Planning for a transition we cannot date
AI capabilities on long tasks are growing quickly. Measurement is hard, forecasts disagree, and the dates below are planning horizons, not predictions.
- A · Historical fact
- B · Empirical evidence
- C · Philosophical proposition
- D · Future scenario
What METR measures
METR is an independent research organisation that evaluates frontier AI systems. Its central metric is the 50% time horizon: the length of task, measured by how long skilled humans take to complete it, that an AI agent finishes successfully half the time. In March 2025 METR reported that this horizon had doubled roughly every seven months since 2019 on its software and research-engineering tasks [44].
The November 2026 update
METR's update of 20 November 2026, "AI R&D-relevant model capabilities are improving rapidly and are getting harder to measure", reports [43]:
- Combining its Time Horizon 1.1 and Time Horizon 2 task sets, time horizons are doubling roughly every 3.61 months (95% confidence interval 2.08–4.94 months).
- A naive extrapolation of that trend would reach a 50% horizon of 200 hours, the longest threshold in METR's task set, around July 2027.
- Models often match or outperform skilled humans on coding problems and metric optimisation where feedback is strong; current public frontier models reach horizons of up to about 27 hours on the verifiable subset of the newer task set.
- Models appear to have worse judgment than human experts on tasks with long feedback loops or where solutions must generalise, and AI judges over-rate work by similarly strong models.
- The composite METR Capability Index shows no obvious general acceleration: weak to moderate evidence against general acceleration, moderate evidence against general deceleration.
Four distinctions keep these results honest:
- Task horizon is not general intelligence. It measures completion of well-specified, mostly software tasks at a 50% success rate.
- A task is not a profession. Jobs bundle many tasks, responsibilities and relationships.
- 50% is not reliability. Delegating safely requires success rates far higher, and METR's horizons at stricter thresholds are much shorter [44].
- Lab results are not workplace results. In METR's 2025 randomised trial, experienced developers took 19% longer with early-2025 AI tools while believing they were faster [45].
What this means for education
The pattern in METR's data (strength where feedback is clean, weakness where it is slow or ambiguous) points directly at the capabilities education should prioritise: independent judgment, verification and epistemic humility, understanding of system limitations, the ability to supervise autonomous agents and to collaborate with them. Bainbridge's "ironies of automation" explain why these cannot be left to the machines: the more automated a system, the more it depends on human supervisors whose skills automation lets wither [49], and people across expertise levels tend to over-trust automated aids [50].
How long a tenfold increase in time horizon takes at different doubling times
Three planning horizons
For strategy, this project uses three horizons. They are planning frames under our working assumption (90% project-level confidence that advanced AI will substantially automate cognitive work), not predicted dates and not a probability attributed to any researcher.
Cognitive acceleration
Increasingly autonomous agents take on long software, research and analysis tasks. Knowledge work reorganises around delegation, verification and supervision.
Potential intelligence explosion
If AI substantially automates AI research itself, capability gains could compound. Scientific discovery, automation and the economy could all accelerate.
Post-labor transition
A scenario in which advanced AI and robotics perform most economically valuable cognitive and physical tasks, and income, status and purpose must be organised without wage labour at the centre.
The intelligence-explosion idea goes back to I. J. Good's 1965 argument that a machine able to improve machines could trigger recursive self-improvement [46]. The AI 2027 scenario sketches such a path in detail [47]. Surveyed AI researchers, by contrast, put full automation of all occupations much later, and their answers swing widely with question wording [48]. METR's own conclusion is cautious: current public evidence is insufficient to rule out drastic acceleration of AI research in the next couple of years.
There are serious reasons to expect friction. David Shapiro, who expects a post-labor transition, argued in January 2026 that grid connections, transformers, memory and chip packaging were slowing deployment in a "digestion phase" before renewed acceleration around 2028 [60]. Physical bottlenecks can delay the economic effects of capabilities that already exist.
Design principle: robustness to timing. Every element of this framework must be worth learning even if the transition is slower, partial or uneven. Judgment, wisdom, craft, health, civic understanding and the arts are valuable in any of these futures. That is why the framework is built around human formation rather than around predictions about which jobs survive.