosecco with, had decamped to Bletchley, 28 governments — including China — signed the Bletchley Declaration, which states the need for international cooperation on AI on the grounds that: “There is potential for serious, even catastrophic, harm, either deliberate or unintentional, stemming from the most significant capabilities of these A.I. models.” This didn’t stop Elon Musk, a day later painting a picture of a utopian future where AI has done away with the need to work and where “we won’t have universal basic income but universal high income”.
This is the interesting thing about AI — no one, including tech insiders, knows where it could go and how dark or utopian it could get. In the meantime, though, the drive to innovate (and create profit) continues apace, fuelled by a global investment boom which has seen ever more eye-watering sums raised by companies (some of which have little more to show than a good idea and ‘AI’ in their name).
Undoubtedly, 2023 was a year of incredible innovations in the world of AI — most strikingly, it will be known as the year that generative AI truly entered the mainstream. At the end of November, Ofcom found that 79% of 13-17-year-olds were using tools such as ChatGPT or Bard on a regular basis, most for fun but a significant minority (just over a third), to help with work and studies. A week later, at the beginning of December, Google DeepMind announced a new era for Gemini — their “largest and most capable AI model” and the “next step in our journey to make AI more helpful for everyone.” The announcement highlighted Gemini’s capacity to draw information from image as well as text data — and went viral for its mind-bogglingly sophisticated answers to seemingly complicated problems.
Google DeepMind is now one of London’s most successful start-ups, ever. From their Kings Cross campus the company — still helmed by Hassabis — has become a key player in the vast existential chess game that is artificial intelligence. Are they in it for the good of humanity? For profit? For power? I decided to spend a day there to understand how one London company plans to shape the future of the world.
***
In his interview for a job at DeepMind, nine years ago, Chief Business Officer Colin Murdoch says he remembers how Shane Legg flipped a laptop round to show him how an AI programme had learned a strategy to beat a popular computer game. “The offices were next to Russell Square tube back then. There are about 170 steps which go down to the tube station and I remember I was just buzzing as I went down them. I think that was my AI ‘Eureka!’ moment when I realised the technology has incredible potential.”
In 2023, Murdoch helped to oversee the merging of ‘Brain’ — Google’s AI team — and DeepMind to form Google DeepMind. Before, the two had operated as separate entities - and there was speculation that the successful launch of ChatGPT had spooked Google into turning the heat up on their own AI operations.
There were also rumours of a tussle within the company about who’d win control of the newly amalgamated teams — a tussle that Hassabis evidently won, much to the benefit of London. The fact is, the presence of Google DeepMind makes London a player on the AI world stage (as the Tony Blair Foundation warned back in June 2023: “the UK’s enterprise is overly dependent on a single US-owned and funded entity, Google DeepMind.”)
“It’s been good,” Murdoch says brightly, of the merger. “These are probably two of the best AI research teams in the world. And we brought them together at a point where the whole progress in the field is really accelerating. So we're a little bigger now, which means we can do more research in a more coordinated way.” Like most people that I meet on campus Murdoch is cultishly positive about DeepMind. I get chatting to someone in the canteen queue (all meals are free at DeepMind), who tells me that the company is “passionate about its employees” as if he's reading from a brochure. On the corridors, there is employee art displayed on the walls; there is also a free gym, a room with DJ decks for social events and snack stations at every turn. Equally, though, the Fort Knox level security hints at the fact that the work being carried out here is of the utmost commercial value.
Despite that, DeepMind’s output isn’t actually judged on how much money its AI products make. In fact, the company only began to turn a profit in 2021 — before that it was profoundly loss making (in 2019, for instance, the company posted losses of £477m up from £470m in 2018); that year Google also waived the repayment of an inter-company loan to the tune of £1.1bn.
The business model is such that the software created at the Kings Cross campus is sold into other Alphabet-owned companies (like Google and YouTube). “We've always operated in a mode where we execute our research agenda, then we provide that back to Google and Alphabet to enhance their products and services,” says Murdoch. This is why Gemini, despite the flashy announcement just a few weeks ago, has already been seeded into many of Google’s products, including Search.
“That's how they mark our scorecard at the end of each period, like ‘how much research progress did you make? How did you help enhance our products and services?’” So what happened in 2021 to suddenly make the company profitable? “I can't get into the specifics of the financials,” he demurs. Perhaps it has something to do with the fact that it was only in 2021 when all the research began to have truly useful applications though, as Murdoch points out, “we're in a long term mission to develop ever more general, more capable AI technology through research and you've got to take a long term view to do that.”
***
For me, the profit question is really only interesting because if — as so many warn — we’re talking about extinction level threat, at what point does the pursuit of profit begin to put us all in harm's way? “I think profit and responsibility have to run in parallel,” says Dawn Bloxwich, director of Pioneering Responsibly at Google DeepMind. She says that she’d class herself as a ‘cautious optimist’ when it comes to AI — “I'm not crazy optimistic but I really believe that our technology can do a lot of good in the world.” As evidence she points to AlphaFold, one of DeepMind’s AI programmes which can predict the 3D shapes of almost all proteins known to science, promising to revolutionise drug development.
‘Pioneering responsibly,’ she explains, is a kind of statement of intent. “It means that we want to move forwards at pace, but also think about the ethical and safety aspects of our work and weave those considerations into every step of the development process.” Basically, her teams are responsible for overseeing the ethical implications of the work being done on site — and putting in appropriate safeguards and policies as new technologies are developed. But it’s not all about existential-level threats, she says.
