---
title: "What Do Experts Say About AGI Timelines?"
slug: expert-surveys-on-agi-timelines
description: "What the main expert surveys on AGI timelines report, from Müller and Bostrom (2012-13) to AI Impacts 2023 and AAAI 2025, and how to read them."
date: 2026-10-26
last_reviewed: 2026-10-26
status: published
type: knowledge-base
image: expert-surveys-on-agi-timelines
image_alt: "Three researchers in a dim server room study a glowing monitor of charts, one pointing at the screen."
image_caption: "Illustration of researchers examining data. It does not show the authors of any survey discussed here. AI-generated image."
tags: [agi, forecasting, expert surveys, timelines, ai impacts]
---

# What Do Experts Say About AGI Timelines?

**In short:** the largest published survey, of 2,778 AI researchers in October 2023, gave a 50% chance that machines outperform humans at every task by 2047 and a 10% chance by 2027. Earlier surveys put the 50% point around 2040 (2012-13) and 2061 (2016). These are aggregated opinions, not measurements, and they move a lot with how the question is worded.

Artificial general intelligence (AGI) has no single agreed definition (see [what is AGI](./what-is-agi-definitions-compared.md) and [definitions and operational criteria for AGI](../vol05_agi/AGI-DEF-001_definitions_and_operational_criteria_for_agi.md)). Surveys therefore ask about operational stand-ins such as "high-level machine intelligence" (HLMI). Every figure below is graded Forecast: it is what respondents said, not what will happen. The encyclopedia's method article is [AGI forecasting](../vol05_agi/AGI-TIME-001_agi_forecasting.md), and the [glossary](/glossary.html) defines the terms.

## Quick reference

| Survey | When asked | Respondents | Question asked about | 50% date | Grade |
|---|---|---|---|---|---|
| Müller and Bostrom | 2012-2013 | 170 of 549 invited (31%), four groups | HLMI: "most human professions at least as well as a typical human" | 2040 (median) | Forecast |
| Grace et al. (2018) | 2016 | 352 of 1,634 NIPS/ICML 2015 authors (21%) | HLMI: "every task better and more cheaply than human workers" | 45 years from 2016 | Forecast |
| Grace et al. (2024) | Oct 2023 | 2,778 authors at six top venues (15%) | Same HLMI wording | 2047 (10%: 2027) | Forecast |
| Grace et al. (2024) | Oct 2023 | Same survey | Full automation of all occupations (FAOL) | 2116 (10%: 2037) | Forecast |
| AAAI Presidential Panel | 2024-25, published Mar 2025 | 475 respondents | Whether scaling current approaches yields AGI | No date asked | Forecast |

## What did the 2023 AI Impacts survey find?

The survey behind "Thousands of AI Authors on the Future of AI" (Katja Grace and six co-authors) was run by AI Impacts from 11 to 24 October 2023. It invited 18,459 working addresses of people who had published at NeurIPS, ICML, ICLR, AAAI, IJCAI or JMLR, and received 2,778 responses, a 15% response rate. The paper was posted on arXiv in January 2024 and published in the *Journal of Artificial Intelligence Research* (volume 84) in October 2025.

HLMI was defined as being "achieved when unaided machines can accomplish every task better and more cheaply than human workers," with the instruction to "think feasibility, not adoption." The aggregate answer, "if science continues undisrupted," was a 10% chance by 2027 and 50% by 2047. The 50% date was 13 years earlier than the authors' 2022 survey (2060). Between the 2016 and 2022 surveys it had moved only one year, from 2061 to 2060.

The same respondents were asked when all occupations would be fully automatable. That gave 10% by 2037 and 50% by 2116, down from 2164 in 2022. The authors write that, insofar as HLMI and full automation of labor describe the same event, the 69-year gap "would seem to be a framing effect."

## What did earlier surveys say?

**Müller and Bostrom (2012-13).** Vincent C. Müller and Nick Bostrom (see [Nick Bostrom](../vol10_people_organizations/BIO-BOSTROM-001_nick_bostrom.md)) sent a short questionnaire to four groups: participants of a 2011 philosophy-and-theory-of-AI conference, participants of two artificial general intelligence conferences in Oxford in December 2012, members of the Greek Association for Artificial Intelligence, and the 100 most-cited AI authors. 170 of 549 responded. The median answers gave HLMI a 10% chance by 2022, 50% by 2040 and 90% by 2075. Respondents expected superintelligence "in less than 30 years thereafter," and put about a one-in-three chance on the outcome being "bad" or "extremely bad." The groups differed: the most-cited authors' median 50% date was 2050, and only 29 of those 100 replied. The 10% point (2022) has passed without HLMI.

**Grace et al. (2018).** In a 2016 survey of authors from NIPS and ICML 2015, the aggregate gave a 50% chance of HLMI within 45 years and 10% within 9 years. Asian respondents expected HLMI in 30 years, North Americans in 74.

## What do the most recent surveys show?

The AAAI 2025 Presidential Panel on the Future of AI Research, published in March 2025, surveyed 475 members of the AI community (about 20% students; 67% academic; 53% North America). It did not ask for a date. Instead, 76% said that "scaling up current AI approaches" to yield AGI is "unlikely" or "very unlikely" to succeed, 77% preferred building systems with an acceptable risk-benefit profile over the direct pursuit of AGI, and 70% opposed halting AGI research until full safety mechanisms exist.

