Public evidence under Article 50: a baseline snapshot of twelve consumer AI products against the EU AI Act transparency clauses

Abstract

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 note at the end of the snapshot and in the README. Article 50 of the EU AI Act began to apply on 2 August 2026. This is a dated snapshot of what twelve consumer AI products publicly evidenced against its transparency clauses on 5 and 6 August 2026, four days in, taken as a baseline for a second run after the Article 50(2) transitional period expires on 2 December 2026. It is not a compliance assessment and no cell should be read as one. Every question is of a single form, whether evidence of a stated thing is publicly available, and every cell carries a URL to an artefact that was opened plus an access date. Eleven items are tied to clause numbers: four on the interactive disclosure under Article 50(1), four on machine-readable marking and detection under Article 50(2), two on the manner and accessibility of disclosure under Article 50(5), and one on signatory status under Article 50(7), reported separately and never added to the totals. The rubric was written and signed off before any company artefact was opened, and a prior-art search ran first and is reported in full. Findings at this date: accessibility of the disclosure is unevidenced across all twelve, with no provider publishing anything that reaches the AI notice; marking across every output type a product generates is evidenced by two of twelve, with text the least covered modality and one provider documenting text marking; five of twelve evidence a disclosure at the point of interaction, of which four name themselves and one says only that it is powered by AI. A finding that emerged from method rather than design is reported separately: the interaction disclosure varies with account state, and of the products carrying such a line, no two were observed carrying it in the same state, a variation that reversed four cells during the work and is logged as a correction. Signing the Code of Practice and marking outputs are shown to be different facts, in both directions. The 50(2) column must be read against the transitional period rather than as a shortfall, which is why that explanation precedes the table. Limits are stated rather than implied: twelve products, one observer, one country, one browser, two days, and interpretive guidance that is still in draft, approved in content on 20 July 2026 and applicable only once adopted in all languages. Every correction made during the work is logged with the withdrawn reading left in place, including two that arose from citing a Guidelines paragraph before reading it. A dependency-free Python script re-prints the tables from the data and fails the file if any row carries a disallowed answer or an unsourced yes or no. Conflict of interest: this work was assisted by an AI model made by Anthropic and Anthropic products are in the scored set. The rubric was fixed before any artefact was opened, that product was scored last, every relevant cell was re-checked, and every quotation from an Anthropic document is verbatim. The results run in both directions. Independent analysis and open-science documentation only, not legal advice.