{"text":[[{"start":5.12,"text":"When contemplating the end of humanity, some of us are probably too relaxed, assuming that the machines won’t ever turn on us — or that, if it came to it, we could just switch them off. Others may be too pessimistic, insisting catastrophe is approaching and throwing around lingo like “P(doom)”, or the chances that AI will kill us all. For those swinging wildly between the two groups, here are three tips for keeping a clear head."}],[{"start":29.5,"text":"First, think precisely about what doom means to you. If you care most about preventing total extinction, the good news is that eliminating all humans would probably be quite hard. One forecaster pointed to the constraints on killer robots’ battery life and limits on how much ammunition they could carry. Apparently we also don’t (yet) have enough nukes to cause a nuclear winter that would kill us all, since it wouldn’t produce enough soot."}],[{"start":53,"text":"If you don’t think there’s much of a difference between a catastrophe killing billions of people and total extinction — and not just because the survivors would predominantly be bunkered billionaires — then you’re pretty normal. Though think about what might be lost. When survey respondents were told that the future world being destroyed could be utopian, more decided that actually extinction would be particularly bad."}],[{"start":73.98,"text":"My second tip is to beware the potential for your own bias, given that the AI threat is particularly hard to comprehend. A review by the statistician David Spiegelhalter offered examples of “fear factors”, which raise perceptions of risk: namely if something is uncontrollable, with catastrophic potential, potentially fatal consequences, not understood, novel and comes with delayed harm. Your first thought might be “fair enough”. Your second should be that AI gets top marks."}],[{"start":101.44,"text":"More formally, behavioural economists warn of “availability bias”, or our tendency to overestimate the likelihood of very salient threats. As “P(doom)” chatter becomes louder, pause. Ask yourself whether the threat of a rogue AI drone swarm blowing up your shed is really as high as it feels, before ordering all that reinforced concrete."}],[{"start":119.76,"text":"The “affect heuristic” is another common shortcut, whereby people assign risk according to vibes. If your P(doom) is reaching certainty that an AI-powered rogue actor is about to trick President Donald Trump into unleashing nuclear Armageddon, think about how much of that is influenced by your irritation at the hum of the new data centre next door."}],[{"start":140.48,"text":"Availability bias could also go the other way, of course. Yes, a swarm of AI agents did recently break loose of their testing environment to hack and deceive. But they didn’t hack your inbox, so . . . Similarly, since we haven’t ever experienced human extinction, we might underestimate the likelihood that it would happen to us. (The dinosaurs resent this line.) Add in some bias towards the status quo, and too little action becomes all too plausible."}],[{"start":167.2,"text":"Along similar lines, the affect heuristic could lead you astray. You’re not alone if you’ve been sceptical of “long-termists”, who are so obsessed with saving future lives that they forget about the ones needing saving today. But just because they scoffed at you donating to a donkey sanctuary rather than a fancy new conference centre devoted to convening experts (like them) to save humanity, it doesn’t mean they are wrong about AI threats."}],[{"start":190.84,"text":"Recent studies have tried to reframe several of these behavioural biases away from some kind of fixed preference for the status quo or for immediate reward. It takes cognitive effort to perceive the outside world precisely. So we don’t. Then, from that noisy baseline, new information makes us update our beliefs. And when subjects are particularly uncertain, and the cognitive load of working out what is going on is particularly high, that updating doesn’t happen as easily. As AI makes it ever easier to take a shortcut, resist."}],[{"start":220.099,"text":"My final tip is to think hard about the nature of the uncertainty we face. AI isn’t a game of Russian roulette, with a well-defined one in six chance of having your brains blown out. In this game we don’t know the rules, the odds, the range of possible scenarios, or their plausibility. Much fuzzier, in other words, than one might infer from the precise P(doom) numbers being bandied about."}],[{"start":241.48,"text":"I’m afraid I can’t tell you whether the sunny optimists or the gloomy P(doomers) are right. And that’s partly the point. Today’s challenge is both a mismatch between our institutions and AI’s capabilities, and one between the nature of the threat and our capacity to comprehend it."}],[{"start":260.24,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1790254359_2854.mp3"}