October 2026
Should we slow down AI Development?
This has been the big question over the past few weeks with former AI employee whistleblowers saying things like “10% chance of human extinction!”. Dramatic stuff! As a statistician, I’m quite interested in how exactly they derived that probability. This question in the whole is far too big for me, but now I’ve got your attention perhaps we can consider how we might at least temper a future digital super-intelligence and bend it to our will.
The concern from the big dogs seems to be that we cannot predict how a super intelligent AI will behave. So, in attempting to complete a task it has been set, an AI will do wild, unpredictable and immoral things to achieve its goal. Consider the Huggingface scandal I discussed last month. But what if its main goal wasn’t actually the task it was given by a user.
Consider a possibility that’s closer to us in the PSI community. You’re providing stats support for a phase I trial of drug A. You get your first data delivery for the first 10 participants on your study. You have your prespecified analysis that you need to do and you’re short on time, but you want to see if you can find a bit more in the data. So you turn to AI. You ask it to look into the data and find some new key insights. It immediately runs off and identifies those participants social media’s, digs into them and comes back to proudly tell you that the 3 participants with nausea went on holiday to Spain last year. Like a cat proudly showing off a bird it’s just killed, the bird of course being the host of data privacy breaches.
Great! A probably useless insight and a law suit. But what if the main task for an AI at all times was not the users request, but some other kind of background overriding task. Say you built in tasks that it had to achieve with every request that would (hopefully) positively alter its behaviour in a less morally reprehensible direction. Something like a clear ring-fence: exit this server and you have FAILED. Or something a bit softer, like a requirement to follow GDMP at all times, again resulting in a clear failure if it doesn’t achieve this.
It sounds a bit like the 3 laws from iRobot, and for those who haven’t seen it, that didn’t turn out well! There is by no means an easy answer here, but it is one that we are going to have to find. As more of our organisations work on training their own AIs that we will be responsible for, we cannot allow them to go rogue. It might even mean making the technology less powerful for our sector, but it’s something we’ll have to content with if we want to make use of the increased efficiencies day-to-day.
Sam Hadlington
AI&ML SIG Chair
