Generative AI Fundamentals Map
Map/Measurement and understanding/Wave 9

Scalable oversight

Evidence people can check without rereading the work.

capabilityEmerging Open on the map →

Facts

Lane
Measurement and understanding
Wave
9 of 12 (build order, not a date)
Needs
17 boxes all the way back through hard links
Unlocks
4 boxes downstream through hard links
Leverage
0.1 points · stands in the way of 1 of 53 stages · rank 55 of 57
Cluster
Oversight and alignment

Evidence as of October 2026

Researchers across the major labs call readable chain of thought a "new and fragile opportunity" for monitoring, and warn it could be trained away.

  1. 01Chain of Thought Monitorability: a new and fragile opportunity
  2. 02Frontier Model Forum brief on chain-of-thought monitorability

Detail boxes 1

Chain-of-thought monitorability capability · Built today · helps, not required

Reading a model’s reasoning in plain language to catch bad intent.

AI 2027 marks the shift to “neuralese” recurrence and memory as the point where monitoring today’s mostly faithful English chains of thought is lost. Researchers across the labs call it a “new and fragile opportunity”.

  1. 01AI 2027 scenario (page truncated when read)
  2. 02Chain of Thought Monitorability: a new and fragile opportunity

Around it

Needs · 5
Unlocks · 5

Brakes

Works against this box

Loops and off-ramps

Openings researchers' judgement, October 2026 scouting

Startup room 34/100 · rank 94 of 124 Research room 63/100 · rank 20 of 124
Crowding2 Lab overlap4 Small-team fit3 Buyer today1 Research openness5
Depth
light
Players
MATS; SPAR; academic groups; lab alignment teams; no startups
Wedge
Oversight protocols that work when the overseer is weaker and the model adversarial; cheap experiments are possible. research
  1. 01MATS: scalable oversight benchmark