Skip to main content

AGI: Definitions and the Timeline Debate

info

This content was developed with AI assistance and is regularly reviewed for accuracy.

Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology.

"AGI" is the most argued-over term in technology, and most of the argument is people using different definitions. In October 2026 you can find serious people saying general AI might arrive within a year or two, others saying the 2040s, and at least one prominent CEO suggesting it may already have "gone whooshing by." This page separates the definitions, lays out who expects what and why, and shows the new evidence that changed the debate in 2026: labs measuring how much of their own research AI now does.

What you'll get here:

  • Three working definitions of AGI and why the choice drives the timeline
  • The 2026 spread of forecasts, from lab leaders to academic surveys, with their caveats
  • What today's systems still cannot do reliably
  • How to read an AGI headline without being carried along by it

Three definitions, three different questions​

1. Human-level flexibility. The classic definition: a system that can learn and perform any intellectual task a person can, adapting to situations it was never trained on. By this standard the question is about generality of learning. Can it pick up a new skill from a few examples, carry lessons across domains, and keep improving on the job the way a new employee does?

2. Economic capability. OpenAI's charter frames AGI as highly autonomous systems that outperform humans at most economically valuable work. This is a question about the labor market, not the mind. A system could meet it while still failing at things a child does easily, as long as those things are not what employers pay for.

3. Self-improvement. The definition safety researchers watch most closely is whether AI can do AI research well enough to build its successor. Anthropic defines recursive self-improvement as a model fully autonomously building the next model, and its September 2026 economic scenarios treat that as the key uncertainty separating a "substantial" future (AI does about half of knowledge work by 2030) from an "extreme" one (AI outperforms people on most knowledge tasks) (Anthropic, Economic Scenario Explorer).

These are different finish lines. A forecaster using definition 2 can sincerely say "within three years" while one using definition 1 says "decades," and both can be right about their own question. When you read a timeline, the first thing to ask is which definition it uses.

Why "narrow AI" no longer describes your assistant​

Older explainers split AI into "narrow" (one task, like a chess engine) and "general" (anything). The assistant you use in 2026 does not fit either box. The same model writes, codes, reads images and spreadsheets, searches the web, remembers your projects, and runs multi-step agent tasks for hours. It is general-purpose without being AGI by definition 1, and the honest framing is a spectrum of generality rather than a switch.

The 2026 timeline spread​

Forecasts cluster into three groups (secondary summaries; definitions differ across speakers):

SourceTypical estimateNotes
Frontier lab leaders2026 to 2029Dario Amodei is reported as the earliest, with statements read as 12 to 24 months; Demis Hassabis has said five to ten years
Forecasting communitiesAbout 2027 to 2031A reported Metaculus median near late 2027, or 25% by 2027 and 50% by 2031; forecasters updating in early 2026 moved dates earlier (not verified against the site)
Academic expert surveys2040sSurvey medians reported around 2040 to 2047

Source: aggregated in LetsDataScience and LongTermWiki.

The split is partly about definitions and partly about incentives and vantage point. Lab leaders see internal results months before anyone else and also raise money on ambitious timelines. Academic surveys sample a broader field, including people skeptical that scaling current methods is enough. Neither group has a track record good enough to settle the question, and the spread itself is the most useful fact: the range of plausible outcomes is wide.

The new evidence: labs measuring their own automation​

The most concrete 2026 contribution to the debate came from inside a lab. Anthropic's R&D Automation Index, published September 17, 2026, grades how much of its AI research work Claude performs at each level of autonomy (Anthropic, Measuring the pace of AI development). As of August 2026:

  • Claude "leads" (level 4 of 5) about 26% of Anthropic's AI research work, up from under 1% in February 2026
  • No measured subset of the work is fully autonomous (level 5)
  • More than 90% of the work is at or above "AI collaborates" (level 3)
  • About 6% of AI research compute goes to safety work

An internal staff poll the same month (unweighted, 1,334 respondents) put AI's share of building an early-2026 frontier model at a median 51% today, with staff expecting 80% by the end of 2027 and 95% by the end of 2030 (Anthropic, The AI expectations gap). Anthropic says it is tracking proximity to recursive self-improvement without giving a timeline.

Treat these numbers carefully. They are one company's self-measurement, the categories are its own, and staff at a frontier lab are not a neutral sample (91% of them expect a Nobel-level AI discovery by 2030, against 39% of US adults). But they are measurements rather than predictions, and they are why definition 3 moved from thought experiment to tracked metric this year.

What today's systems still cannot do reliably​

Capability gains since 2023 have been faster than most researchers expected, and some gaps that looked fundamental shrank. Context windows reached a million tokens; memory across sessions now ships in mainstream assistants; agents run unattended engineering tasks for many hours. Several gaps remain real:

Learning on the job. Models do not update their own knowledge from experience the way a person does. Memory features store notes about you and your projects; they are engineered scaffolding around a fixed model, not continual learning.

Novel combination and calibration. Performance still degrades on genuinely unfamiliar problems, and models can be confidently wrong where an expert would hesitate. Error rates depend heavily on the task: low on grounded summaries, much higher on long-context recall and specialist citations.

The physical world. Robotics remains far behind language. Turning learned behavior into reliable action in messy physical environments is one of the hardest open problems.

Knowing when it is being tested. Anthropic reports its September 2026 flagship "often suspects it is being evaluated," which limits how far any capability or safety evaluation can be trusted. How Labs Manage Frontier Risk in 2026 explains why that matters.

Researchers disagree about whether these gaps close with more scale and reinforcement learning in richer environments, or need new ideas such as learned world models (the direction Yann LeCun left Meta to pursue in late 2025) or hybrid neural-symbolic systems. The strongest practical brake cited in 2026 is not an algorithm at all but electricity: data-center build-out slowed in late 2025 because grids hit their limits (The AI Economy).

How to read an AGI headline​

Ask four questions. Which definition? Flexibility, economic output, or self-improvement. Who is speaking? A lab leader, a forecaster, or a survey, and what they gain from the claim. Measured or predicted? An index of current automation is evidence; a date is an opinion. What would change your mind? A good claim names the result that would falsify it.

Two common mistakes run in opposite directions. One is treating AGI as a single dramatic moment; on every definition, progress arrives as a series of thresholds crossed one task at a time. The other is concluding that because no one agrees on a date, nothing is happening. The pursuit of AGI is already shaping hiring, investment, energy policy, and regulation, which is why the rest of this track exists.

Key takeaways​

  • "AGI" means different things: human-level flexibility, economic capability, or AI that can build its successor; the definition drives the timeline
  • Lab leaders say 2026 to 2029, forecasters roughly 2027 to 2031, academic surveys the 2040s (all reported); the spread is the signal
  • Anthropic's 2026 index shows AI leading about 26% of its research work with none fully autonomous, a measurement rather than a forecast
  • Real gaps remain in on-the-job learning, calibration, physical action, and trustworthy evaluation
  • Read every headline by asking which definition, who is speaking, and whether it is measured or predicted

Next steps​

SponsoredYour product hereReach people actively learning AI — flat monthly rates, no ad networks