AI 2027: a scenario with two endings

A narrative can make a possible future vivid. It can also make a forecast look more settled than its authors intend. AI 2027 is useful to read with two documents open: the original scenario and the authors’ later milestone forecasts.

What the original says

Published in April 2025, the scenario follows a rapid path through AI development and branches into a race ending and a slowdown ending. Those branches ask readers to consider how decisions change outcomes. They do not describe two events that are both predicted to occur. Read the original scenario for the sequence and its assumptions.

The five authors are Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean. The original ledger omitted Dean. The authors’ individual expectations differ, so “the authors predict 2027” is too crude a description.

What changed in 2026

The 2 April 2026 update moved Kokotajlo’s median for an Automated Coder from late 2029 to mid-2028, and Lifland’s from early 2032 to mid-2030. This milestone describes a company preferring AI coding to its human software engineering workforce. Their update points to agentic coding evidence and revised assumptions. It is not a median date for human extinction.

Four questions hidden inside “when will AI arrive?”

Can a system write useful code? Can it replace a broad engineering workflow? Can it perform at top-expert level across cognitive tasks? Can people keep it under effective control? Those questions require different evidence. A result on one does not automatically answer the others.

This ledger’s editorial rule is to retain the original scenario year while recording revisions beside the relevant milestone. Moving the scenario to whichever year is currently fashionable would erase the history. Treating a coding update as an apocalypse-date revision would misdescribe it.

What would count as being right?

The most informative comparison is a list of defined milestones, observed dates and reasons for any mismatch. A scenario can be early on one step and late on another. A probability forecast should be assessed as a probability forecast; one outcome cannot, on its own, establish whether an entire forecasting method was well calibrated.

Why the risk still matters

The concern is the chain of events: stronger agents might accelerate research, institutions might race to deploy them, and oversight might fail to keep pace. Each link needs its own evidence. The AI risk explainer distinguishes experiments, observed incidents and possible future loss of control.

The productive reading exercise is to ask which assumptions you accept, which you reject, and what new observation would change your mind. That leaves room for concern, disagreement and useful preparation without pretending the ending has already been entered into a calendar.