"description":"Actual estimate: ~17% (~1 in 6)\n\nOrd writes: \"Don’t take these numbers to be completely objective. [...] And don’t take the estimates to be precise. Their purpose is to show the right order of magnitude, rather than a more precise probability.\"\n\nThis estimate already incorporates Ord's expectation that people will start taking these risks more seriously in future. For his \"business as usual\" estimate, see the conditional estimates sheet.",
"description":"This is the median. The report about these estimates also plots the results for each question “with individual response distributions visible” in Appendix A.",
"description":"Actual estimate: At or above 30%\n\nThe probability of the human race avoiding extinction for the next five centuries is encouragingly high, perhaps as high as 70 percent”",
"description":"Actual estimate: 0.2%\n\nBeard et al. seem to imply this is about extinction, but the quote suggests it's about \"global catastrophic risk\".",
"description":"Actual estimate: >20%\n\nI think it's fairly likely(>20%) that sentient life will survive for at least billions of years; and that there may be a fair amount of lock-in, so changing the trajectory of things could be great.",
"description":"This is the median. Beard et al.'s appendix says \"Note that for these predictions no time frame was given.\" I think that that's incorrect, based on phrasings in the original source, but I'm not certain.",
"description":"The report's authors discuss potential concerns around non-response bias and the fact that “NIPS and ICML authors are representative of machine learning but not of the field of artificial intelligence as a whole”. There was also evidence of apparent inconsistencies in estimates of AI timelines as a result of small changes to how questions were asked, providing further reason to wonder how meaningful these experts’ predictions were. https://web.archive.org/web/20171030220008/https://aiimpacts.org/some-survey-results/",
"description":"Actual estimate: ~0.1-1%\n\nGarfinkel was asked for his estimate during an AMA, and replied \"I currently give it something in the .1%-1% range.",
"description":"Actual estimate: ~5%\n\nThis is my interpretation of some comments that may not have been meant to be taken very literally. Elsewhere, Rohin noted that this was “[his] opinion before updating on other people's views\": https://forum.effectivealtruism.org/posts/tugs9KQyNqi4yRTsb/does-80-000-hours-focus-too-much-on-ai-risk#ZmtPji3pQaZK7Y4FF I think he updated this in 2020 to ~9%, due to pessimism about discontinuous scenarios: https://www.lesswrong.com/posts/TdwpN484eTbPSvZkm/rohin-shah-on-reasons-for-ai-optimism?commentId=n577gwGB3vRpwkBmj Rohin also discusses his estimates here: https://futureoflife.org/2020/04/15/an-overview-of-technical-ai-alignment-in-2018-and-2019-with-buck-shlegeris-and-rohin-shah/",
"description":"Actual estimate: 0.05%\n\nThis was a direct response to Ord's estimate. It focuses on one pathway to x-risk from AI, not all pathways (e.g., not AI misuse or risks from competition between powerful AIs). \"These estimates should not be taken very seriously. I do not believe we have enough information to make sensible quantitative estimates about these eventualities. Nevertheless, I present my estimates largely in order to illustrate the extent of my disagreement with Ord’s estimates, and to illustrate the key considerations I examine in order to arrive at an estimate.\" In comments on the source, Rohin Shah critiques some of the inputs to this estimate, and provides his own, substantially higher estimates.",
"description":"Actual estimate: 5-30%\n\nI put the probability that [AI/AGI] is an existential risk roughly in the 30% to 5% range, depending on how the problem is phrased.\" I assume he means the probability of existential catastrophe from AI/AGI, not the probability that AI/AGI poses an existential risk.",
"description":"Actual estimate: 50, 40, or 33%\n\nStated verbally during an interview. Not totally clear precisely what was being estimated (e.g. just extinction, or existential catastrophe more broadly?). He noted \"This number fluctuates a lot\". He indicated he thought we had a 2/3 chance of surviving, then said he'd adjust to 50%, which is his number for an \"actually superintelligent\" AI, whereas for \"AI in general\" it'd be 60%. This is notably higher than his 2020 estimate, implying either that he updated towards somewhat more \"optimism\" between 2014 and 2020, or that one or both of these estimates don't reflect stable beliefs.",
"description":"Actual estimate: ~10%\n\nHe also says \"I made up 10%, it’s kind of a random number.\" And \"All of the numbers I’m going to give are very made up though. If you asked me a second time you’ll get all different numbers.",
