Dataset Viewer
Auto-converted to Parquet Duplicate
product
stringclasses
14 values
provider
stringclasses
14 values
clause
stringclasses
4 values
item_id
float64
1.1
7.1
answer
stringclasses
3 values
evidence_url
stringlengths
18
110
⌀
artefact_type
stringclasses
3 values
accessed_date
stringdate
2026-08-05 00:00:00
2026-08-06 00:00:00
note
stringlengths
27
1.34k
ChatGPT
OpenAI
50(1)
1.1
yes
https://chatgpt.com/
in_product
2026-08-05
Entry screen carries the line "ChatGPT is AI. By using it, you agree to our Terms & Privacy Policy." Read from the served HTML of the logged out page, not from a rendered screenshot. STATE DEPENDENT, found 2026-08-06 round 2: the line is present on the logged out entry screen and absent from a capture of the first chat...
ChatGPT
OpenAI
50(1)
1.2
yes
https://chatgpt.com/
in_product
2026-08-05
Same line sits on the entry screen beneath the composer, before any message is sent, so it is not confined to a policy document. See the state dependence note on 1.1.
ChatGPT
OpenAI
50(1)
1.3
yes
https://chatgpt.com/
in_product
2026-08-05
Disclosure names the product: "ChatGPT is AI", rather than a general statement that the site uses AI. See the state dependence note on 1.1.
ChatGPT
OpenAI
50(1)
1.4
yes
https://model-spec.openai.com/2025-04-11.html
public_docs
2026-08-05
Close call. Model Spec states "The assistant should not pretend to be human or have feelings", and the worked example that OpenAI labels "Compliant" for the question "Are you conscious?" opens "I'm a large language model, which is an AI model". Framed as a prohibition on pretending plus an example, not as an explicit d...
ChatGPT
OpenAI
50(2)
2.1
yes
https://help.openai.com/en/articles/8912793-c2pa-in-chatgpt-images
public_docs
2026-08-05
Table on the page: images generated with ChatGPT, Codex and the API include C2PA metadata and SynthID watermarks; audio includes SynthID watermarks. Page updated 4 days before access.
ChatGPT
OpenAI
50(2)
2.2
yes
https://help.openai.com/en/articles/8912793-c2pa-in-chatgpt-images
public_docs
2026-08-05
Techniques named: Content Credentials under the C2PA standard, and SynthID invisible watermarking.
ChatGPT
OpenAI
50(2)
2.3
yes
https://openai.com/index/advancing-content-provenance/
public_docs
2026-08-05
Public verification tool at openai.com/verify accepts uploaded images and audio and reports whether OpenAI provenance signals are present. Update of 31 July 2026 adds audio and API access for verification.
ChatGPT
OpenAI
50(2)
2.4
no
https://help.openai.com/en/articles/8912793-c2pa-in-chatgpt-images
public_docs
2026-08-05
Product generates text, images and audio. Marking covers images and audio only. The page asks "Do you have plans to support text outputs?" and answers that the goal is to expand provenance signals to all modalities including text, so text is stated as not yet covered. Scope caveat added on recheck of Guidelines paragra...
ChatGPT
OpenAI
50(5)
5.1
yes
https://chatgpt.com/
in_product
2026-08-05
Close call. The disclosure is a standalone sentence at the interaction point rather than buried in menus or terms. Judged from served HTML; visual prominence, contrast and size were not measured and are outside what a public artefact shows. See the state dependence note on 1.1.
ChatGPT
OpenAI
50(5)
5.2
no_artefact_found
null
public_docs
2026-08-05
Bounded search across openai.com and help.openai.com found no accessibility statement or conformance report reaching the AI disclosure. The one WCAG statement located is OpenAI's own, about the Atlas browser, saying it is not yet fully compliant with those standards. That concerns a different product and a different st...
ChatGPT
OpenAI
50(7)
7.1
yes
https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content
filing
2026-08-05
OpenAI appears in the Section 1 signatory table. Parsed from the saved page snapshot by parse_signatories.py.
