AI Futures Project publishes 'AI 2040: Plan A'

The AI Futures Project — Daniel Kokotajlo, Thomas Larsen, Eli Lifland and Romeo Dean, again with prose from blogger Scott Alexander — published “AI 2040: Plan A,” a follow-up to their 2025 “AI 2027” scenario. Where AI 2027 modelled an underregulated race toward an intelligence explosion, Plan A is explicitly normative: the authors call it “the least bad plan we currently know of” rather than a forecast, describing how governments could still navigate to a good outcome from where the field stood in mid-2026.

The scenario’s central mechanism is a US-China agreement establishing what the authors call a trustless regulatory regime — verification through chip tracking and datacentre auditing rather than mutual trust — that lets both countries keep developing AI through the 2030s but pause deliberately at roughly human-genius-level systems to solve alignment before attempting superintelligence, pushing that milestone out to around 2040 rather than the early-2030s date the authors judge otherwise likely. The document projects the coordinated path would still deliver large economic gains, including a universal dividend the authors modelled rising from $25,000 to $1.6 million a year in current-dollar terms, alongside disease cures and AI-assisted governance.

Alexander framed the document’s purpose as a benchmark for evaluating political proposals — asking whether a given AI policy fits any coherent vision of how the whole trajectory ends — rather than as a literal policy blueprint. As with AI 2027, the specific figures were illustrative rather than confident predictions, and the plan depended on a degree of US-China cooperation that had no real precedent in either country’s AI policy to that point.