← Research

Does the finding replicate outside the panel built around AI employers?

The question. Run the same comparison separately in four sub-corpora assembled for different, and in one case opposite, reasons. Which term groups move in all of them?

What it found

Delegation is the largest non-definitional effect in the study and it rises in every arm where it can be measured, including The Muse, which was not assembled around AI. Experimentation rises everywhere too. Cross-functional briefing is flat in all four, the most stable null here. Diligence does not replicate: governance is flat and ethics inverts in The Muse. So the AI premium does not fall on one column of the four-Ds diagram; it falls on Delegation and Tinkering Disposition, which sit in different columns.

The numbers

Term groupSub-corpusAIEM elementAI nNon-AI nLift
AI oversightThe MuseDiscernment2221527.51
AI oversightRemote boardsDiscernment1,83812113.45
AI oversightATS panelDiscernment6271310.88
handing work to a systemATS panelDelegation1,9166310.49
experimentation, pilotsThe MuseTinkering Disposition4582124.64
ambiguityHacker NewsTolerance for Ambiguity58164.46
handing work to a systemThe MuseDelegation3541834.08
evaluation languageHacker NewsDiscernment3151043.65
ethics, bias, responsible AIHacker NewsDiligence61223.27
experimentation, pilotsRemote boardsTinkering Disposition7222252.90
handing work to a systemRemote boardsDelegation2,5378562.81
governanceHacker NewsDiligence93492.37
ambiguityRemote boardsTolerance for Ambiguity9283272.28
experimentation, pilotsATS panelTinkering Disposition1,0261302.25
interpretationHacker NewsDiscernment73382.25
experimentation, pilotsHacker NewsTinkering Disposition51311.96
handing work to a systemHacker NewsDelegation25161.79
verificationHacker NewsDiscernment2561761.74
ownershipRemote boardsInternal Locus of Evaluation3,6512,0611.68
interpretationATS panelDiscernment5,0888351.67
ethics, bias, responsible AIATS panelDiligence9111301.61
evaluation languageATS panelDiscernment7,3441,5701.49
ownershipHacker NewsInternal Locus of Evaluation4343721.42
ambiguityThe MuseTolerance for Ambiguity5077691.36
ownershipATS panelInternal Locus of Evaluation3,0367821.36
interpretationRemote boardsDiscernment17,40010,8641.32
evaluation languageRemote boardsDiscernment34,17522,7501.32
ownershipThe MuseInternal Locus of Evaluation1,2891,9771.32
verificationATS panelDiscernment6,3881,6351.29
governanceATS panelDiligence1,6154031.22
ethics, bias, responsible AIRemote boardsDiligence6554561.22
ambiguityATS panelTolerance for Ambiguity1,4173331.20
evaluation languageThe MuseDiscernment5,91910,6951.11
cross-functional briefingRemote boardsDescription6,8626,5801.04
cross-functional briefingATS panelDescription7,7332,7221.02
governanceRemote boardsDiligence1,9111,5781.01
verificationRemote boardsDiscernment31,18128,7090.99
cross-functional briefingThe MuseDescription5,38112,3640.94
verificationThe MuseDiscernment4,0879,4850.90
governanceThe MuseDiligence1,5003,3290.88
interpretationThe MuseDiscernment2,5465,9060.87
cross-functional briefingHacker NewsDescription5347200.87
ethics, bias, responsible AIThe MuseDiligence5891,7930.62

Never sum across sub-corpora. The columns are different populations.

How it was computed

Identical method to the Greenhouse analysis, computed separately per sub-corpus. Length strata are formed within each sub-corpus and discipline, so a discipline is compared against its own length distribution inside its own panel.

How to read it

Sub-corpora are never pooled and there is deliberately no combined row: they are different populations with AI shares from 16% to 70%, not strata of one. A dash means the cell did not clear suppression in that sub-corpus. Delegation rests on a single ratified term and its non-AI arms are thin, so read it with its cell counts rather than alone.

Source view: competency_lift_by_corpus · published 2026-09-12