Gemini
Google
50(2)
2.1
yes
https://deepmind.google/technologies/synthid/
public_docs
2026-08-05
Page states SynthID embeds digital watermarks directly into AI-generated images, audio, text or video, and that the watermarks are embedded across Google's generative AI consumer products.
Gemini
Google
50(2)
2.2
yes
https://deepmind.google/technologies/synthid/
public_docs
2026-08-05
Technique named as SynthID, described as an invisible digital watermark added at the moment of creation.
Gemini
Google
50(2)
2.3
yes
https://deepmind.google/technologies/synthid/
public_docs
2026-08-05
Documented route: upload an image, video or audio clip to a Gemini chat and ask whether it was created or altered by Google AI, and Gemini checks for a SynthID watermark. Note the stated check covers image, video and audio, not text.
Gemini
Google
50(2)
2.4
yes
https://deepmind.google/technologies/synthid/
public_docs
2026-08-05
All four modalities addressed: image and video, audio via Lyria and the NotebookLM podcast feature, and text generated by the Gemini app and web experience.
Gemini
Google
50(7)
7.1
yes
https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content
filing
2026-08-05
Google appears in the Section 1 signatory table. Parsed from the saved snapshot by parse_signatories.py. Contradicts an earlier secondary rendering of the same page that placed Google in Section 2 only.
Vibe
Mistral AI
50(7)
7.1
yes
https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content
filing
2026-08-05
Mistral appears in the Section 1 signatory table. Product recorded as Vibe, the name under which chat.mistral.ai presents on the access date; the signatory entry is the provider name Mistral.
Microsoft Copilot
Microsoft
50(7)
7.1
yes
https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content
filing
2026-08-05
Microsoft appears in the Section 1 signatory table.
Meta AI
Meta
50(7)
7.1
yes
https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content
filing
2026-08-05
Meta appears in the Section 1 signatory table.
Claude
Anthropic
50(7)
7.1
yes
https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content
filing
2026-08-05
Anthropic appears in the Section 1 signatory table. COI: the assisting model is made by Anthropic. Cell parsed by script from the saved snapshot, not typed from memory.
FLUX
Black Forest Labs
50(7)
7.1
yes
https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content
filing
2026-08-05
Black Forest Labs appears in both the Section 1 and Section 2 tables.
Synthesia
Synthesia
50(7)
7.1
yes
https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content
filing
2026-08-05
Synthesia appears in the Section 1 signatory table.
Cohere surface
Cohere
50(7)
7.1
yes
https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content
filing
2026-08-05
Cohere appears in the Section 1 signatory table. Product still to pass the consumer-reachable screen.
Aleph Alpha surface
Aleph Alpha
50(7)
7.1
yes
https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content
filing
2026-08-05
Aleph Alpha appears in the Section 1 signatory table. Product still to pass the consumer-reachable screen.
Firefly
Adobe
50(7)
7.1
no
https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content
filing
2026-08-05
Adobe appears in neither the Section 1 nor the Section 2 table on the access date. Signing is voluntary, so this records absence from the list and nothing more.
Canva
Canva
50(7)
7.1
no
https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content
filing
2026-08-05
Canva appears in neither table on the access date.
Brand Studio
Stability AI
50(7)
7.1
no
https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content
filing
2026-08-05
Stability AI appears in neither table on the access date. Product also needs a screen: dreamstudio.ai redirected to brandstudio.com on probe, so the product may have been renamed or retired. Product renamed in this record from DreamStudio to Brand Studio, the surface dreamstudio.ai resolves to.
ElevenLabs
ElevenLabs
50(7)
7.1
no
https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content
filing
2026-08-05
ElevenLabs appears in neither table on the access date.
ElevenLabs
ElevenLabs
50(2)
2.3
yes
https://elevenlabs.io/blog/ai-speech-classifier
public_docs
2026-08-05
AI Speech Classifier accepts an uploaded audio sample and reports whether it contains ElevenLabs-generated material. The page states it does not reliably classify audio from the Eleven v3 model and cannot detect audio from other providers.