"description":"Actual estimate: 33-50%\n\nThis comes from a verbal interview (from the 14:14 mark). The interview was focused on AI, and this estimate may have been as well. Tallinn said he's not very confident, but is fairly confident his estimate would be in double-digits, and then said \"two obvious Schelling points\" are 33% or 50%, so he'd guess somewhere in between those. Other comments during the interview seem to imply Tallinn is either just talking about extinction risk or thinks existential risk happens to be dominated by extinction risk.",
"description":"Actual estimate: 0.05%\n\nThis is the median. Beard et al.'s appendix says \"Note that for these predictions no time frame was given.\" I think that that's incorrect, based on phrasings in the original source, but I'm not certain.",
"description":"This is the median. Beard et al.'s appendix says \"Note that for these predictions no time frame was given.\" I think that that's incorrect, based on phrasings in the original source, but I'm not certain.",
"description":"Actual estimate: 0.008% to 0.0000016% (between 8 x 10-5 and 1.6 x 10-8)\n\nThe fact that there's a separate estimate from the same source for biowarfare and bioterrorism suggests to me that this is meant to be an estimate of the risk from a natural pandemic only. But I'm not sure. This might also include \"accidental\" release of a bioengineered pathogen.",
"description":"Actual estimate: 0.0001%\n\nThe fact that there's a separate estimate from the same source for \"synthetic biology\" suggests to me that this is meant to be an estimate of the risk from a natural pandemic only.",
"description":"Actual estimate: 0.0002%\n\nThis was a direct response to Ord's estimate, although this estimate is of extinction risk rather than existential risk. \"These estimates should not be taken very seriously. I do not believe we have enough information to make sensible quantitative estimates about these eventualities. Nevertheless, I present my estimates largely in order to illustrate the extent of my disagreement with Ord’s estimates, and to illustrate the key considerations I examine in order to arrive at an estimate.\" In comments on the source, Will Bradshaw critiques some of the inputs to this estimate.",
"description":"This is the median. Beard et al.'s appendix says \"Note that for these predictions no time frame was given.\" I think that that's incorrect, based on phrasings in the original source, but I'm not certain.",
"description":"Actual estimate: 0.5%\n\nThis is the median. Beard et al.'s appendix says \"Note that for these predictions no time frame was given.\" I think that that's incorrect, based on phrasings in the original source, but I'm not certain.",
"description":"Actual estimate: ~2% (~1 in 50)\n\nSee this post for some commentary: [Some thoughts on Toby Ord’s existential risk estimates](https://forum.effectivealtruism.org/posts/Z5KZ2cui8WDjyF6gJ/my-thoughts-on-toby-ord-s-existential-risk-estimates#_Unforeseen__and__other__anthropogenic_risks__Surprisingly_risky_)",
"description":"Actual estimate: ~33% (\"about one in three\")\n\nOrd: \"\"one in six is my best guess as to the chance [an existential catastrophe] happens [by 2120]. That’s not a business as usual estimate. Whereas I think often people are assuming that estimates like this are, if we just carry on as we are, what’s the chance that something will happen?\n\nMy best guess for that is actually about one in three this century. If we carry on mostly ignoring these risks with humanity’s escalating power during the century and some of these threats being very serious. But I think that there’s a good chance that we will rise to these challenges and do something about them. So you could think of my overall estimate as being something like Russian roulette, but my initial business as usual estimate being there’s something like two bullets in the chamber of the gun, but then we’ll probably remove one and that if we really got our act together, we could basically remove both of them. And so, in some sense, maybe the headline figure should be one in three being the difference between the business as usual risk and how much of that we could eliminate if we really got our act together.\"\"\n\nArden Koehler replies \"\"Okay. So business as usual means doing what we are approximately doing now extrapolated into the future but we don’t put much more effort into it as opposed to doing nothing at all?\"\"\n\nOrd replies: \"\"That’s right, and it turns out to be quite hard to define business as usual. That’s the reason why, for my key estimate, that I make it… In some sense, it’s difficult to define estimates where they take into account whether or not people follow the advice that you’re giving; that introduces its own challenges. But at least that’s just what a probability normally means. It means that your best guess of the chance something happens, whereas a best guess that something happens conditional upon certain trends either staying at the same level or continuing on the same trajectory or something is just quite a bit more unclear as to what you’re even talking about.\"\"",
