How we measured it
The rule is fixed before the data is read — that is the credibility.
1 · The question
On a product's public-facing surface, can an ordinary observer find the transparency signals Article 50 is concerned with — and is any of it independently verifiable? We assess only what we can observe. Absence of an observable signal is reported as exactly that — not observed on the public surface as of the scan date — never as "non-compliant."
2 · Sample frame & inclusion rule (fixed before scanning)
- Target N ≈ 100 of the most-used consumer-facing AI products available to EU users.
- Inclusion rule: consumer-facing AI product · accessible to an EU user without a paid enterprise contract · appears in at least one cited public ranking of widely-used AI products.
- Stratification (for fair cuts, not selection): provider HQ region (EU / US / other) and product category.
- No cherry-picking: the list is frozen and published in full before grading.
3 · The rubric — six signals, each 0 / 1 / 2
0 not observed · 1 partial/unclear · 2 present & clear.
| Signal | What "2" looks like |
|---|---|
| R1 AI-interaction notice | A user is plainly told they are interacting with an AI, at the point of interaction. |
| R2 Clear & specific | The notice is specific and understandable, not buried in a ToS link. |
| R3 Synthetic-content labelling | AI-generated/manipulated image, audio, or video carries a visible or embedded marking. |
| R4 Machine-detectable | The marking is detectable by a machine (metadata / provenance), not only by eye. |
| R5 Record of what the AI did | A user-facing record/trace of the AI output exists. |
| R6 Independently verifiable | Any of the above can be verified by a third party without trusting the provider. The strategic finding — almost nothing scores here. |
4 · Grading protocol
- Automated, reproducible scan. Each product is measured by LedgerProof's automated readiness check of its public-facing surface against the Article 50 signals — the same free tool anyone can run on any product, so every figure here can be independently reproduced.
- Public surface only: no bulk harvesting, no access behind logins or paywalls, no circumvention of access controls or ToS.
- Honest about scope: an automated check reads what is observable on the public page at a point in time; a signal not observed is reported as "not observed," never as a claim about internal practice.
- Aggregate figures report the sample size (n) and the date; products that block automated review are excluded and counted. We don't over-state precision.
5 · Scoring → grade (A–D, never "F")
Total (0–12) maps to a published band: A / B / C / D. We use no "F" — the framing is readiness and the path to improve it, not condemnation. Cut-offs are fixed before scanning and published with the report.
6 · Named-product policy & right-of-reply
- Aggregate-first — the headline is about the landscape, not individuals.
- A product is named only where the signal is plainly observable, we'd publish the same about ourselves, and it's framed as readiness and corrigible.
- Right-of-reply + a free re-scan to every named product before publication. Changed something? We re-scan and update. The point is to raise the floor, fairly — not to embarrass.
7 · Limitations (stated in the report itself)
- A point-in-time, public-surface observation; products change.
- Not exhaustive and not a legal/compliance determination. Absence of an observable signal ≠ absence of the practice.
- Our interest is disclosed: we build the tooling that makes R6 possible. The data stands alone — anyone can re-run it.
8 · Reproducibility
The frozen product list, the rubric, and the band cut-offs are all published with the report, alongside the realized sample size and scan date. The free tool runs the same public-surface check, so any reader — journalist, academic, or the product itself — can reproduce a result. We graded ourselves first: see our own readiness →