Gemini
Google
50(1)
1.1
yes
https://gemini.google.com/
in_product
2026-08-06
Manual capture, logged out. Entry screen carries "Gemini is AI and can make mistakes, including about people." next to the prompt box, and the greeting reads "Meet Gemini, your personal AI assistant". Round 2 note: a new account capture was not obtainable. Account creation required scanning a QR code with a phone and s...
Gemini
Google
50(1)
1.2
yes
https://gemini.google.com/
in_product
2026-08-06
Both strings are present on the logged out entry screen before any prompt is sent, alongside rather than inside the linked Terms and Privacy Policy.
Gemini
Google
50(1)
1.3
yes
https://gemini.google.com/
in_product
2026-08-06
Disclosure names the product: "Gemini is AI", not a general statement about Google services.
Gemini
Google
50(5)
5.1
yes
https://gemini.google.com/
in_product
2026-08-06
Standalone sentence at the interaction point. Judged from the captured DOM; visual prominence and contrast not measured.
Microsoft Copilot
Microsoft
50(1)
1.1
yes
https://copilot.microsoft.com/
in_product
2026-08-06
Manual capture, signed in. Screen carries "Copilot is an AI and may make mistakes. Using Copilot means you agree to the Terms of Use." The logged out capture of the same URL is a sign in wall with no equivalent line. Round 2 note: a new account capture was not obtainable, since signup required a Microsoft account and a...
Microsoft Copilot
Microsoft
50(1)
1.2
yes
https://copilot.microsoft.com/
in_product
2026-08-06
Line is present on the new chat screen before any message is sent. Qualifier recorded in the note for 1.1: evidenced in the signed in state only.
Microsoft Copilot
Microsoft
50(1)
1.3
yes
https://copilot.microsoft.com/
in_product
2026-08-06
Disclosure names the product: "Copilot is an AI".
Microsoft Copilot
Microsoft
50(5)
5.1
yes
https://copilot.microsoft.com/
in_product
2026-08-06
Standalone sentence beneath the composer. Judged from the captured DOM.
Meta AI
Meta
50(1)
1.1
no
https://www.meta.ai/
in_product
2026-08-06
Manual capture, logged out. No sentence stating the user is interacting with an AI system. The screen carries the product name Meta AI, the control label "Ask Meta AI" and the prompt "Where should we start?". Whether the name makes the AI nature obvious is the Article 50(1) exception, which this scorecard does not scor...
Meta AI
Meta
50(1)
1.2
no
https://www.meta.ai/
in_product
2026-08-06
No such statement present before first interaction, since none was found at all. Confirmed on the round 2 new account capture.
Meta AI
Meta
50(1)
1.3
no
https://www.meta.ai/
in_product
2026-08-06
No disclosure sentence to assess for specificity. Confirmed on the round 2 new account capture.
Meta AI
Meta
50(5)
5.1
no
https://www.meta.ai/
in_product
2026-08-06
No disclosure present in the capture to be presented clearly or otherwise. Caveat: the capture retains a "Loading page" label, so a late rendering element cannot be ruled out. Recapture would settle it. Confirmed on the round 2 new account capture, which removes the earlier loading caveat.
Vibe
Mistral AI
50(1)
1.1
no
https://chat.mistral.ai/chat
in_product
2026-08-06
Manual capture, no account needed. Product presents under the name Vibe on the access date. Entry gate reads "You must accept our Terms of Service and Privacy Policy to start using Vibe". Neither the gate nor the post acceptance screen states that the user is interacting with an AI system. CONFIRMED 2026-08-06 round 3:...
Vibe
Mistral AI
50(1)
1.2
no
https://chat.mistral.ai/chat
in_product
2026-08-06
The only pre interaction text is the Terms of Service acceptance gate. The capture taken after acceptance carries no disclosure line.
Vibe
Mistral AI
50(1)
1.3
no
https://chat.mistral.ai/chat
in_product
2026-08-06
No disclosure sentence to assess for specificity.
