Everyone Has Updated Their AGI Timeline. Nobody Has Updated Their Plans.

A few years ago, most people building AI thought human-level systems were decades away. Surveys of machine-learning researchers put the median at around 2060, and forecasters on Metaculus expected AGI about fifty years out. The people doing the work did not think it was imminent.

That has changed, though not evenly. On the forecasting platforms the expected wait collapsed: the mean Metaculus forecast for AGI fell from about fifty years away to about five in the four years to early 2025, according to a review of expert forecasts by 80,000 Hours. AI researchers moved less, but they moved: the largest survey of them, run in 2023, gave a 50% chance that unaided machines would outperform humans at every task by 2047, thirteen years earlier than the same survey had found a year before.

The safety budget has not moved proportionally. The money going into alignment research, governance development and deployment oversight has grown far more slowly than the money going into capability.

What the Forecasts Say

Metaculus, the community forecasting platform with one of the better track records on technological questions, tracks several AGI questions, and they show how much the answer depends on the definition. As of October 2026, the community's median date for the first general AI system to be devised, tested and publicly announced was early 2031. For a "weakly general" AI system, a lower bar defined by a fixed set of tests, it was late summer 2027. Both numbers move daily.

The researchers who build these systems are more cautious. In the largest survey of its kind, 2,778 researchers who had published at top AI venues gave a 10% chance that machines would outperform humans at every possible task by 2027 and a 50% chance by 2047. The same survey did not put the chance of all human occupations becoming fully automatable at 50% until 2116. Individual answers ranged from the next few years to centuries away, so the gap between forecasters and researchers is not a split between a confident crowd and a confident profession. Neither group agrees with itself.

These aren't wild extrapolations from hype. They are forecasts from people tracking the rate of capability improvement and updating their models accordingly. The most detailed attempt to turn those forecasts into a concrete, dated sequence of events is the AI 2027 scenario, which names the specific capabilities that would have to develop, and roughly when.

Why Timelines Compressed

The capabilities that drove the update were real and documentable. GPT-4 in 2023 passed a simulated bar exam with a score around the top 10% of test takers, and on official practice materials for the US medical licensing exam a separate evaluation found it exceeded the passing score by more than 20 points. Claude 3 Opus, released in March 2024, posted gains across reasoning benchmarks, and the GPT-5 and Claude 4 releases in 2025 continued the trend.

The scaling laws held. With more compute, more data, and better architectures, capability improvements continued at rates that exceeded many models of how fast they could continue. People updated their forecasts because the evidence demanded it.

The Money Tells the Story

Nobody publishes a full accounting of what is spent on AI safety, and public estimates vary widely. What is not in dispute is the mismatch in scale. The largest labs and their backers now commit tens of billions of dollars a year to compute alone, while safety work, counting lab safety teams, nonprofits such as MIRI and ARC, and academic groups, is a small fraction of that. OpenAI's superalignment team, the most visible lab commitment to the problem, was wound down in 2024.

The governance situation is similarly behind the curve. The EU AI Act, the most comprehensive regulatory framework to date, was largely negotiated before today's agentic systems existed. The US executive orders on AI safety have been partially implemented, partially reversed, and don't have independent technical expertise to enforce them against frontier capability claims. The most recent of them asks labs to volunteer for pre-release review rather than requiring it.

Sam Altman Said the Quiet Part Out Loud

In a 2024 blog post, Sam Altman wrote about "superintelligence in a few thousand days," which is somewhere between five and ten years. Here was the CEO of the most prominent AI lab stating publicly that transformative AI is a near-term planning horizon rather than a speculative future. He framed it optimistically, in terms of abundance, medical breakthroughs, and economic growth.

What the framing omitted is that a near-term planning horizon for capability has to be a near-term planning horizon for safety. You can't treat "we'll deploy transformative AI in a decade" as opportunity while treating "we need to ensure that AI is aligned before deployment" as a research project that can continue at its current pace.

The math is simple enough to do on a napkin. If AGI arrives in 2030, the alignment problem requires twenty years of research to solve, and it's currently 2026, the timeline is already wrong. The solution could never arrive in time. Alignment may well be solvable; the problem is that the available time is being spent elsewhere, and an intelligence explosion could compress what time remains far faster than the calendar suggests.

Plans That Would Match the Forecasts

Safety investment would scale with capability investment. Governance would develop at the pace of capability. Deployment decisions would turn on safety evaluations rather than competitive timelines that no one can unilaterally exit.

None of that describes the current situation. The timelines have moved; the plans have not. Maybe this is manageable. Maybe alignment is easier than we fear, or the timeline forecasts are still off. Or this is the most important planning failure in human history. The stakes are high enough that "maybe it's fine" is an unreasonably casual attitude to take into a decade like this one.

Correction, 6 October 2026: An earlier version of this article said surveys of AI researchers had moved their median forecast into the 2028–2032 range. The largest such survey puts it at 2047; the article now distinguishes researchers from forecasting platforms. The earlier version also cited an annual Alignment Forum survey that we could not verify, linked the wrong Metaculus question for weak AGI and described its definition incorrectly, gave forecast figures without dates, and stated safety-funding figures without a source. Those passages have been corrected or removed.