Earlier this week, AI researcher Jacob Coxon announced his resignation from Anthropic, claiming that “no other human activity poses this level of danger,” and that those developing the technology think it could kill us all by the end of the decade. The day before, Labour MP Alex Sobel introduced the Artificial Superintelligence Security Bill, which was drafted by campaign group ControlAI, claiming that the heads of AI companies “put the chance it ends all human life between 10-25%”, and demanding Government controls to prohibit the development of superintelligent AI systems.

Considering the catastrophes the human race has survived to date — earthquakes, famines, floods, plagues, nuclear bombs, volcanoes and ice ages — it  seems unlikely that even a superintelligent AI could end all human life. But we should not ignore the prodigious growth of its powers that will transform our world. It is now starting to demonstrate its capacity not for consciousness, but for agency. Set it a simple task, and it appears to use its initiative to find a solution, with no regard for human laws or morals.

Governments should be anticipating the need to make AI users accountable for its impacts, developing laws to govern specific applications. The Medicines and Healthcare Products Regulatory Agency (MHRA) today called for better regulation of AI products used by doctors to take notes or analyse scans. But responsibility for the future harms of AI, as much as for its benefits, will lie with the humans setting it in motion.

So far, the cybersecurity failures that have resulted in major damage and disruption have been the work of human hackers and scam artists — sometimes aided by rudimentary AI tools, but clearly traceable back to malicious agents of the familiar, flesh-and-blood kind. When AI systems find a path of least resistance that involves hacking a national institution, disrupting a national energy grid, or bringing transport systems to a halt, it may be less clear who should end up in court, or pay compensation to the victims.

It’s easy to see why Sobel wants to see Government action, which prevents the development of technology capable of seizing control of Britain’s national infrastructure, or which is able to neutralise or circumvent human authority. Given how dependent we are on technology, we are increasingly vulnerable to artificially intelligent agents whose natural habitat is those same digital systems.

For example, think of the chaos caused by a brief failure of the UK Air Traffic Control system this week, or by the fire which disabled a single electricity substation and put Heathrow Airport out of action. Last year, a failure of Spain’s electricity grid to cope with unusual fluctuations of supply and demand plunged the entire Iberian Peninsula into blackout for days on end. What could a superintelligent AI do to such fragile infrastructure?

But this is the point: governments already need to take critical infrastructure more seriously. Rather than trying to protect it through futile attempts to prevent the development of self-improving and increasingly autonomous AI, authorities must take responsibility for ensuring it can withstand unforeseen problems. If one of the world’s busiest airports can be immobilised thanks to a fire caused by simple maintenance failures, a future superintelligent machine is the least of our worries.

---

Timandra Harkness presents the BBC Radio 4 series, FutureProofing and How To Disagree. Her book, Technology is Not the Problem, is published by Harper Collins.

TimandraHarknes

---