Comment Permalink

Thanks! We've edited the text to include both the FAOL estimate that you mention, and the combined estimate that Vasco mentions in the other reply. (The changes might not show up on site immediately, but will soon.) To the extent that people think FAOL will take longer than HLMI because of obstacles to AI doing jobs that don't come from it not being generally capable enough, I think the estimate for HLMI is closer to an estimate of when we'll have human-level AI than the estimate for FAOL. But I don't know if that's the right interpretation, and you're definitely right that it's fairer to include the whole picture. I agree that there's some tension between us saying "experts think human-level AI is likely to arrive in your lifetime" and this survey result, but I do also still think that that sentence is true on the whole, so we'll think about whether to add more detail about that.

When do experts think human-level AI will be created?

by Vishakha Agrawal, Algon

Jan 23 min read 9

33

AI safetyExistential riskForecastingAI forecastingArtificial intelligence

Frontpage

This is a linkpost for https://aisafety.info/questions/5633/When-do-experts-think-human-level-AI-will-be-created

This is an article in the featured articles series from AISafety.info. AISafety.info writes AI safety intro content. We'd appreciate any feedback.

The most up-to-date version of this article is on our website, along with 300+ other articles on AI existential safety.

On the whole, experts think human-level AI is likely to arrive in your lifetime.

It’s hard to precisely predict the amount of time until human-level AI.^[1] Approaches include aggregate predictions, individual predictions, and detailed modeling.

Aggregate predictions:

AI Impacts’ 2023 survey of machine learning researchers produced an aggregate forecast of 50% by 2047 (compared to 2059 in their 2022 survey).
As of June 2024, Metaculus^[2] has a median forecast of 2031 for “the first general AI system” and a median forecast of 2027 for “weakly general AI”. Both these timeline forecasts have been shortening over time.
This website combines predictions from different forecasting platforms into a single (possibly inconsistent) timeline of events.
In January 2023, Samotsvety’s forecasters estimated 50% probability of AGI by 2041 with a standard deviation of 9 years.

Individual predictions:

In a 2023 discussion, Daniel Kokotajlo, Ajeya Cotra and Ege Erdil shared their timelines to Transformative AI. Their medians were 2027, 2036 and 2073 respectively.
Paul Christiano, head of the US AI Safety Institute, estimated in 2023 that there was a 30% chance of transformative AI by 2033.
Yoshua Bengio, Turing Award winner, estimated “a 95% confidence interval for the time horizon of superhuman intelligence at 5 to 20 years” in 2023.
Geoffrey Hinton, the most cited AI scientist, also predicted 5-20 years in 2023, but his confidence is lower.
Shane Legg, co-founder of DeepMind, estimated a probability of 80% within 13 years (before 2037) in 2023.
Yann LeCun, Chief AI Scientist at Meta, thinks reaching human-level AI “will take several years if not a decade. [...] But I think the distribution has a long tail: it could take much longer than that.”
Leopold Aschenbrenner, an AI researcher formerly at OpenAI, predicted in 2024 that AGI happening around 2027 was strikingly plausible.
Connor Leahy, CEO of Conjecture, gave a ballpark prediction in 2022 of a 50% chance of AGI by 2030, 99% by 2100. A 2023 survey of employees at Conjecture found that all of the respondents expected AGI before 2035.
Holden Karnofsky, co-founder of GiveWell, estimated in 2021 that there was “more than a 10% chance we'll see transformative AI within 15 years (by 2036); a ~50% chance we'll see it within 40 years (by 2060); and a ~⅔ chance we'll see it this century (by 2100).”
Andrew Critch, an AI researcher, estimated in 2024 that there was a 45% chance of AGI by the end of 2026.

Models:

A report by Ajeya Cotra for Open Philanthropy estimated the arrival of transformative AI (TAI) based on “biological anchors”.^[3] In the 2020 version of the report, she predicted a 50% chance by 2050, but in light of AI developments over the next two years, she updated her estimate in 2022 to predict a 50% chance by 2040, a decade sooner.
Tom Davidson's take-off speeds model somewhat extends and supersedes Ajeya Cotra's bio-anchors framework, and offers an interactive tool for estimating timelines based on various parameters. The scenarios it offers as presets predict 100% automation in 2027 (aggressive), 2040 (best guess), and never (conservative).
Matthew Barnett created a model based on the “direct approach” of extrapolating training loss that as of Q1 2025 outputs a median estimate of transformative AI around 2033.

These forecasts are speculative,^[4] depend on various assumptions, predict different things (e.g., transformative versus human-level AI), and are subject to selection bias both in the choice of surveys and the choice of participants in each survey.^[5] However, they broadly agree that human-level AI is plausible within the lifetimes of most people alive today. What’s more, these forecasts generally seem to have been getting shorter over time.^[6]

33 Reactions

More posts like this

Comments9

Sorted by

New & upvoted

Click to highlight new comments since: Today at 9:44 PM

YarrowApr 235

Metaculus accepts predictions from just anybody, so Metaculus is not an aggregator of expert predictions. It’s not even a prediction market.

AlgonMay 21

Metaculus is not an aggregator of expert predictions

Yeah, that's a good point. TY! It would be wrong to claim that Metaculus predictions are somehow "expert predictions". We'll change the article to make it clearer that we're not claiming that.

