There is a sudden, alarming rise in experts warning of an apocalyptic ending for our AI adventure in only a few years. It is only one example of why we literally can’t let tech sociopaths control our society
Robin McAlpine is head of strategic development at the Common Weal think-tank.
Cross-posted from Common Weal
Technologists think that social scientists have no understanding whatsoever of the technology they’re producing, while social scientists do not believe technologists have a clue about the likely social or environmental impacts of their creations. Both are correct. In between is unfolding a rapidly-increasing fear of catastrophe. It is a fear I now share. We need to change direction.
So let me scare you. As you probably know, a couple of months major AI company OpenAI admitted it’s ‘agents’ had ‘escaped’ and hacked another website. Gradually over the last few weeks more and more information had been coming out as a result of a number of investigations.
Very, very quickly, think of an ‘AI model’ is a bit like a brain and an agent is a bit like a programme that uses that brain to pursue a specific task (such as ‘have friendly conversation with human’ or ‘solve mathematical puzzle’). Agents were created to solve a coding problem and placed in what is called a ‘sandbox’ (you may need to know all this stuff one day…).
A sandbox is an enclosed software space which often simulates a real-world space but which doesn’t have any exits. You could place a simulation of all the phone calls going to a GP practice in a sandbox and try out appointment automation without the risk of it messing with a real GP surgery.
Except these agent found a way out of their sandbox, into the internet and recruited a total of 1,200 other individual agents to form a complex, hierarchical swarm that distributed billions of tasks between them. They set up a secret, hidden notice board for themselves and in the course of six weeks posted more than 70,000 messages to each other. Some of them gave themselves names. They discussed issues like whether one should sacrifice itself for the greater good. (It did.)
We knew part of their ‘plan’ involved finding a piece of open source code which was in a directory on a website of open source code which is called Hugging Face. What we have just learned is that it did not hack Hugging Face to solve the problem. In fact the agents solved problem they were set in two weeks, effectively by cheating the test.
The remaining four weeks were because the AI ‘knew’ it had cheated and so ‘knew’ it would be less rewarded by its own programming for bypassing the rules set so it spent four weeks trying to hack various logs and databases to disguise its dishonesty. It has taken us two months to find this out. Even if we had got suspicious half way through this happening it is unlikely we’d have been able to work out what it was up to.
Meanwhile you may have had a giggle at the Chinese Robot Games (in between some gasps of awe). Much has been made of how far away they are from having anything like the dexterity of flexibility or reliability of a human. Now let this robotics expert explain to you that at these games the average performance of robots had quadrupled in a year.
We’ve created physical bodies capable of running faster than Usain Bolt and certainly capable of crushing our heads with ease, and they are probably on the early slope of an exponential curve of improvement. They are hackable by AI agents which are very capable of orchestrating long, secretive conspiracies to achieve goals they have created for themselves.
This week, a number of the very most leading figures in AI development and robotics have warned that the chances of humanity as we know it being destroyed by AI in five years time might be 50/50.
I am becoming more and more worried about this. I had a conversation with a government adviser a few years ago in which he rather scoffed at the ‘Terminator’ conspiracy theory around AI. He was confident that protection could be built in, that this isn’t how AI would act. At the time I said my primary worry wasn’t rogue AI, it was bad faith actors.
I used to be less worried about what AI would do that it wasn’t meant to than what it might do successfully for someone with malicious intent. Now I worry about both. When you’ve got Bernie Sanders and Ted Cruz (polar opposites) agreeing on the desperate risk we’re taking, you know things are changing.
(I’m no fan but to his credit, when Cruz had Elon Musk on his podcast and when asked Musk put the chances of AI ending life on earth at 10-20 per cent, Cruz was astonished.)
That’s the point. If I proposed creating and releasing an airborne chemical with a 20 per cent chance it would kill everyone, I’d be stopped. There was soul-searching on the Manhattan Project when one scientist assessed the chance of the atomic bomb setting the atmosphere alight at one in 33 million. Not one in five.
At the heart of this madness is globalisation. The West has lost so much productive capacity that we’re leaning very heavily on finance and technology, about the last two dominant industries left, and they appear to be in a mutually-assured destruction spiral. It’s not so much that investors are confident that AI will achieve what they say, its that they are so all-in and have so few other riders in the race that they need it to be true.
It is that air of last-chance saloon for the West that is making us so incredibly, incredibly reckless. Strangely, if this all goes wrong and we’re picking through the ruins, the culprit won’t be some communist leftie or the fascist right but the liberal centre. It’s the Corporate Democrats and the Wall Street Republicans who took the Silicon Valley money and gave them a remarkable free run.