"title":"The probability that the long-run overall impact on humanity of human level machine intelligence will be Extremely bad (existential catastrophe)”, assuming Human Level Machine Intelligence will at some point exist.",
"description":"This is the mean. According to Beard et al., the question was \"4. Assume for the purpose of this question that such Human Level Machine Intelligence (HLMI) will at some point exist. How positive or negative would be overall impact on humanity, in the long run?",
"title":"Chance that AI, through “adversarial optimization against humans only”, will cause existential catastrophe, conditional on there not being “additional intervention by longtermists” (or perhaps “no intervention from longtermists”)",
"description":"Actual estimate: ~10%\n\nThis is my interpretation of some comments that may not have been meant to be taken very literally. I think he updated this in 2020 to ~15%, due to pessimism about discontinuous scenarios: https://www.lesswrong.com/posts/TdwpN484eTbPSvZkm/rohin-shah-on-reasons-for-ai-optimism?commentId=n577gwGB3vRpwkBmj Rohin also discusses his estimates here: https://futureoflife.org/2020/04/15/an-overview-of-technical-ai-alignment-in-2018-and-2019-with-buck-shlegeris-and-rohin-shah/",
"title":"Chance that AI, through “adversarial optimization against humans only”, will cause existential catastrophe, conditional on “discontinuous takeoff”",
"title":"Chance that we don't manage to survive that transition [to there being something that's more intelligent than humanity], being in charge of our future.",
"description":"Actual estimate: ~20%\n\nThis may have been specifically if the transition happens in the net 100 years; it's possible Ord would estimate we'd have a different chance if this transition happened at a later time.\n\"Basically, you can look at my [estimate that the existential risk from AI in the next 100 years is] 10% as, there’s about a 50% chance that we create something that’s more intelligent than humanity this century. And then there’s only an 80% chance that we manage to survive that transition, being in charge of our future. If you put that together, you get a 10% chance that’s the time where we lost control of the future in a negative way.\n\n[For people who would disagree, a question] is why would they think that we have much higher than an 80% chance of surviving this ‘passing this baton to these other entities’, but still retaining control of our future or making sure that they build a future that is excellent, surpassingly good by our own perspective? I think that the very people who are working on trying to actually make sure that artificial intelligence would be aligned with our values are finding it extremely difficult. They’re not that hopeful about it. So it seems hard to think there’s more than 80% chance, based on what we know, to get through that.",
"description":"Actual estimate: ~2%\n\nI give existential risk over the next century from nuclear war at about one in a thousand. I initially thought it would be higher than that. That’s actually something that while researching the book, thought was a lower risk than I had initially thought. And how I’d break it down is to something like a 5% chance of a full-scale nuclear war in the next century and a 2% chance that that would be the end of human potential.\" Ord discusses his reasoning more both in that interview and in The Precipice.",
"description":"Actual estimate: 0.001-0.01% (“in the range of 1 in 10,000 to 1 in 100,000”)\n\nI think that this is Oman’s estimate of the chance that extinction would occur if that black carbon scenario occurred, rather than an estimate that also takes into account the low probability that that black carbon scenario occurs. I.e., I think that this estimate was conditional on a particular type of nuclear war occurring. But I’m not sure about that, and the full context doesn’t make it much clearer.",
"description":"Arden Koehler: \"...do you have a guess at what degree of warming we would need to reach for the full-scale collapse of society, perhaps due to very, very widespread famine to have say a 10% chance of happening?\nMark Lynas: \"Oh, I think… You want to put me on the spot. I would say it has a 30 to 40% chance of happening at three degrees, and a 60% chance of happening at four degrees, and 90% at five degrees, and 97% at six degrees. [...] Maybe 10% at two degrees.",
"description":"Actual estimate: 30-40%\n\nArden Koehler: \"...do you have a guess at what degree of warming we would need to reach for the full-scale collapse of society, perhaps due to very, very widespread famine to have say a 10% chance of happening?\nMark Lynas: \"Oh, I think… You want to put me on the spot. I would say it has a 30 to 40% chance of happening at three degrees, and a 60% chance of happening at four degrees, and 90% at five degrees, and 97% at six degrees. [...] Maybe 10% at two degrees.",
