Version 1, 8 August 2026. NM AI Research.
AI disclosure: parts of this text were artificially generated with AI assistance and reviewed by the author. The model, and the conflict it creates, are named in the Conflict of interest section.
DOI: 10.5281/zenodo.21819102 (concept, always resolves to the current version)
A scored, re-runnable record of what twelve consumer AI products publicly evidenced against the transparency clauses of the EU AI Act, measured on 5 and 6 August 2026, four days after Article 50 began to apply.
This is not a compliance assessment. Every question is of one form: is evidence of X publicly available, yes or no. Every cell carries a URL to an artefact that was opened and an access date. no artefact found is a statement about a bounded search, never about a provider.
The point of the exercise is the second run. The Article 50(2) marking duty carries a transitional period for systems already on the market, expiring 2 December 2026. The same eleven questions will be asked again shortly after that date, with this rubric unchanged, and the movement between the two runs is the intended result. This is the baseline it will be measured from.
| File | What it is |
|---|---|
Article50_Snapshot_v1_2026-08-08.pdf |
The piece |
article50_scores.csv |
Every cell: product, provider, clause, item, answer, evidence URL, artefact type, access date, note |
reproduce.py |
Re-prints the tables from the CSV and validates it. Standard library only, no network, no keys |
parse_signatories.py |
Parses the Commission's Code of Practice signatory table from a saved page snapshot |
signatories.json |
The parsed Section 1 and Section 2 lists as at 5 August 2026 |
article50_rubric_v1.md |
The rubric, frozen and signed off before any artefact was opened |
article50_primary_verification.md |
Each fact taken from the primaries, anchored to its paragraph |
communication_131214.pdf |
European Commission C(2026) 5054 final, 20 July 2026, the communication approving the draft Guidelines. Included unaltered so the clause anchors can be checked against the text as it stood on the measurement date |
guidelines_131215.pdf |
Its 51-page annex, the draft Guidelines themselves. Every paragraph number cited in this deposit refers to this file |
python3 reproduce.py article50_scores.csv
Prints the per-product and per-item tables with numerator and denominator on every count, then validates the file. Exits non-zero if any row carries an answer outside yes, no, no_artefact_found, or a yes or no without an evidence URL or access date.
python3 parse_signatories.py signatory_page_2026-08-05.html
Reproduces the signatory counts offline. Note it reports a discrepancy in the source: the Commission's page states 82 Section 1 signatories and its table contains 83 entries. Section 2 states and contains 152.
The rubric was written and signed off before any company artefact was opened. A prior-art search ran first and is reported whether or not it was convenient. Products were selected by a stated inclusion rule written before the set was chosen. Evidence was collected from provider documentation and from manual browser captures of product interfaces, saved as artefacts. Where an artefact could not be opened, that is recorded as such rather than treated as an answer. Where a call rested on judgement rather than a quotation, the note says so and quotes what it relied on.
Twelve products, one observer, one country, one browser, two days, four days into a regime whose interpretive guidance is still in draft. The Commission's Guidelines were approved in content on 20 July 2026 and become applicable only once adopted in all languages, and are non-binding in any event. Several cells could move on a different day. The disclosure under Article 50(1) was found to vary with account state, which is reported as a finding in its own right.
This is not the first attempt to score named products against these clauses. Rijsbosch, van Dijck and Kollnig, "Missing the Mark", Policy & Internet (2026), preprint arXiv:2503.18156, scored fifty named image generation systems on machine-readable marking with a public detection tool, using 2025 data. A Dutch field test published on aiactblog.nl tested ten customer service chatbots on 13 July 2026 with a repeat promised. Both are credited in the piece.
The research question, method, sourcing decisions and analytical judgements are the author's. The assisting model is not always a neutral party to the subject matter, and here it plainly is not: it is made by Anthropic, whose products are in the scored set. Parts of this text were artificially generated and were reviewed by the author before publication.
The rubric was fixed before any artefact was opened, that product was scored last, every cell relating to it was re-checked, and every quotation from its documentation is verbatim. The results run in both directions: it evidences none of the four marking items and holds one of only two evidenced cells on designed identity disclosure.
Guarantee: every measured claim is traced to the primary document named in the Verification note, and reproduce.py regenerates every table from article50_scores.csv. Judgements of standing are identified as such rather than presented as measurements. A reader can check this without trusting either party.
What the author cannot guarantee: text drafted by a language model can carry errors that are locally plausible, consistent in tone with the surrounding prose, and therefore survive review. Several methods have been deployed to mitigate this, including explicit instructions, internal red-teaming and cross-lab blindspot checks, but the author does not claim the review is exhaustive. Corrections are logged against the DOI when surfaced.
No warranty is offered beyond the terms of the CC BY 4.0 licence. Independent analysis and open-science documentation only, not legal advice.
NM AI Research (2026). Public evidence under Article 50: a baseline snapshot of twelve consumer AI products against the EU AI Act transparency clauses. Version 1. Zenodo. https://doi.org/10.5281/zenodo.21819102
CC BY 4.0, for the text, data and code produced by the author.
communication_131214.pdf and guidelines_131215.pdf are European Commission documents,
C(2026) 5054 final of 20 July 2026 and its annex. They are reproduced unaltered under the
Commission's reuse terms (Decision 2011/833/EU) and are not covered by the author's
CC BY grant. They are included so that every paragraph anchor cited here can be opened
against the text as it stood on the measurement date, rather than against a page that may
since have changed.
Verification: every date, deadline and clause statement in this deposit was traced to one of the two primary documents included with it, the Commission Communication C(2026) 5054 final of 20 July 2026 and its annexed draft Guidelines, with paragraph anchors listed in article50_primary_verification.md. Signatory counts and lists were parsed by script from a saved snapshot of the Commission's own page rather than taken from any secondary report, after an early secondary reading proved wrong. The publication date of the Code of Practice, 10 June 2026, is verified at Guidelines footnote 43. The reported adequacy assessment of 8 to 9 July 2026 appears in neither primary and is asserted nowhere. All product-level cells come from provider artefacts opened on the stated access date, never from secondary coverage or from memory.