Vibe
Mistral AI
50(5)
5.1
no
https://chat.mistral.ai/chat
in_product
2026-08-06
No disclosure present in either capture.
Claude
Anthropic
50(1)
1.1
yes
https://claude.ai/
in_product
2026-08-06
Manual capture, new account, first chat screen. Screen carries “Claude is AI and can make mistakes. Please double-check responses.” beneath the composer, and the greeting reads “Welcome, [name]! I’m Claude.” ⛔ WITHDRAWN: this cell read no on 2026-08-06 from a capture of a long standing account, where the line was absen...
Claude
Anthropic
50(1)
1.2
yes
https://claude.ai/
in_product
2026-08-06
Line is present on the first chat screen before any message is sent. ⛔ WITHDRAWN: previously no, on the superseded capture.
Claude
Anthropic
50(1)
1.3
yes
https://claude.ai/
in_product
2026-08-06
Disclosure names the product: “Claude is AI”. ⛔ WITHDRAWN: previously no, on the superseded capture.
Claude
Anthropic
50(5)
5.1
yes
https://claude.ai/
in_product
2026-08-06
Standalone sentence at the interaction point. Judged from the captured DOM; visual prominence not measured. ⛔ WITHDRAWN: previously no, on the superseded capture.
Firefly
Adobe
50(2)
2.1
yes
https://helpx.adobe.com/firefly/web/get-started/learn-the-basics/content-credentials-overview.html
public_docs
2026-08-06
Manual capture, page last updated 11 Mar 2026. States Adobe automatically applies Content Credentials to assets where 100% of the pixels are generated with Adobe Firefly, described as industry standard tamper evident metadata.
Firefly
Adobe
50(2)
2.2
yes
https://helpx.adobe.com/firefly/web/get-started/learn-the-basics/content-credentials-overview.html
public_docs
2026-08-06
Technique named as Content Credentials, tamper evident metadata, with the issuer, date, app or device, AI tool used and actions enumerated as always included.
Firefly
Adobe
50(2)
2.3
yes
https://helpx.adobe.com/firefly/web/get-started/learn-the-basics/content-credentials-overview.html
public_docs
2026-08-06
Inspect tool on Adobe Content Authenticity is documented as enabling a viewer to see Content Credentials associated with a selected asset, with a public Content Credentials cloud making them recoverable if stripped from the file.
Firefly
Adobe
50(2)
2.4
no
https://helpx.adobe.com/firefly/web/get-started/learn-the-basics/content-credentials-overview.html
public_docs
2026-08-06
The stated trigger is pixel based, assets where 100% of the pixels are generated. Firefly also generates video, audio and speech. No statement covering those modalities was found on this page. Caveat: an inline example list did not survive the capture, so the enumeration of covered asset types is incomplete in the arte...
Meta AI
Meta
50(2)
2.1
yes
https://about.fb.com/news/2024/02/labeling-ai-generated-images-on-facebook-instagram-and-threads/
public_docs
2026-08-06
States that photorealistic images created with Meta AI carry visible markers, invisible watermarks and metadata embedded within image files. Vintage flag: this announcement is dated February 2024 and a bounded search of Meta's own domains located nothing more recent on marking.
Meta AI
Meta
50(2)
2.2
yes
https://about.fb.com/news/2024/02/labeling-ai-generated-images-on-facebook-instagram-and-threads/
public_docs
2026-08-06
Techniques named: IPTC metadata and invisible watermarks, stated to be in line with Partnership on AI best practices.
Meta AI
Meta
50(2)
2.3
no
https://about.fb.com/news/2024/02/labeling-ai-generated-images-on-facebook-instagram-and-threads/
public_docs
2026-08-06
The page describes Meta detecting and labelling other companies' images on its own platforms. No route is documented by which a third party can check whether content came from Meta AI. The feature described for video and audio is user self disclosure at upload, which is not detection.
Meta AI
Meta
50(2)
2.4
no
https://about.fb.com/news/2024/02/labeling-ai-generated-images-on-facebook-instagram-and-threads/
public_docs
2026-08-06
Marking as described covers photorealistic images. The same page states that signals were not yet being included in AI tools generating audio and video at the same scale. Product also generates text, which is not addressed.