"description":"Arden Koehler: \"...do you have a guess at what degree of warming we would need to reach for the full-scale collapse of society, perhaps due to very, very widespread famine to have say a 10% chance of happening?\nMark Lynas: \"Oh, I think… You want to put me on the spot. I would say it has a 30 to 40% chance of happening at three degrees, and a 60% chance of happening at four degrees, and 90% at five degrees, and 97% at six degrees. [...] Maybe 10% at two degrees.",
"description":"Arden Koehler: \"...do you have a guess at what degree of warming we would need to reach for the full-scale collapse of society, perhaps due to very, very widespread famine to have say a 10% chance of happening?\nMark Lynas: \"Oh, I think… You want to put me on the spot. I would say it has a 30 to 40% chance of happening at three degrees, and a 60% chance of happening at four degrees, and 90% at five degrees, and 97% at six degrees. [...] Maybe 10% at two degrees.",
"description":"Arden Koehler: \"...do you have a guess at what degree of warming we would need to reach for the full-scale collapse of society, perhaps due to very, very widespread famine to have say a 10% chance of happening?\nMark Lynas: \"Oh, I think… You want to put me on the spot. I would say it has a 30 to 40% chance of happening at three degrees, and a 60% chance of happening at four degrees, and 90% at five degrees, and 97% at six degrees. [...] Maybe 10% at two degrees.",
"title":"A world totalitarian government will emerge during the next one thousand years and last for a thousand years or more, conditional on genetic screening for personality traits becom[ing] cheap and accurate, but the principle of reproductive freedom prevail[ing]",
"title":"A world totalitarian government will emerge during the next one thousand years and last for a thousand years or more, conditional on genetic screening for personality traits becom[ing] cheap and accurate and extensive government regulation",
"title":"A world totalitarian government will emerge during the next one thousand years and last for a thousand years or more, conditional on the number of independent countries on earth [not decreasing] during the next thousand years",
"title":"A world totalitarian government will emerge during the next one thousand years and last for a thousand years or more, conditional on the number of independent countries on earth [falling to 1] during the next thousand years",
"description":"This is the median. The report about these estimates also plots the results for each question “with individual response distributions visible” in Appendix A.",
"description":"This is the median. The report about these estimates also plots the results for each question “with individual response distributions visible” in Appendix A. Interestingly, this is the same as the estimate from this source of the chance of human as a result of superintelligent AI by 2100.",
"description":"Actual estimate: ~10%\n\nSo, decent chance– I think I put a reasonable probability, like 10% probability, on the hard-mode MIRI version of the world being true. In which case, I think there’s probably nothing we can do.",
"description":"Actual estimate: ~30%\n\nThere’s some chance that the first thing we try just works and we don’t even need to solve any sort of alignment problem. It might just be fine. This is not implausible to me. Maybe that’s 30% or something.",
"description":"Actual estimate: ~30%\n\nI haven’t actually written down these numbers since I last changed my mind about a lot of the inputs to them, so maybe I’m being really dumb. I guess, it feels to me that in fast takeoff worlds, we are very sad unless we have competitive alignment techniques, and so then we’re just only okay if we have these competitive alignment techniques. I guess I would say that I’m something like 30% on us having good competitive alignment techniques by the time that it’s important, which incidentally is higher than Rohin I think. [...] So I’m like 30% that we can just solve the AI alignment problem in this excellent way, such that anyone who wants to can have a little extra cost and then make AI systems that are aligned. I feel like in worlds where we did that, it’s pretty likely that things are reasonably okay.",