That result does not contradict the 2047 figure. One asks whether today's methods, scaled up, are enough; the other asks when machines will outperform humans by any route. The scaling question is covered in [AI scaling laws explained](./ai-scaling-laws-explained.md).

AI Impacts keeps a [list of AI timeline surveys](https://wiki.aiimpacts.org/ai_timelines/predictions_of_human-level_ai_timelines/ai_timeline_surveys/ai_timeline_surveys), last updated 9 January 2024, and it names no survey of comparable size after the 2023 one. As of 26 October 2026 we have not found a successor elsewhere either, but absence from one list is not proof that none exists.

## How should you read these numbers?

Four cautions apply to all of them.

| Caution | Evidence grade | What it means |
|---|---|---|
| Framing changes answers | Well Supported | In 2023, respondents asked "probability by year X" gave a 50% HLMI date 34 years out; those asked "year for probability P" gave 17 years |
| Definitions differ | Established | "Most professions" (2012-13), "every task" (2016, 2023) and "all occupations" yield different dates |
| Samples are self-selected | Preliminary | Response rates of 10% to 65% leave room for participation bias |
| Experts forecast poorly in general | Contested | The 2023 authors cite evidence that subject-matter experts "perform poorly" at forecasting |

The framing effect is the largest. The Grace team combines both framings with equal weight and says "we do not know if one framing is more accurate than the other." A headline date is therefore partly a product of survey design.

Samples also change. The 2016 survey covered NIPS and ICML only; 2023 added four venues. The authors report that NeurIPS and ICML participants in 2023 held "very similar opinions to the full sample," which helps, but each survey still speaks only for people who chose to answer. The broader methodology is in [forecasting AI progress: methods and calibration](../vol12_forecasts/FCST-METHOD-001_forecasting_ai_progress_methods_and_calibration.md).

## Why do expert forecasts differ so much?

Within a single survey, individual answers span decades; the 2023 paper describes "a wide range of predictions." Several causes are documented in the sources above:

- **Definitions.** A machine that can do every task is a different target from one that does most jobs as well as a typical person.
- **Region and community.** In 2016, Asian respondents' median was 44 years earlier than North Americans'. In 2012-13, attendees of AGI conferences were more optimistic than the most-cited authors.
- **Timing of the survey.** The 13-year shift happened between two surveys one year apart; the paper reports it but does not test what caused it.
- **Methods beyond surveys.** Compute-trend models and forecasting platforms produce their own estimates; see [the uncertainties that matter most](../vol12_forecasts/FCST-UNCERT-001_the_uncertainties_that_matter_most.md) and the [forecast register](../reference/factbook/FCST-REG-001_forecast_register.md).

Surveys are best read as a snapshot of what a research community believes, useful for tracking change, not as a prediction of the date. For what might follow AGI, see [intelligence explosion vs singularity](./intelligence-explosion-and-singularity.md).

## Frequently asked questions

**When do AI researchers think AGI will arrive?**
In the October 2023 survey of 2,778 researchers, the aggregate gave a 50% chance that machines outperform humans at every task by 2047 and 10% by 2027. Individual answers varied by decades.

**Why did the 2023 survey date move so much?**
The 50% date moved 13 years earlier than in 2022 (2060 to 2047). The authors report the shift but do not establish why.

**Do surveys ask about AGI directly?**
Usually not. They use defined stand-ins such as high-level machine intelligence or full automation of labor, which give very different dates.

**Do most researchers think scaling current models will reach AGI?**
No, according to the AAAI 2025 panel survey: 76% of 475 respondents said scaling up current approaches is unlikely or very unlikely to yield AGI.

**Are expert surveys reliable predictions?**
They are measurements of opinion. Answers shift with question framing, definitions and who responds, so treat any single date as a Forecast with wide uncertainty.

## Sources

- Grace, K., Stewart, H., Sandkühler, J. F., Thomas, S., Weinstein-Raun, B., Brauner, J., Korzekwa, R. C.: [Thousands of AI Authors on the Future of AI](https://arxiv.org/abs/2401.02843) (arXiv 2024; *Journal of Artificial Intelligence Research* 84, 2025)
- Grace, K., Salvatier, J., Dafoe, A., Zhang, B., Evans, O.: [When Will AI Exceed Human Performance? Evidence from AI Experts](https://arxiv.org/abs/1705.08807) (*Journal of Artificial Intelligence Research* 62, 2018)
- Müller, V. C., Bostrom, N.: [Future Progress in Artificial Intelligence: A Survey of Expert Opinion](https://www.nickbostrom.com/papers/survey.pdf) (in *Fundamental Issues of Artificial Intelligence*, Springer Synthese Library, 2016)
- AAAI: [AAAI 2025 Presidential Panel on the Future of AI Research](https://aaai.org/wp-content/uploads/2025/03/AAAI-2025-PresPanel-Report-FINAL.pdf) (March 2025)
- AI Impacts: [AI Timeline Surveys](https://wiki.aiimpacts.org/ai_timelines/predictions_of_human-level_ai_timelines/ai_timeline_surveys/ai_timeline_surveys) (wiki page, last updated 9 January 2024)

*Status: draft, reviewed 26 October 2026 against the primary sources above. New survey rounds can shift these dates sharply, and we could not confirm whether a successor to the 2023 AI Impacts survey has published results.*