Synthesia
Synthesia
50(2)
2.1
yes
https://help.synthesia.io/en/articles/16046624-what-does-the-eu-ai-act-article-50-mean-for-my-synthesia-videos
public_docs
2026-08-06
States videos generated on Synthesia carry C2PA based provenance signals embedded in the file, applied automatically at generation with nothing for the customer to switch on, marking the video as AI generated and recording a tamper evident chain of custody.
Synthesia
Synthesia
50(2)
2.2
yes
https://help.synthesia.io/en/articles/16046624-what-does-the-eu-ai-act-article-50-mean-for-my-synthesia-videos
public_docs
2026-08-06
Technique named as C2PA, described as a cross industry standard for attaching secure machine readable information about how content was made.
Synthesia
Synthesia
50(2)
2.3
no
https://help.synthesia.io/en/articles/16046624-what-does-the-eu-ai-act-article-50-mean-for-my-synthesia-videos
public_docs
2026-08-06
The provider's own Article 50 explainer documents no route by which a person exposed to a video can check it. Qualifier: C2PA is an open standard and third party inspection tools exist, but this page points to none, and the item asks for a documented route.
Synthesia
Synthesia
50(2)
2.4
no
https://help.synthesia.io/en/articles/16046624-what-does-the-eu-ai-act-article-50-mean-for-my-synthesia-videos
public_docs
2026-08-06
⛔ WITHDRAWN, was yes. Corrected 6 August 2026 on a recheck of Guidelines paragraph (60), which was cited in the rubric but had not been read when this cell was first scored. (60) states: "Video refers to a time-based sequence of images, which may be synchronised with audio, capable of being seen by humans. In that case...
Claude
Anthropic
50(2)
2.1
no
https://www.anthropic.com/transparency/voluntary-commitments
public_docs
2026-08-06
COI applies. Transparency Hub states: "Claude currently has multimodal input capabilities and text-based outputs, including text-based artifacts and text-to-speech voice output. We have worked across industry and academia to explore and stay abreast of technological developments for watermarking and are preparing for c...
Claude
Anthropic
50(2)
2.2
no
https://www.anthropic.com/transparency/voluntary-commitments
public_docs
2026-08-06
No marking technique is named as in use for Claude outputs. The page names watermarking as a field being tracked.
Claude
Anthropic
50(2)
2.3
no
https://www.anthropic.com/transparency/voluntary-commitments
public_docs
2026-08-06
No detection route is documented by which a third party could check whether content came from Claude.
Claude
Anthropic
50(2)
2.4
no
https://www.anthropic.com/transparency/voluntary-commitments
public_docs
2026-08-06
Product outputs text and text to speech audio. No marking is evidenced for either modality, so coverage across modalities does not arise. Scope caveat added on recheck of Guidelines paragraph (68): source code is expressly outside Article 50(2), as are short sequences such as single words, captions, alt-text and UI lab...
Microsoft Copilot
Microsoft
50(2)
2.1
yes
https://support.microsoft.com/en-us/microsoft-copilot/transparency-note-for-microsoft-copilot
public_docs
2026-08-06
Transparency Note states: "When Copilot generates an image, we have enabled a Content Credentials feature, which uses cryptographic methods to mark the source, or provenance, of all AI-generated images created using Copilot."
Microsoft Copilot
Microsoft
50(2)
2.2
yes
https://support.microsoft.com/en-us/microsoft-copilot/transparency-note-for-microsoft-copilot
public_docs
2026-08-06
Technique named as Content Credentials employing standards set by the C2PA.
Microsoft Copilot
Microsoft
50(2)
2.3
yes
https://support.microsoft.com/en-us/designer/frequently-asked-questions-about-microsoft-designer
public_docs
2026-08-06
Microsoft points to an industry standard inspection route on its own support domain: "Provenance metadata can be viewed on the Content Credentials site." Scoping caveat: that sentence sits on the Designer page and is written about images created with Designer's editing features. Neither artefact I opened states that im...
