AI doomsday headlines are naturally sparking fear, but the worst thing we could possibly do right now is slow down, writes David Grainger
The claim that we face a 10 per cent risk of AI destroying humanity within a decade has caused quite a response. Even the CEOs of Anthropic and rival OpenAI have joined calls for a slowdown in AI research while our understanding of these AI-powered agents and the damage they could cause catches up.
For many, this seems prudent. Maybe even essential. But it may be the wrong response entirely. Slowing down might even increase the risk to humanity that AI poses.
Why would AI agents destroy us? They don’t have motives like we do. The most benign reason is misinterpretation of an instruction – the cause of the unexpected hacking at Hugging Face.
A much more likely reason though is someone could build an agent with the express purpose of causing destruction. It is easy to demonise the technology, but whatever harm it can do, humans are worse.
AI can amplify the ability of bad actors to cause chaos on an unprecedented scale – whether to create a new and deadly biological weapon or turn off your water and power. In the age of AI, your enemies are more powerful than they have ever been. And that power is asymmetric – the more sophisticated your society, the more it depends on systems that AI can take over. The creators of AI are also, therefore, the most vulnerable.
The solution? If you listen to the CEOs at Anthropic and OpenAI, then it is to slow down. And maybe throw in a “kill switch” or two.
Why slowing down won’t work
Let’s kill the “kill switch” first. Rogue agents created by a hostile power could gain control of systems then disable any switch and cause irreversible damage before their target knew anything of the attack.
What about the slowdown? Surely that seems sensible?
Unfortunately, a containment strategy, like the world has mostly sustained for nuclear weapons for half a century, can never work for AI agents. The cat is already out of the bag – the know-how to create them is in the public domain. And unlike a nuclear bomb, which requires complex and difficult to obtain physical infrastructure such as giant centrifuges to enrich uranium, AI agents need little more than a home computer.
Vast data centres are only required where there are huge parallel demands for AI to solve many different tasks. Training and deploying specialised agents with very harmful intent requires no particularly large and expensive computing infrastructure. And smuggling a few gigabytes of computer code will be infinitely easier than crossing borders with enriched nuclear material.
No matter what we do, our enemies will have something approaching the capabilities we have. We cannot assume their technology will remain insufficient to seriously harm us, if indeed it is insufficient today.
That paints a doomsday scenario indeed. Our saviour, of course, is AI.
Who can build security systems for our critical infrastructure that AI cannot easily penetrate? Certainly not humans, as Hugging Face now knows to its cost. We cannot even fully understand the emergent behaviours of AI agents, which readily create complex systems as challenging to comprehend as life itself. If our understanding of human biology, after a millennium of investigation, is anything to go by then understanding what AI can and cannot do is not going to be quick or easy, if it is possible at all.
But the best AI can create systems that protect against the second best, and perhaps to a large extent, against itself. The world that is dawning faster than many ever imagined features a new arms race. It’s one we must win.
In biology, our immune systems evolve to defeat pathogens such as bacteria and viruses, then those pathogens evolve to evade our immune systems, and so on. It is a battle which is never, and will never be, won.
The same has happened with military technologies. You build a better attack, so I build a better defence. No-one ever gains absolute victory, just a strictly time-limited advantage.
What then would be the worst thing we could possibly do right now? Yes, slow down. We would only squander our edge, the one thing that can protect us from what is coming. And come it will, whether we slow down or not. In the infinite monkey cage of evolution, speed wins.
It’s easy to be scared. It’s easy to think that AI is really different from anything and everything that has come before. But it is not. It’s just another tool that stretches the ability of humankind for good and evil. From the first bone tools of our neolithic ancestors to the first gun, AI is just the next tool or weapon off the taxi rank. Controlling it requires the same time-tested strategy – run as quickly as we can.