A New Scientist column tests p(doom) against three risks that can be counted

What separates a computed probability from an elicited one is whether there is something to count. The lottery has a sample space. About 1 in 4.9 UK National Lottery tickets wins some prize, which the column gives as 20.4 per cent, so 79.6 per cent lose ^[4][5].

Road deaths are messier, and the column is candid about it. The World Health Organization puts annual road traffic deaths slightly above 1 million, and dividing by 8 billion people gives 0.0125 per cent for a single year ^[6][1]. The column rounds that to roughly 0.01 per cent and then says plainly that it is wrong for any individual ^[7]. "More than 90% of road traffic deaths occur in low- and middle-income countries", the WHO says, as quoted in the piece ^[8]. Hold the same annual rate for 70 years and it compounds to about 0.87 per cent, near 1 in 115 ^[2]. The factor of 70 is the length of the window.

Climate is the harder case, and the column lists why: warming is a continuum, there is only one world to observe, and the 2100 figure depends on choices not yet made ^[9]. Tested models supply the missing sample by generating many "parallel worlds" whose statistics can be compared across policies ^[10]. Under current policies, the UN Environment Programme's Emissions Gap Report gives a 100 per cent chance of exceeding 1.5C this century, 92 per cent for 2C and 20 per cent for 3C; on the strongest assumptions of climate action, those fall to 79, 22 and 0 per cent ^[11][12]. The 2C number moves 70 percentage points between the two cases ^[3].

The asteroid comparison needs three inputs, which the column names: how many asteroids there are, the chance one hits Earth, and the damage it would do ^[13]. NASA has found more than 90 per cent of those wider than 1 kilometre, the bigger ones being both easier to spot and more deadly, and their future positions can be predicted ^[14].

So the 0-to-more-than-95-per-cent spread among leading commentators ^[3] is evidence about the estimation procedure, not about machines. That width is what a quantity looks like when no shared method exists to derive it: every answer is admissible because none can be checked against anything. None of that means the risk is small. A hazard nobody can quantify has not been quantified as zero, and the same range that holds the zeros holds the 95s ^[3]. A figure produced this way should not be handled like the UNEP percentages, which arrive with error bars and a stated dependence on policy ^[11].

One more caution about the headline number. When coverage reports a greater-than-10-per-cent chance that AI could kill all humans ^[1], the column's answer to what that means is that it means very little, because the question is not a scientific one ^[2]. The road-traffic example shows what is missing before any such figure can be compared with another: over what period, and to whom.