“For each technology we have a spectrum of risks that we're looking at,” explains Bloxwich, “these include potential biases, how a technology could be used for misinformation and disinformation, impact on underrepresented groups — as well as whether these technologies can persuade and manipulate.” Persuasion, she says, is, in fact, a growing threat within the AI sphere — and one that isn’t yet easy to plan for. “None of the questions we’re grappling with are simple. For a start, they're not always things that people have done a lot of in depth research on in an AI context,” says Bloxwich.
Still, in 2023, researchers at Stanford University’s Polarization and Social Change Lab and the Institute for Human-Centered Artificial Intelligence (HAI), found that “AI-generated persuasive appeals were as effective as ones written by humans in persuading human audiences on several political issues.” Though the written appeals — produced by ChatGPT 3 and covering polarising topics like a smoking ban, gun control and a carbon tax — only moved the dial a few points (i.e. only a few people were persuaded to change their views after reading the materials), extrapolated out to population-wide data set, it was clear that AI manipulation of an election would be entirely possible. “AI has the potential to influence political discourse, and we should get out in front of these issues from the start,” urged the researchers. In October, ChatGPT founder Sam Altman, tweeted a similar sentiment, saying that he expected “ai to be capable of superhuman persuasion well before it is superhuman at general intelligence, which may lead to some very strange outcomes.”
i expect ai to be capable of superhuman persuasion well before it is superhuman at general intelligence, which may lead to some very strange outcomes
— Sam Altman (@sama) October 25, 2023
The problem for AI developers is that manipulation can be good if it’s used for good — “so, for instance, persuading people to change bad habits, to eat better, perhaps,” says Bloxwich. She says that her teams are always working to develop more robust tests for the persuasiveness of technologies though she points out that, when it comes to the kinds of influence we hope AI will have on our world, the question is really about what kind of society we want to live in. “I've had quite a few conversations with philosophers externally,” she says, “and it keeps coming up: ‘what is the future that we want to [create] and whose future are we creating here?’”
None of the questions we’re grappling with are simple. For a start, they're not always things that people have done a lot of in depth research on in an AI context
Tools like Gemini — or indeed ChatGPT, AlphaFold and any number of others — have begun to construct our everyday reality, but they’re often governed by rules and parameters that reflect the social setting which have created them. “I think ultimately when it comes to AI, we as a society with lots of different external groups need to have a conversation about what we want our future to look like. And which rules it should be governed by. That's not our job, as a tech company; we can obviously contribute and be a part of it, but it shouldn't be our decision.” She points to the increase in international summits and the growth of governance that we’re beginning to see “I think we'll see governments step in as well, over time, to bring in more regulation” but in the meantime, it does seem that it falls to a handful of private companies to construct the moral codes of these machines.
***
Last year wasn’t just a significant one for AI breakthroughs — there were also a growing number of dissenting voices which came out in criticism of tech’s big leap into the unknown. This included Mustafa Suleyman, one of the original co-founders of DeepMind whose book The Coming Wave laid out in stark detail the transformations that AI will bring about, many of them unwelcome. Part of the problem with the current rate of innovation, he argues in the book, is the degree of “pessimism aversion” that he sees within the AI field. Basically, researchers aren’t ready to admit that things may be getting out of hand.
Pushmeet Kohli is Vice President of Research at Google DeepMind and has been working for the company since 2017. He’s softly spoken and quietly impressive; he considers the idea of “pessimism aversion” when I put it to him. “Well, as things come out from the lab they interact with a messy world, right? So how will they behave in that messy world? An appropriate amount of safety and reliability evaluation needs to be invested to understand those behaviours. But I don't think researchers in the field are shying away from this.” He argues that AI researchers aren’t motivated by pure curiosity, and to hell with the consequences. “People are really, I think, focused on understanding the implications, understanding the risks, they're not shying away from them.”
AI researchers are really focused on understanding the implications, understanding the risks, they're not shying away from them
Despite operating at the cutting edge of AI research, he’s fairly conservative about his estimates for when we’ll reach the holy grail of artificial general intelligence (AGI) (basically meaning that an algorithm can operate with the same level of mental dexterity as a human). “A number of problems remain,” he says. “Not only in matching the competencies of human reasoning, but also, even when you have this powerful tool, how do you align it effectively, verifiably and safely with what society and what individuals users want out of it?” It’s the ‘meeting the messy world’ element: researchers — like Bloxwich — are still grappling with how to create effective tests for these algorithms — and unleashing them, Kohli argues, is a number of scientific steps beyond that. “To give an example, if you have an AI system which might be used for healthcare, right? How do you make it interpretable? And who should be able to interpret it's reasoning? Is it the doctors, is it the designers or is it the patients? And there are different forms of interpretability — what might be obvious to a clinician, or a machine learning person, might not be obvious to a patient. Making sure that these systems are deployed safely in the real world is a whole research problem in its own right.”
Even truly useful AI assistants and chatbots, he says, are yet some technological leaps away. “Think about when you speak to a person, you don't want them to give you an answer if they’re not an expert, and know nothing about the question you’ve asked. But of course, some people do that — machine learning systems are kind of the same. We have to teach them to be responsible — to, to say, ‘here's something that I know about and that I can tell you,’ but then admit ‘here's something I don't know about, so take this with a huge sort of tablespoon of salt.’ Building systems which understand their own uncertainty is important.”
Despite all that, he’s optimistic about AI’s capabilities — and what the next twelve months could bring. There are so many innovations happening within the building, he tells me, that it’s almost impossible to single one out as particularly exciting, though he does enjoy tools like AlphaFold and the more biological research that Google DeepMind is engaged in. “We as a species are so sophisticated, so complex — the fact that we are able to reason, the fact that we are able to scale in this way, fascinates me. And it's exciting to finally have a tool that can shed light into the underlying mechanisms of life.”
