Research opportunity · Who can know about AI? Epistemic dependence and jurisdiction over the field
Do developers' usage statistics hold up when someone else classifies the conversations?
In short: has this been done?
Related work: Partly addressed. Related studies cover parts of this question; what they leave open is set out below. Only a first-pass search has been done, so it is not yet known whether a fuller search would find a study that does the whole design.
How far checked: First-pass screen only. A first-pass AI web search on 16 October 2026. No deeper check yet. Every search and check so far was done by AI; no person or domain expert has checked them yet.
The searches, the nearest work found and what the search could not reach
Follow this opportunity: Atom feed of RO-LD08-03 updates (new versions, and revisions of the claims it rests on).
Question
When the same chatbot conversations are classified independently of the developer, with a validated scheme and human-coded benchmarks, how far do headline usage measures such as the automation share, the work share and task success move away from the figures developers publish?
Gap type: Instrument stabilityOne PhD chapter or paperNeeds a budget (medium or higher)Related work: Partly addressed
Why it matters
Usage statistics from AI developers are now cited as evidence about how people use AI, but each is produced by the developer's own model and checked in-house. One developer's automation share moved from 49% to 45% when a different model classified the same data. An independent study of public logs borrowed developer-linked classifier prompts without validating them, so independent data alone do not supply an independent instrument. Without an independent instrument it is unknown how much of the published picture is the instrument.
The debate lists this among what we still do not know: How stable are vendor usage measures when the classifier, the validation and the data owner change?
Evidence that motivates it
- In its September 2025 Economic Index report, Anthropic reported that the share of automation in Claude.ai conversations depended in part on which model did the classifying: re-running the same August 2025 data with Claude Sonnet 3.7 instead of Claude Sonnet 4 gave 45% automation instead of 49%, though the rise in directive use remained.
How this study uses it: The automation share moved from 49% to 45% when another model classified the same conversations. - Anthropic’s first Economic Index report classified Claude conversations as 57% augmentation and 43% automation; the classifications were made by Claude and checked by the authors on 150 hand-validated conversations drawn from those for which users had submitted feedback, with 90.7% agreement on the augmentation–automation label and 86% to 95.3% acceptable task assignments by level of detail.
How this study uses it: Classifications made by the developer's model and checked on 150 conversations chosen from those with user feedback. - Anthropic’s January 2026 Economic Index report added measures such as “task success,” which “measures Claude’s assessment of whether Claude completes tasks successfully”; for Claude.ai the classifiers were checked against a human researcher on a small set of transcripts for which users had given feedback, and the report calls them “directionally accurate”; it found augmentation rising to 52% of Claude.ai conversations.
How this study uses it: Task success is the model's own assessment of whether it completed the task. - Analysing 37,845 Australian conversations from WildChat, a public dataset of ChatGPT logs collected outside OpenAI, university researchers classed 61.1% as work-related against 37.3% in OpenAI’s published global figures; to classify them they adopted the classifier prompts and categories of OpenAI-co-authored work.
How this study uses it: An independent study of public logs reused the developer-linked classifier prompts without a validation of its own.
A possible design
Take a public or donated conversation corpus, build a human-coded benchmark sample stratified by topic and language, and classify the corpus with several models and prompt sets, including the prompts developers publish. Report each headline measure with its spread across classifiers and its agreement with the human benchmark, and compare with the developers' published figures for the same period.
Data needed
A licensed public or donated conversation corpus; published developer taxonomies and prompts; trained human coders; access to several classifier models.
Feasibility
One PhD chapter or paper. Medium; human coding of the benchmark sample and model inference costs. Public and donated corpora differ from the developers' user bases, so the study measures instrument sensitivity, not the true population share.
Proposal pack: this page and its references on one printable page, with BibTeX.
Has this been done?
Partly done already: related work covers some of this question (first-pass screen of 16 October 2026; not yet checked in depth).
The developer's own model-swap check and one independent study of public logs show that the measures move, but no study was found that classifies the same corpus with several instruments against a human benchmark and compares the result with developers' published figures.
How strong this evidence is
- Who searched and checked: no checker is recorded beyond the first-pass screen, which was part of the AI drafting (Claude (SocioLens Project Lead), AI-drafted from the LD-08 debate). Every check so far was done by AI; no person or domain expert has checked it yet.
- What was read: 1 of the 3 nearest works below was judged from an abstract, summary or search listing only; it is marked. Not every entry says how much of the work was read.
- Where the search could not reach: Screened by web searches and an arXiv record lookup on the date shown; an R package that re-reads the published Economic Index aggregates appeared in search results but was not opened. Absence of a hit is not proof that the work does not exist.
- Before building a project on it, run your own search in a scholarly database such as Scopus, Web of Science, Google Scholar or SSRN.
Nearest work found
- Anthropic, Economic Index report, September 2025 (LD08-SL-C18)
Full text read for this debate. The developer reran one week of data (4–11 August 2025) with an older model; one comparison, in-house. - WildChat study of Australian conversations (LD08-SL-C22)
Full text read for this debate. Independent data, but the classifier prompts come from developer-linked work and are not validated separately. - Yan, Nguyen, Su, Lieffers & Le, "ShareChat: A Dataset of Chatbot Conversations in the Wild," arXiv:2512.17843 (2025) Judged from an abstract, summary or search listing only
Abstract read. 142,808 publicly shared conversations from five platforms, a possible corpus for the design; not a classifier comparison.
Search queries used
independent replication Anthropic Economic Index classification Clio validation human annotation automation augmentationWildChat classifier validation work-related share OpenAI taxonomy independent reclassificationarXiv id_list: 2512.17843
Cite this page
SocioLens project. (2026, October 16). Do developers' usage statistics hold up when someone else classifies the conversations? [Research opportunity RO-LD08-03, version v1]. SocioLens. https://sociolens-web.pages.dev/opportunities/ro-ld08-03/v1/
@misc{sociolens_ro_ld08_03_v1,
author = {{SocioLens project}},
title = {Do developers' usage statistics hold up when someone else classifies the conversations?},
year = {2026},
date = {2026-10-16},
howpublished = {\url{https://sociolens-web.pages.dev/opportunities/ro-ld08-03/v1/}},
note = {Research opportunity RO-LD08-03, version v1}
}The link points to this version, whose record never changes. For a finding, cite the original study as well: SocioLens summarizes research and does not replace it.
Provenance
RO-LD08-03 v1. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-08 debate on 2026-10-16. Not yet reviewed. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.