Selected Pay Contrasts in Outsourced Human Work, 1890 to 2026: A Source-Linked Dataset and a Measurement Assessment

Human data for AI is often described as a new kind of work. This note tests a narrower historical claim: that markets for outsourced, piece-measured human work split into a low-price commodity tier and a smaller, better-paid expert tier, and that the AI human-data market of 2017 to 2026 follows that path. It compiles a source-linked dataset of 37 pay observations, 15 expert-to-commodity contrasts, 6 market timelines, 13 regulation-table rows and 6 intermediary-margin indicators across five cases: industrial home work in the UK and US (c.1850 to 1940), US clerical and typing work (1874 to 1976), Indian offshore business process outsourcing (1993 to 2021), online microwork (2005 to 2023), and human data for AI (2017 to 2026). Every number in the tables is produced from the deposited CSVs by one standard-library script. Cases, measures, a coding rule for when an expert tier exists, and five falsification criteria were registered before any source was read. The result is less than the question asked for, and the negative part is the more useful part. A named higher-skill category appears in all five selected cases, and the pre-registered contrast of at least two times is present in four of them and absent in clerical work. Comparing worker pay with worker pay inside one country, the eight historical midpoints run from 1.59 to 7.92 and the four US AI midpoints from 1.50 to 5.86, so the AI range overlaps the historical one but is not contained in it: the best-identified AI pair, the only one drawn from a single source in a single year, is 1.50, below both the pre-registered threshold and the historical minimum. After the second source check no pair in either group has an unverified observation, so the comparison restricted to verified pairs is the same comparison; the page carrying the 1.59 endpoint, which had failed at the first check, answered normally at the second, although it remains a relay whose own original was not found. That comparison was assembled at the verification stage, after the earlier result had been seen, so it is reported as exploratory rather than as a confirmatory output of the registered design, and with eight historical pairs across three cases neither containment nor non-containment is worth much. The lag from market emergence to a documented expert tier falls across the cases (56, 56, 11, 6 and 5 years), but the measure is a documentation lag, because the later endpoint is the earliest dated source that could be read, so it tracks the survival and indexing of sources as much as the market, and no timing claim is made from it. The regulation criterion cannot decide anything, because the AI case inherits the same legal instrument as the microwork case and its lag only restates the choice of start year. The AI rows do contain an onshore-expert and offshore-commodity contrast, but no cross-country pairs were built for the historical cases and the ILO's own survey documents a cross-country earnings gap in one of them, so what is particular to the AI rows is the alignment of tier with country, which is a description of the compiled rows and not evidence that the alignment is historically new. Every load-bearing row went through two source checks and now carries a terminal status. The first check counted 41 such rows in the data as it then stood, with 21 pay observations in the ratio pairs, and read 32 of them back at their cited sources, could not reach 2, and left 7 unchecked because no reported number rested on them. The pairs have since gained three pay observations and lost one, so the same definition now counts 43 rows. The second check closed the 9 the first had left open, gave a terminal status to every row in that current 43-row set, re-opened what the first had corrected, and reversed one of its findings: the first had reported that a 2002 Karnataka Act did not exist and recoded a regulation row on that ground, whereas the Act exists and the second check read its text, so the recoding is now presented as a coding judgment with the alternative coding's numbers given alongside it. Between them the checks corrected one regulation row, re-dated one job posting, re-attributed two payout shares to the analyst and press sources that actually state them, restored a revenue clause the first check had wrongly deleted, and withdrew one cross-country ratio because no source states the period its money figure covers, reporting the two figures in their own units instead. The rows that remain unverified are that Kenyan rate, the two market-emergence years that no source names, one margin row whose host refused on the day of the second check, and one margin row derived by us. The second check also added a script and a manifest that record, for every load-bearing row, the URL, retrieval time, HTTP status, SHA-256 and byte length of the bytes received and whether the quoted passage was present in them, so that a reader can repeat the check instead of trusting it. The note was drafted, analysed and reviewed by AI systems and no human has verified the data rows against their sources; the AI use statement sets this out in full, including the reversal above. The review rounds do not cover this version: eight rounds, one by the drafting model family and seven by an independent model, all ran before the second source check and read none of the 26,721 bytes of prose that folding it in added, so what they assessed was the earlier draft. A ninth round was run on the fold-in and raised three High, three Medium and one Low finding; this version applies them and that round did not see it, so no review has read the note as deposited. The deposit includes the dataset, the script, the pre-registration, both source checks, the snapshot manifest and those eight earlier review rounds, including the round that judged the note not depositable, so that a reader can audit the account rather than take it on trust.