What are the chances that artificial intelligence could destroy humanity? Until recently, this wasn’t a question that most people worried about, yet in recent weeks, the discussion has broken through to the mainstream. You may have seen headlines saying there is , but what does that actually mean?
The honest answer is: very little. No one can produce a scientifically meaningful probability of an AI apocalypse – what people in AI circles call p(doom) – because it isn’t a scientific question. That is demonstrated by the fact that leading commentators in the field give numbers for p(doom) – covering the full probability spectrum tells us nothing.
Does that mean you shouldn’t worry about AI taking over the world? I can’t answer that question, because it depends on a number we have no reasonable way to derive. What I can do is compare it with other scenarios. Let’s start with an easy one: what is the chance of you losing the lottery?
Advertisement
First, as with any question of probability, we have to make some assumptions. I will assume you have actually bought a ticket and that you are playing the UK National Lottery. The odds of winning any prize, from a free play up to the jackpot, are about , or 20.4 per cent. That means your odds of losing are 79.6 per cent, which is how the lottery makes money – most players lose.
Let’s move to something more serious: what are your odds of dying in a road traffic collision? According to the World Health Organization, , so the odds of this happening to you are around 1 million divided by 8 billion, or roughly 0.01 per cent.
That is wrong, of course. The odds for you personally are highly dependent on many factors, including where you live – “more than 90% of road traffic deaths occur in low- and middle-income countries”, says the WHO – but we can use the figure as a ballpark. What matters is that the odds are low enough that we don’t worry about it all the time, but high enough that we take some precautions to mitigate against the possibility, from wearing seatbelts to looking both ways when we cross the road.
Spectrum of risk
These examples are about individual risks, but an AI apocalypse would happen to us all. Take climate change, perhaps the most pressing global risk. This is a far more complicated threat to assign a probability to for a few reasons. Unlike losing the lottery or being killed in a road traffic collision, climate change isn’t a binary outcome, but a continuum of warming above pre-industrial levels. It also cannot be calculated from known statistics, because we only have one world and the outcome lies in the future. Finally, the amount of global warming we will experience in 2100 depends on what we do between now and then.
None of this means we can’t assess the risk scientifically, however. Using extensively tested climate models, we can produce many “parallel worlds” in order to calculate statistics that show us how different climate policies are likely to result in different outcomes.
According to the , published last year, under current policies, there is a 100 per cent chance of exceeding 1.5°C of warming this century, a 92 per cent chance of exceeding 2°C and a 20 per cent chance of exceeding 3°C. These figures also come with error bars, but broadly give you a picture of where we are heading. On the strongest assumptions of climate action, the chances of exceeding 1.5°C, 2°C and 3°C fall to 79 per cent, 22 per cent and 0 per cent respectively. Clearly, these are actions worth taking.
Climate change is a relatively slow risk, however, compared with an AI suddenly wiping out the world. A better comparison might be an asteroid strike. We know that asteroid apocalypses can happen – just ask the dinosaurs – but what are the risks? To understand that, we need to know a few things: how many asteroids there are, the chance that one impacts Earth and the damage it could cause.
Fortunately, these numbers are all very well quantified. As you might expect, larger asteroids are both easier to spot and more deadly. , we have found more than 90 per cent of all asteroids above 1-kilometre wide (if you want to know how we can be sure, I explained the technique previously here).
What’s more, we are able to predict their future positions with high accuracy, so we know there is either no or an extremely low chance of one hitting Earth within the next century. Chicxulub, the asteroid that wiped out the dinosaurs, was about 10 kilometres wide, so the fact that we are essentially safe from anything an order of magnitude smaller and up is good news. The latest estimates suggest that an impact by an object 1 kilometre or larger only happens .
That isn’t to say we are completely safe from asteroids, because there are many more smaller asteroids and so the chances of them hitting are higher. An important size threshold is 140 metres, often dubbed a “city killer”. We haven’t found all of these, but we are working on it, and the latest estimates suggest asteroids above this size hit around every 20,000 years.
A unique apocalypse
In some sense, asteroids are reassuring, because we can see them coming, but what about the threat of a deadly virus? As we all know too well from our experiences of the covid-19 pandemic, a virus spreading around the world has the potential to appear from nowhere and upend life as we know it. But can we predict the chances of this happening scientifically?
Looking at the statistics, in roughly the past century, the largest pandemic was the 1918 flu, which killed an estimated 50 million people. The HIV pandemic, which began in the 1980s and is still ongoing, has killed more than 40 million people. Estimates for deaths from covid-19 vary, depending on how figures are counted, but the WHO reports a figure of over 7 million. Various other flu pandemics and regional infectious disease epidemics have killed smaller numbers of people, from hundreds to thousands to millions.
In other words, we should expect the most severe events to happen less frequently, and less severe events to happen more often. This is exactly the same pattern as the asteroid risk. Unlike asteroids, pandemics are much harder to observe before they strike, so their risk estimation is more like climate change, using models to examine a range of scenarios. While estimates vary, is that there is a 2 to 3 per cent chance of a covid-19-level pandemic every year, meaning we should expect one every 33 to 50 years.
What I have hopefully illustrated is that estimating the probability of an AI apocalypse is unlike any of the other scenarios I have discussed. Whether it is the lottery, a deadly traffic collision or an asteroid strike, we can apply statistics and scientific reasoning to answering the question. For p(doom), there is no scientific framework.
Where does that leave us? On an individual level, you should stop listening to the numbers being thrown about until they have serious backing – something that AI researchers should be working on constructing, rather than merely prophesying. For policy-makers – who must ideally plan for all scenarios that could lead to harm, even in the face of uncertainty – I would suggest considering regulation, but not rushing to action.
Ultimately though, your risk perception will come down to personal preference. For example, I disagree with Nvidia boss Jensen Huang, who recently said there was a . I would put the figure at some number, likely to be small, but greater than zero, simply because I don’t think it is ruled out by the laws of physics. Beyond that, I can’t say anything stronger – but compared with risks such as climate change or the next global pandemic, it isn’t something that keeps me up at night.