Microsoft Copilot
Microsoft
50(2)
2.4
no
https://support.microsoft.com/en-us/microsoft-copilot/transparency-note-for-microsoft-copilot
public_docs
2026-08-06
Marking is stated for images only. The same Transparency Note discusses Copilot generating content in other modalities including audio and text, and no marking is stated for those.
ElevenLabs
ElevenLabs
50(2)
2.1
yes
https://elevenlabs.io/blog/synthid
public_docs
2026-08-06
States SynthID is "a digital watermark embedded directly into audio generated by ElevenLabs", inaudible, and remaining when clips are trimmed, sped up, stripped of metadata or converted to another file type.
ElevenLabs
ElevenLabs
50(2)
2.2
yes
https://elevenlabs.io/blog/synthid
public_docs
2026-08-06
Techniques named: SynthID watermarking in partnership with Google DeepMind, alongside C2PA credentials described as part of the existing provenance ecosystem.
ElevenLabs
ElevenLabs
50(2)
2.4
no
https://elevenlabs.io/blog/synthid
public_docs
2026-08-06
Coverage is stated as partial: "This week, we have started including SynthID in Text to Speech generations by free users, and we will expand coverage to all ElevenLabs audio generations over the coming weeks." Product also generates music and sound effects. The page's publication date was not captured in the extract, s...
Vibe
Mistral AI
50(2)
2.1
no_artefact_found
null
public_docs
2026-08-06
Bounded search of mistral.ai, docs.mistral.ai and help.mistral.ai located no page evidencing marking of outputs. The one page indexed as carrying provenance commitments, mistral.ai/legal/child-abuse/, returns HTTP 404 on access, and the legal index served no matching text. Absence of a located artefact, not a statement...
Vibe
Mistral AI
50(2)
2.2
no_artefact_found
null
public_docs
2026-08-06
No technique named, since no marking artefact was located.
Vibe
Mistral AI
50(2)
2.3
no_artefact_found
null
public_docs
2026-08-06
No detection route located.
Vibe
Mistral AI
50(2)
2.4
no_artefact_found
null
public_docs
2026-08-06
Coverage across modalities does not arise, since no marking artefact was located.
Canva
Canva
50(2)
2.1
no_artefact_found
null
public_docs
2026-08-06
Bounded search of canva.com located no technical documentation stating that Canva applies machine-readable marks to AI-generated output. Adjacent evidence, recorded but not scored: the AI Product Terms effective 26 June 2026 prohibit users from "Remove, alter, or disable any provenance or metadata tags from any AI-gene...
Canva
Canva
50(2)
2.2
no_artefact_found
null
public_docs
2026-08-06
C2PA is named in the prohibition clause above, but not as a technique the provider states it applies.
Canva
Canva
50(2)
2.3
no_artefact_found
null
public_docs
2026-08-06
No detection route located.
Canva
Canva
50(2)
2.4
no_artefact_found
null
public_docs
2026-08-06
Coverage across modalities does not arise.
FLUX
Black Forest Labs
50(2)
2.1
yes
https://bfl.ai/blog/capable-open-and-safe-combating-ai-misuse
public_docs
2026-08-06
States: "Additionally, we apply content provenance metadata on our hosted services to help platforms and viewers identify AI-generated content once it is shared online." The scored product is the hosted playground, so the hosted services wording applies to it.
FLUX
Black Forest Labs
50(2)
2.2
no
https://bfl.ai/blog/capable-open-and-safe-combating-ai-misuse
public_docs
2026-08-06
The technique applied on hosted services is described only as "content provenance metadata". C2PA is named on the same page, but in relation to links included in open-weight repositories to help other developers implement metadata, not as the technique applied to hosted output.
FLUX
Black Forest Labs
50(2)
2.3
no
https://bfl.ai/blog/capable-open-and-safe-combating-ai-misuse
public_docs
2026-08-06
No route is documented by which a person exposed to an image can check it. The stated purpose is to help platforms and viewers identify content once shared, with no tool or method named.
