The exact text we saved

This is the version exactly as it was saved, character for character. The SHA-256 hash below is a 64-character code worked out from that text. Run the same text through SHA-256 and you get the same code; change a single letter and the code comes out completely different. Each version also stores the previous version’s hash, which chains them together so no version can be swapped out quietly.

TRUVACE RECORD VERSION
record: TRV-2026-2484
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-12T07:40:58.541161Z
status: published
lens: g_space
sector: labor
headline: TaskNet: Building Occupational Task Inventories from Worker Interviews at Scale
dek: Understanding how AI is reshaping work requires an accurate account of the tasks workers perform. Workers are often the first to see how their jobs are changing, as new responsibilities and tools often appear in everyday work before these changes reach formal descriptions. Interviews with workers can help capture these emerging changes in job tasks. Yet existing methods support standardized ratings across workers, while open-ended conversations with workers remain costly to collect and synthesize at scale. We in…
gain_title: In a study of 132 workers in six occupations, an AI interviewer produced 202 task statements that all passed O*NET analyst review and covered 89% of tasks reported by held-out workers, versus 79% for O*NET.
problem_title: (none)
trace_subject: (none)
gain_reading: In a study of 132 workers in six occupations, an AI interviewer produced 202 task statements that all passed O*NET analyst review and covered 89% of tasks reported by held-out workers, versus 79% for O*NET.
gain_evidence: all of which were accepted by O*NET Center analysts | It covers 89% of tasks reported by held-out workers, compared with 79% for O*NET
problem_reading: (none)
problem_evidence: (none)
quick_read: Stanford researchers built TaskNet, a system in which an AI interviewer talks with workers about their jobs and turns the answers into task lists. Across six occupations, 132 workers gave 2,726 task reports, which became 202 task statements. Analysts at the National Center for O*NET Development accepted every one, compared with roughly 65% of the write-in task suggestions workers submit through O*NET's current process.

Measured against the O*NET 31.0 database, TaskNet covered 75% of its 118 published tasks, added 79 tasks with no O*NET counterpart and refined 58 others. In interviews with 57 additional workers held out from the build, it covered 89.3% of reported tasks versus 78.6% for O*NET, and it was ahead in five of six occupations.

The choice of task list moves AI exposure estimates, which are used to guide workforce policy and training. Under TaskNet, exposure scores rose for Lawyers, Software Developers and Elementary School Teachers and fell for Personal Care Aides. The authors say much of the shift comes from newly documented coordination and oversight tasks. The paper is an NBER working paper and has not been peer reviewed.
limitation: Workers were recruited through a single AI data platform, about twenty per occupation and not matched to the U.S. workforce, and the design cannot separate why TaskNet's exposure estimates differ from O*NET's.
tag: Evidence-backed gain
key_points: The AI interviewer worked with 132 workers (20 to 25 per occupation); median interview length was 29 minutes, and 318 other workers gave 7,923 ratings of task relevance and importance. | Workers reviewed 2,798 task proposals made after their interviews and confirmed 94.2% unchanged, revised 5.3%, deleted 0.5% and added 163 tasks of their own. | Of the 202 TaskNet tasks, 79 (39%) had no O*NET counterpart, from 15% for Dishwashers to 52% for Software Developers, and 180 (89%) were rated core tasks versus 82% of O*NET's tasks for these jobs. | On held-out worker reports, coverage gains over O*NET were largest for Software Developers (17.0 points), Emergency Medicine Physicians (16.9) and Personal Care Aides (14.6); Dishwashers were tied at 85.7%. | Against 12,474 activities from 2,186 job postings, TaskNet covered 72.5% versus 70.2% for O*NET, but trailed O*NET for Elementary School Teachers (8.8 points lower) and Dishwashers (7.4 points lower). | Lawyers saw the largest exposure increase (9.6 points on RL feasibility, 0.16 on Eloundou-β), followed by Software Developers (9.1, 0.07); Personal Care Aides fell by 6.5 points and 0.11, and Dishwashers were unchanged. | About 53% of new tasks fell into invisible-work themes such as oversight and quality assurance, coordination, and teaching; 20% of new tasks depended on a recent tool or practice, against 3% of shared tasks.
rundown: The system works in three stages. An AI interviewer runs a semi-structured conversation and probes for tasks the worker has not mentioned by comparing the reported list with a model-inferred list for the role. Reports are then clustered into task statements using a task definition drawn from hierarchical task analysis and functional job analysis, and only tasks reported by at least two workers are kept, following O*NET's emerging-tasks rule.

Validation had two parts. Two O*NET Center analysts reviewed 229 system-generated statements and judged all of them valid tasks, though their feedback led the authors to rewrite 21% and classify 11% as adjacent work belonging to other occupations. Incumbents then rated the 204 remaining statements, and two were dropped for being relevant to fewer than a quarter of workers, leaving 202.

Compared with O*NET, TaskNet splits broad tasks into finer ones. For example, Lawyers' 'represent clients in court' became drafting motions and briefs, preparing exhibits and witnesses, and presenting arguments. TaskNet refined 28 O*NET tasks into 58 statements, and one Emergency Medicine Physician coordination task split into communicating patient status (more AI-exposed) and supervising residents (not exposed). Thirty O*NET tasks had no TaskNet counterpart, and 37% of those were specialized or dated, such as probating wills or setting up banquet tables.

The authors say the exposure differences may stem from the interview method, from more recent task data, or from different worker samples, and their design cannot separate the three. They note that newly documented oversight and coordination tasks score as highly exposed to AI but depend on human accountability and tacit knowledge, so exposure may reflect complementarity as much as substitution.
sources:
- peer_reviewed | NBER Working Paper | https://doi.org/10.3386/w35987 | 2026-11-01
prev: 0000000000000000000000000000000000000000000000000000000000000000

SHA-256 hash

1d59610159f287cacbc6608120135fa41344ba2bce51bdfe0d12d07155dcedad

Previous version’s hash

0000000000000000000000000000000000000000000000000000000000000000