AI Impacts survey: 2047 for HLMI, up from 2060, and what it asked

The AI Impacts expert survey gives a 50% chance of “high-level machine intelligence” (HLMI) by 2047, according to its aggregate forecast of 1,714 responses^{1, 2}. That is 13 years earlier than the 2022 survey, which AI Impacts now restates as 2060^{1, 2}.

The 2047 figure is not a date for AGI in the loose sense, and it is not the only date the survey produced^{1, 2, 3, 4}. This article uses only the 2022 AI Impacts survey and its 2023 successor^{1, 2, 3, 4}.

What the 2023 survey is

The 2023 Expert Survey on Progress in AI was published as a preprint in January 2024^{1, 2, 4}. The survey received 2,778 responses from researchers who had published in six top AI venues^{1, 2, 4}. The main survey group was run 15–24 October 2023, a 15% response rate^{1, 2, 4}.

The authors say it ran after the launch of ChatGPT and GPT-4⁴. That detail comes from the preprint alone⁴.

The 2022 survey drew 738 responses from authors at NeurIPS and ICML, a 17% response rate, according to AI Impacts^{1, 3}.

What “HLMI” asks

Both surveys used the same wording for HLMI^{1, 3, 4}. It means unaided machines accomplishing every task better and more cheaply than human workers^{1, 3, 4}. Tasks where being human is intrinsically advantageous, such as jury membership, are ignored^{1, 3, 4}.

The question is about feasibility, not adoption^{1, 3, 4}. Respondents were also asked to assume human scientific activity continues without major negative disruption^{1, 3, 4}.

Two dates, not one

The 2023 survey asked a second, occupation-based question: “full automation of labor” (FAOL), meaning all occupations being fully automatable^{1, 2}. Its aggregate forecast gives a 50% chance by 2116, based on 774 responses, versus 2164 in 2022^{1, 2}.

The 2047 figure is therefore the task-based HLMI date, not the occupation-based one^{1, 2}. The FAOL date is about 69 years later^{1, 2}. Participants asked about FAOL tended to give significantly longer timelines than those asked about HLMI^{1, 2}.

The table below sets the two surveys side by side on sample, wording and forecast years.

Item	2022 survey	2023 survey	Source
Responses and venues	738 responses, NeurIPS and ICML authors (17% response rate)	2,778 responses, authors from six top AI venues (15% response rate); preprint	1, 2, 3, 4
HLMI definition (task-based)	Unaided machines doing every task better and more cheaply than human workers; feasibility, not adoption	Same wording	1, 3, 4
HLMI: 50% chance by	2059 (37 years out); restated as 2060 in the 2023 wiki	2047 (1,714 responses)	1, 2, 3
HLMI: 10% chance by	Not in the claims given	2027	2, 4
FAOL (all occupations fully automatable): 50% by	2164	2116 (774 responses)	1, 2
Why 2047 is not the FAOL date	Different question	FAOL is about 69 years after the HLMI figure; those asked FAOL tended to give longer timelines	1, 2
Method	Not detailed in the claims given	Three year-probability pairs per person, gamma curve fitted, average curve; framing varied by half of respondents. Authors say framing can greatly influence answers	1, 4
Context and sample effect	Run before ChatGPT and GPT-4 launched is not stated; not in claims	Run 15–24 October 2023, after ChatGPT and GPT-4; authors say the wider sample did not appear to notably affect opinion	1, 2, 4
Task questions identical in both surveys (32)	Baseline	21 moved earlier, 11 later; average 50th-percentile year 1.0 years earlier	4

How the number is built

The headline year is not a simple median of what people typed^{1, 4}. Each respondent gave three year-and-probability pairs, and a gamma distribution (a flexible curve for waiting-time data) was fitted to each respondent’s answers^{1, 4}.

The preprint’s abstract also reports a 10% chance of HLMI by 2027 alongside the 50% chance by 2047, again conditional on science continuing undisrupted^{2, 4}. That figure rests on the preprint abstract only^{2, 4}.

Framing was varied. Half of respondents got a fixed-years question, asking for a probability by a given year^{1, 4}. Half got a fixed-probabilities question, asking for a year at a given probability^{1, 4}. The authors note that framing can greatly influence responses^{1, 4}. They also note that AI researchers are experts in AI research, not in forecasting^{1, 4}.

Why 2059 and 2060

The 2022 survey page reported a 50% chance of HLMI at 37 years out, which is 2059^{1, 3}. The 2023 material restates the 2022 figure as 2060^{1, 3}. The sources we have do not explain the one-year difference, and only AI Impacts material backs it^{1, 3}.

AI Impacts calls the shift 13 years, which uses 2060 as the 2022 figure^{1, 2}. The 2022 survey page itself gave 2059^{1, 3}. The sources do not state a shift measured against 2059^{1, 3}. Either way, the direction is the same: the 2023 forecast is earlier^{1, 2}.

What changed between the surveys

The sample and the world both changed. The 2023 survey had around four times as many participants as 2022 because it expanded from two venues to six⁴. The authors say this did not appear to have a notable effect on opinion⁴. That is their own assessment, and the preprint is the only source we have for it⁴.

The task-level results suggest a smaller shift than the headline^{1, 2, 4}. Across the 32 task questions identical in 2022 and 2023, the aggregate prediction moved earlier for 21 and later for 11⁴. On average the 50th-percentile year shifted 1.0 years earlier⁴. This also comes from the preprint alone⁴.

The claims we hold do not explain why the HLMI date moved so much further than the typical task date, nor why researchers revised their views^{1, 2, 4}.

How far to trust the 2047 figure

• It is a forecast. The aggregate projects a 50% chance of high-level machine intelligence by 2047^{1, 2}. It is not a prediction that machines will do every task by then^{1, 2, 3, 4}.
• It is sensitive to wording. The same survey gives a date about 69 years later when asked about occupations instead of tasks^{1, 2}.
• It is sensitive to framing. The authors say the way a question is posed can greatly influence answers^{1, 4}.
• It reflects expertise in AI, not forecasting. The authors say so themselves^{1, 4}.
• The comparison across years is imperfect. The venues, the sample size and the 2022 baseline year (2059 or 2060) all differ or are unclear^{1, 3, 4}.

The most defensible reading is narrow. Among published AI researchers surveyed in October 2023, the aggregate forecast for a task-based, cost-conditional milestone was 2047, earlier than in 2022^{1, 2, 3, 4}.

This article was researched and drafted with the help of AI tools, then fact-checked against its sources and edited by a human before publication. Every source is linked below. If you spot an error, please tell us and we will correct it.

Sources

1. 2023 Expert Survey on Progress in AI [AI Impacts Wiki] — wiki.aiimpacts.org, accessed 2026-10-03
2. [2401.02843] Thousands of AI Authors on the Future of AI — arxiv.org (preprint, not peer reviewed), accessed 2026-10-03
3. 2022 Expert Survey on Progress in AI – AI Impacts — aiimpacts.org, accessed 2026-10-03
4. Thousands of AI Authors on the Future of AI — arxiv.org (preprint, not peer reviewed), accessed 2026-10-03