He repeats the “A recent survey showed that half of AI researchers give AI at least 10% chance of causing human extinction” claim, but the survey in question finds that this is true of respondents, who were 17% of those the survey was sent to. So this claim is misleading as phrased.
The survey seems to have taken reasonable steps to account for responder-bias, and IIRC at least I couldn’t tell any obvious direction in which respondents were biased. Katja has written some about this here: https://twitter.com/KatjaGrace/status/1643342692905254912
Response rates still seem good to mention when mentioning the survey, but I don’t currently believe that getting a survey with a higher response rate would change the results. Might be worth a bet?
but the survey in question finds that this is true of respondents, who were 17% of those the survey was sent to
Only 20% of the respondents gave a response to that particular question (thanks to Denreik for drawing my attention to that fact, which I verified). Of the initially contacted 4271 researchers, 738 gave responses (17% of 4271), and 149 (20% of 738) gave a probability for the “extremely bad” outcome on the non-trick version of the question (without the “human inability to control” part).
He repeats the “A recent survey showed that half of AI researchers give AI at least 10% chance of causing human extinction” claim, but the survey in question finds that this is true of respondents, who were 17% of those the survey was sent to. So this claim is misleading as phrased.
The survey seems to have taken reasonable steps to account for responder-bias, and IIRC at least I couldn’t tell any obvious direction in which respondents were biased. Katja has written some about this here: https://twitter.com/KatjaGrace/status/1643342692905254912
Response rates still seem good to mention when mentioning the survey, but I don’t currently believe that getting a survey with a higher response rate would change the results. Might be worth a bet?
Fair enough, didn’t know about those steps. That does update me towards this being representative.
For reference, https://aiguide.substack.com/p/do-half-of-ai-researchers-believe is a recent blog post about the same claim. After fact-checking, the author is “not convinced” by the survey.
Additionally, here https://twitter.com/tdietterich/status/1651096428935254016 according to one survey participant, “it was obvious from question formulation that they were not interested in an unbiased answer.”
Only 20% of the respondents gave a response to that particular question (thanks to Denreik for drawing my attention to that fact, which I verified). Of the initially contacted 4271 researchers, 738 gave responses (17% of 4271), and 149 (20% of 738) gave a probability for the “extremely bad” outcome on the non-trick version of the question (without the “human inability to control” part).