LO
AI researcher — how we know
The page itself gives the judgements. This one gives what they rest on: which technologies bear on the work, how the estimate moved since language models reached the public, and the method behind both.
Which technologies matter here#
Four separate signals. They are deliberately not added together — a job exposed to two technologies is not twice as exposed.
How it got here#
The index is not a static number. This is where it would have sat at each capability checkpoint since ChatGPT — reconstructed, and labelled as such.
—— this stretch contains a verified event- - - no event in this stretch — reconstruction only0 = no task exposed, 100 = every task exposed
● 3 verified events for this occupation, plotted at the date it happened — the parts of the line near a marker are anchored to something checkable.
The one curve on this site reconstructed against an employer's own published count rather than only against capability releases, and it is worth saying that the employer sells the capability. The low start is real: in late 2022 writing the training code, the harness and the cluster plumbing was hand work, and the tooling a researcher had was a better autocomplete. The climb from 2023 is that build phase being handed over, and the 2025-2026 section is where the handover stops being assistance and becomes delegation — by mid-2026 the disclosure puts the ratio at 3.1 agent-workdays for every workday of human labour. It flattens near the top rather than continuing, and the mechanism for the flattening is in the same document: the share of agent output going to high-level planning stays minimal, and over half of successful long-horizon tasks still take a human intervention. Read the height as how much of the doing has moved, and the flattening as where the deciding still sits — not as a ceiling anyone has demonstrated.
A flat line is not a forecast of safety. It says which tasks automation has reached so far — the occupations that moved least here are the ones where the constraint is physical or regulatory, and both of those can change.
Written about this#
These pieces argue from the same records this page holds, and each of their sections names what it rests on.
Method and sources#
- Assessment date
- 2026-09-13
- Basis of the task judgements
- 5 evidence-backed · 1 platform inference · 0 not enough evidence
- Verified events
- 5