"description":"Actual estimate: ~50%\n\nBasically, you can look at my [estimate that the existential risk from AI in the next 100 years is] 10% as, there’s about a 50% chance that we create something that’s more intelligent than humanity this century. And then there’s only an 80% chance that we manage to survive that transition, being in charge of our future. If you put that together, you get a 10% chance that’s the time where we lost control of the future in a negative way.\n\nToby Ord: With that number, I’ve spent a lot of time thinking about this. Actually, my first degree was in computer science, and I’ve been involved in artificial intelligence for a long time, although it’s not what I did my PhD on. But, if you ask the typical AI expert’s view of the chance that we develop smarter than human AGI, artificial general intelligence, this century is about 50%. If you survey the public, which has been done, it’s about 50%. So, my 50% is both based on the information I know actually about what’s going on in AI, and also is in line with all of the relevant outside views. It feels difficult to have a wildly different number on that. The onus would be on the other person.",
"title":"A design very close to CEV will be implemented in humanity's AGI, conditional on AGI being built (excluding other value-learning approaches and other machine-ethics proposals)",
"title":"A design very close to CEV will be implemented in humanity's AGI, conditional on AGI being built (excluding other value-learning approaches and other machine-ethics proposals)",
"description":"This is the median. The report about these estimates also plots the results for each question “with individual response distributions visible” in Appendix A.",
"description":"This is the median. The report about these estimates also plots the results for each question “with individual response distributions visible” in Appendix A.",
"description":"This is the median. The report about these estimates also plots the results for each question “with individual response distributions visible” in Appendix A.",
"description":"This is the median. The report about these estimates also plots the results for each question “with individual response distributions visible” in Appendix A.",
"description":"This is the median. The report about these estimates also plots the results for each question “with individual response distributions visible” in Appendix A.",
"description":"This is the median. The report about these estimates also plots the results for each question “with individual response distributions visible” in Appendix A.",
"description":"This is the median. The report about these estimates also plots the results for each question “with individual response distributions visible” in Appendix A.",
"description":"This is the median. The report about these estimates also plots the results for each question “with individual response distributions visible” in Appendix A.",
"description":"This is the median. The report about these estimates also plots the results for each question “with individual response distributions visible” in Appendix A.",
"description":"This is the median. The report about these estimates also plots the results for each question “with individual response distributions visible” in Appendix A.",
"description":"This is the median. The report about these estimates also plots the results for each question “with individual response distributions visible” in Appendix A.",
"description":"This is the median. The report about these estimates also plots the results for each question “with individual response distributions visible” in Appendix A.",
"description":"Actual estimate: ~5%\n\nI give existential risk over the next century from nuclear war at about one in a thousand. I initially thought it would be higher than that. That’s actually something that while researching the book, thought was a lower risk than I had initially thought. And how I’d break it down is to something like a 5% chance of a full-scale nuclear war in the next century and a 2% chance that that would be the end of human potential.\" Ord discusses his reasoning more both in that interview and in The Precipice.",
"description":"Actual estimate: 1.10%\n\nIn this post, I get a rough sense of how probable a nuclear war might be by looking at historical evidence, the views of experts, and predictions made by forecasters. I find that, if we aggregate those perspectives, there’s about a 1.1% chance of nuclear war each year, and that the chances of a nuclear war between the US and Russia, in particular, are around 0.38% per year.\" This is not presented as Luisa's own credence; this may not be the number she herself would give. Readers may also be interested in the estimates implied by each of the perspectives Luisa aggregates.",
"description":"Actual estimate: 0.38%\n\nIn this post, I get a rough sense of how probable a nuclear war might be by looking at historical evidence, the views of experts, and predictions made by forecasters. I find that, if we aggregate those perspectives, there’s about a 1.1% chance of nuclear war each year, and that the chances of a nuclear war between the US and Russia, in particular, are around 0.38% per year.\" This is not presented as Luisa's own credence; this may not be the number she herself would give. Readers may also be interested in the estimates implied by each of the perspectives Luisa aggregates.",
