A survey of 2,778 AI researchers: 38.4% support faster development and 34.7% support slower development, AI development's pace will keep accelerating, and more
The “2023 Expert Survey on Progress in AI” shows that the scientific community has no consensus on the risks and opportunities of AI …
The DecoderMatthias Bastian
Context & Ripple Effects
Earlier coverage framed rapid AI progress as a choice between faster social adaptation and an enforceable slowdown; this researcher survey shows that divide also runs through the technical community itself, rather than separating experts from critics. The earlier adaptation-versus-slowdown debate therefore lacks a clear expert-side resolution.
It also sits alongside a longer-running tension in which support for AI development has been coupled with demands for careful management. Public support paired with calls for careful AI management makes the split among researchers relevant to how legitimacy for future governance is argued.
First-order effects
The survey gives neither faster-development advocates nor advocates of restraint a decisive claim to expert consensus, with the two camps separated by only a few percentage points.
Expectations of continuing acceleration make the disagreement more consequential for researchers, labs, and policymakers: the dispute concerns the pace of a process respondents expect to intensify, not a static technology agenda.
Second-order effects
Competing policy and industry camps can cite the same expert community to support opposing priorities, shifting debate toward concrete governance choices rather than appeals to a unified scientific view.
If this split persists as development accelerates, AI governance is likely to be built around managing durable disagreement and dual-use trade-offs rather than waiting for expert unanimity.
The structural challenge is aligning deployment incentives with safety and public legitimacy while views on the appropriate pace of progress remain divided.
The trend: AI industrialization is advancing faster than consensus on its acceptable pace, making governance a continuing coordination problem rather than a one-time expert verdict.
Study by @KatjaGrace et al of nearly 3,000 AI researchers finds that the estimated AI timelines have shrunk. And many researchers think that the risk of extinction from AI is substantial. However, there's a lot of disagreement and also some inconsistency between questions. [image…
We just ran probably the biggest survey of AI researchers ever! 2778 participants from six top AI venues answered questions from fourteen topics regarding the future of AI. Preprint: https://aiimpacts.org/... Six interesting things in pictures: [image]
@primalpoly ... Constant nagging and finger pointing makes what may be well reasoned stances, entirely unpalatable. You can do better. Excluding everyone else and creating an “us vs them” dichotomy is only going to lead to you being treated as a pariah.
.@KatjaGrace et al asked 2,500+ AI researchers about risks & expectations Thoughtful survey design - comparing to 2016 survey, randomizing participants into alternative wordings to check framing effects, and getting more in-depth info from smaller samples [image]
New study suggests 40+% of AI researchers have at least “substantial concern” that labour automation over the next 30 years will leave most people economically powerless. [image]
This comparison on 2022 and 2023 shows that, but if you go even further back to the original analysis in 2016 it gets more fun. https://arxiv.org/... [image]
Great work, but I already see accounts freaking about mean P(doom) 16.2% ... i) the median P(doom) is 5% and ii) mean P(doom) barely changed from 15.7% in 2022
📢📊 Important new survey results from @AIImpacts, surveying 2,700+ AI researchers. Some key points: ⚠️ The median respondent believes there's a 5% or greater chance of AI causing human extinction; one-third to half of respondents believe there's over a 10% chance. 😨 A majority...
Many people would like to see a truce between Yann LeCun and myself. I am among them. The world would benefit from civil, moderated debates between us, and I really enjoyed the 2017 debate we had together. At bottom I will explain the tiny action that could lead to a truce.... [i…
I'm excited to dig into the results of the survey, but I'll start by shouting out the really excellent graphic design in the blogpost. These graphs are a really nice way of showing the highlighted results. https://blog.aiimpacts.org/...
Forecasting future events ⏩ is crazy difficult Instead of mastering calibration techniques 👨🔬, waiting and gathering more info ⏰ can be more impactful (though one can't always wait) The space of possible futures is enormous. Time can narrow it down considerably: ⬛️ ➡️ ▪️.
Seems clear to me these respondents aren't thinking in terms of a coming several centuries long innovation pause as per: https://www.overcomingbias.com/ ...
I expected a new survey of AI researchers would find their forecasts for AGI had moved sooner. But I didn't expect a 13 year drop in just one year. (After ~no movement the last 6 years.) Forecasts for “high level machine intelligence”: 2023 survey: 2047 2022 survey: 2060 2016...
This is super interesting on AI risk. Think it would be good if other fields made more attempts to conduct scientific surveys of expert opinion. (Disclosure: I did a very small bit of unpaid consulting on this survey.)
Whole survey is wild throughout. The answers, they do not make sense, nor do they together live in a remotely consistent world. I should be used to it, but I'm still boggled.
Survey of AI researchers expects ability to do ALL human jobs faster and cheaper by 2047, but ‘full automation of labor’ only at ~2113, retail salesperson by ~2095. These seem like... completely different and incompatible predictions? We're going to sit on AI for generations?
Is Yann LeCun @ylecun a highly cited machine learning researcher? Yes. Is he also an arrogant, bad-faith bully who lies about his critics, avoids serious debate, dismisses AI risks, & prioritizes his financial interests in @meta over humanity's survival? Also, yes.
AI Impacts conducted a survey of 2,778 AI researchers who had published peer-reviewed research in the prior year in six top AI venues (NeurIPS, ICML, ICLR, AAAI, IJCAI, JMLR). The survey took place in the fall of 2023 after the launch of ChatGPT and other top models, Future of...…
@primalpoly ... Geoffrey, liking/disliking a person aside, it's not hard to see that pointing out limitations and flaws about some approach X is NOT necessarily seriously scholarly work, nor original thought that needs quotation. E.g., I have long held that LLMs cannot perform De…
New @AIImpacts survey from my friend @KatjaGrace and her team! Read the whole thing, but my 2 big takeaways: 1) Their AI timelines have shortened. The this year's survey (red line) is >5 years sooner than last years for almost any probability. [image]
Interesting study. Now, I am very sceptical of the skill of experts in topic X to predict future of X (not same skill), but just seeing what people inside a field think is also informative. If practitioners overall become more optimistic on performance, that is interesting.
The fact that there is broad-based practitioner concern about AI safety risk (although with a lot of variation from person to person) and a quickening of AI timelines is significant. You'll still get the occasional media report framing these as fringe positions. But they're not.
The median researcher thinks AGI is 5% likely to be a catastrophic risk to humanity. For context, nuclear plants aim for 0.005% chance of major radiation leak & airlines aim for things like 0.5 fatalities per million takeoffs. With good risk management, we can get #AI there too.
“surveyed 2,778 researchers who had published in top-tier artificial intelligence (AI) venues, ... chance of all human occupations becoming fully automatable was forecast to reach 10% by 2037, and 50% as late as 2116 (compared to 2164 in the 2022 survey)”