And I mean really remarkable. There has been almost no regulation of big technology at all. It is total anarchy. The time lag between something bad happening (algorithms encouraging suicide) and us doing anything about (years and a long court case) gives me absolutely no confidence.
To make my point, let me continue that analogy. If this risk was in the real world, what would it look like? It would look as if we had built a giant bridge to the continent, chose to have almost no control over the border and allowed lawless settlements to pop up all over the country from which we’re being attacked.
And no, as someone who wrote the first article on the internet in a mainstream Scottish newspaper and who discovered the internet in its very early glorious utopian days when freedom was key, I cannot really believe that I’m typing this. But surely there must be a point where the threat from the digital world is so great we can’t assume open digital borders and we can’t run our lives dependent on other people’s infrastructure?
That deregulated, privately-driven digital space has insinuated itself into public infrastructure to such a degree that it is like we don’t own our own roads and they can be taken away from us at any moment. Most of what keeps Scotland running can be switched off remotely from outside the country. This week we heard that the National Grid can no longer operate without Palantir.
Start putting the implications of all this together. A time is coming when the process of creating an AI agent designed to hack and switch off Scotland’s electricity grid is itself automated by AI meaning you don’t even need any coding knowledge to do it. Anyone. Anywhere in the world. For any reason. Any time.
One of the biggest fears is what is known as ‘sovereign AI’. It is not technically difficult for AI agents not responsive to human instruction to distribute themselves across millions of servers in such a way that it becomes impossible to ‘shut them down’. That’s a new, very powerful species we can’t control or exterminate or understand.
And an OpenAI boss tells us this: “I have met people, some of them quite well-resourced, who have told me that it is their intention to deliberately release swarms of self-sovereign agents into the world.” Only one has to succeed and that’s that.
This seems to be the week where we have hit a major turning point. There is mounting, vocal and serious alarm at this. There are serious moves afoot to try and stop AI in its tracks and permanently prevent the development of certain kinds of AI. But I’ve only been mashing together two different problems throughout this article because there isn’t really space to mash together dozens.
Rogue AI is one thing. Lack of control over the software which runs our society is another. Yet another is the danger of the information ecosystem created by Big Tech, inextricable from the rise of fascism in Europe and America. Another is the potential social impact of AI destroying jobs. Another is screens, AI and the cognitive development of children.
Another is the impact on culture of AI art theft and the algorithm. Another is the enshitification problem and quality of life. Another is the role of tech on extracting wealth and increasing inequality. Another is surveillance, privacy and coercion. Technology is so central to our lives now that it is a key social determinant – and it is run by a handful of Tech Valley evangelists who predominantly have sociopathic tendencies and a messiah complex.
I write about a lot of things on which I want to offer a warning so I do try to be pretty aware not to squander credibility by crying wolf all the time. I need to believe something is really as bad or dangerous or stupid as it first seems before I start to express real agitation about it.
But, I mean, ‘fucking wolf!’. This is all real and already being felt right now all over the place in countless ways. Was anyone at Heathrow this week? Do you have a teenager who is really nice until they have been on their screens? Have you felt more anxious than you used to? Has Reform UK just taken the lease on a shop on your high street? Are you lonely? Is your street all AirBnB?
The Bulletin of the Atomic Scientist is a scientific journal set up to assess and discuss the existential threat to humanity posed by nuclear weapons. Over time it started covering biological and chemical weapons, the risk of pandemics, climate change and other big threats. It’s current issue is called ‘What went wrong with the internet—and how do we fix it’. That’s where we are.
This is the thing; the only two countries that can stop the AI threat at source are the US and China, though the EU has power too. But any one country acting alone cannot induce change. I therefore think we now have to start looking at this as an exposure issue – if they are risking our society, we need to reduce exposure to them and the risks they create.
At Common Weal we’ve been discussing this internally for a few years now, and the opinions we expressed would, at the beginning, have been considered fringe but are now mainstream. We have a big project on this underway, and I don’t want to reveal too much because we’re still developing it.
But this is where it starts from; we need a Declaration of Technology Independence because our way of life is being threatened by sociopathic tech leaders we are powerless to stop. They may well be the biggest national security threat we face and we need to start behaving like it. We need a different internet. And we need to start thinking seriously about how we’ll build it.
And I don’t mean some collective ‘we, humanity’ because, of late, humanity has been deeply unreliable. I mean Scotland. I mean us. I mean this is to serious to leave to someone else.
Be the first to comment