That said, we don't use the word prediction market. And AFAICT, Metaculus has a favourable track record compared to actual prediction markets. And probably to other mechanisms for aggregating info, too. So I think there's value to referencing them.

So perhaps the following, maybe as a footnote?

While Metaculus allows non-domain-experts, the bulk of their forecasters, to make predictions, Metaculus' track record compares favourably to that of many AI experts and superforecasters. Hence the inclusion on this list.

What do you think?

YarrowMay 41

I think if you don't note in the body of the text, rather than just in a footnote, that just anybody can predict anything on Metaculus, then this will inevitably be misleading to anyone who doesn't already know what Metaculus is, since you imply in the post that it's an aggregator of expert predictions when you claim:

On the whole, experts think human-level AI is likely to arrive in your lifetime.

And then go on to list Metaculus as support for this claim. That implies Metaculus is aggregator of expert predictions.

Also, you include Metaculus in a long list of expert predictions without noting that it's different from the other items on the list, which reinforces the implication that it's an aggregator of expert predictions.

I think you should also explain what Samotsvety is in the body of the text and what its forecasters' credentials are.

Invoking "experts" and then using the term this loosely feels misleading.

I think it also bears mentioning the strange feature of the 2023 AI Impacts survey where there's a 69-year gap between the AI experts' prediction of "high-level machine intelligence" and "full automation of labour" (50% chance by 2116). This is such an important (and weird, and confusing) fact about the survey that I think it should be mentioned anytime that survey is brought up.

This is especially relevant since you say:

On the whole, experts think human-level AI is likely to arrive in your lifetime.

And if you think human-level AI means full automation of labour rather than high-level machine intelligence, then a 50% chance by 2116 (91 years from now) is not within the current life expectancy of most adults alive today or even most teenagers.

GregorBlnJan 816

Thank you very much for the review and aggregation of all these forecasts! Very nice!

I just have one point to add:

As the first aggregate prediction, you mention the AI Impacts’ 2023 survey of machine learning researchers. Your post gives the impression that it produced an aggregate forecast of 50% by 2047 for human-level AI. I think this is at least imprecise, if not incorrect.

AI Impacts asked about the timing of human-level performance by asking some participants about how soon they expect “high-level machine intelligence” (HLMI) and asking others about how soon they expect “full automation of labor” (FAOL). The resulting aggregate forecast gave a 50% chance of HLMI by 2047 and a 50% chance of FAOL by 2116. In your post, you ignore that AI Impacts uses two different concepts for human-level AI and just report the aggregate forecast for HLMI under the headline of human-level AI.

I think this is unfortunate because this difference matters. One of your main points is that you claim that experts think human-level AI is likely to arrive in your lifetime. However, most of us will probably not be alive in 2116.

Vasco Grilo🔸Jan 85

Great point, Gregor! Tom Adamczewski has done an analysis which combines the answers to the questions about tasks and occupations. Here is the mainline graph.

Tom aggregates the results from the different questions in the most agnostic way possible, which I think is the best one can do.

I achieve this by simply including answers to both questions prior to aggregation, i.e. no special form of aggregation is used for aggregating tasks (HLMI) and occupations (FAOL). Since more respondents were asked about tasks than occupations, I achieve equal weight by resampling from the occupations (FAOL) responses.

Here is how Tom suggests people describe the results.

Experts were asked when it will be feasible to automate all tasks or occupations. The median expert thinks this is 20% likely by 2048, and 80% likely by 2103. There was substantial disagreement among experts. For automation by 2048, the middle half of experts assigned it a probability between 1% and a 60% (meaning ¼ assigned it a chance lower than 1%, and ¼ gave a chance higher than 60%). For automation by 2103, the central half of experts forecasts ranged from a 25% chance to a 100% chance.²

YarrowMay 43

I find this completely inscrutable. I'm not saying there's anything wrong with it in terms of accuracy, it's just way too in the weeds of the statistics for me to decipher what's going on.

For example, I don't know what a "middle half" or a "central half" is. I looked them up and now I know they are statistics terms, but it would be a lot of work for me to try to figure out what that quoted paragraph is trying to say.

Is AI Impacts going to run this survey again soon? Maybe they can phrase the questions differently in a new survey to avoid this level of confusion between different levels of AI capabilities.

Vasco Grilo🔸May 42

Hi Yarrow,

Assuming there were only 4 experts predicting full automation by 2048 with 10 %, 20 %, 30 %, and 40 % chance, the middle half of experts would be predicting full automation by 2048 with 20 % to 30 % chance. If there were 1 k experts each predicting the probability of full automation by multiple dates, one could infer 1 k predictions for the probability of full automation for any given date, order such preductions by ascending order, and then report the 25th and 75th percentile predictions as the lower and upper bound of the middle half of predictions.

StevenKaasJan 82

Alex (Αλέξανδρος)Jan 71

This timeline is very interesting, and it leads to a question beyond just when: what happens when AI is capable of independent thought and goal-setting? We're focused on mitigating risks, which is crucial. However, we should also consider the moral implications of creating beings capable of:

a) thinking independently, beyond merely fulfilling human-designed requests

b) setting their own goals

How can we ensure a future where humans and advanced AI can co-exist, minimizing suffering for both and maximising the potential benefits of collaboration – from scientific discovery to solving global challenges?