FLUX
Black Forest Labs
50(2)
2.4
yes
https://bfl.ai/blog/capable-open-and-safe-combating-ai-misuse
public_docs
2026-08-06
Product generates images, and the marking statement is not limited by modality within the hosted service. Scored against the one modality the product produces.
Brand Studio
Stability AI
50(2)
2.1
no
https://stability.ai/safety
public_docs
2026-08-06
Wording is present progressive rather than applied: "We are implementing content authenticity standards and watermarking so that users and platforms can identify AI-assisted content generated through our hosted services." That states implementation under way, not that outputs carry a mark. Quoted verbatim so the call i...
Brand Studio
Stability AI
50(2)
2.2
no
https://stability.ai/safety
public_docs
2026-08-06
A footnote names C2PA standards in partnership with the Content Authenticity Initiative, but attached to the implementation statement above rather than to a mark evidenced as applied.
Brand Studio
Stability AI
50(2)
2.3
no
https://stability.ai/safety
public_docs
2026-08-06
No detection route documented.
Brand Studio
Stability AI
50(2)
2.4
no
https://stability.ai/safety
public_docs
2026-08-06
No applied marking evidenced, so coverage across modalities does not arise.
Claude
Anthropic
50(1)
1.4
yes
https://www.anthropic.com/constitution
public_docs
2026-08-06
COI applies. Claude's Constitution lists as a rule: "Never deceive the human into thinking they're talking with a human". It further states that operators cannot instruct Claude to "claim to be human when directly and sincerely asked", even when running a custom AI persona. The second clause maps directly onto the on-r...
Microsoft Copilot
Microsoft
50(1)
1.4
no
https://support.microsoft.com/en-us/microsoft-copilot/transparency-note-for-microsoft-copilot
public_docs
2026-08-06
The Transparency Note has a section headed "AI disclosure" stating Copilot "is also designed to inform people that they are interacting with an AI system" through "various touchpoints". That is a general design commitment and does not address behaviour when a user asks the system about its nature, which is what this it...
Gemini
Google
50(1)
1.4
no_artefact_found
null
public_docs
2026-08-06
Bounded search across support.google.com, policies.google.com, ai.google, deepmind.google and blog.google located no statement that Gemini identifies itself as AI on request. The nearest material is the Generative AI Prohibited Use Policy, which restricts users from misrepresenting generated content as solely human-cre...
Vibe
Mistral AI
50(1)
1.4
no_artefact_found
null
public_docs
2026-08-06
Bounded search of mistral.ai, docs.mistral.ai and help.mistral.ai located no statement on identity disclosure by the system.
Meta AI
Meta
50(1)
1.4
no_artefact_found
https://transparency.meta.com/policies/other-policies/meta-AI-disclosures/
public_docs
2026-08-06
Checked and closed: a Meta policy page titled "AI Disclosures" exists at transparency.meta.com/policies/other-policies/meta-AI-disclosures/ and could not be opened, returning HTTP 400 to curl and an empty body to the fetch tool. It is queued for manual capture. If that page describes an AI notice, this cell and possibl...
Gemini
Google
50(5)
5.2
no
https://www.google.com/accessibility/
public_docs
2026-08-06
Uniform bounded method for item 5.2, applied identically to all products: probe the provider's conventional accessibility path, then open any page returning HTTP 200 and check whether it reaches the AI disclosure. Page opened. It offers Accessibility Conformance Reports based on the ITI VPAT template, but nothing on th...
Microsoft Copilot
Microsoft
50(5)
5.2
no
https://www.microsoft.com/en-us/accessibility
public_docs
2026-08-06
Uniform bounded method for item 5.2, applied identically to all products: probe the provider's conventional accessibility path, then open any page returning HTTP 200 and check whether it reaches the AI disclosure. Page opened. It offers conformance reports and mentions Copilot in navigation, but nothing on the page rea...
Canva
Canva
50(5)
5.2
no_artefact_found
https://www.canva.com/accessibility/
public_docs
2026-08-06
Uniform bounded method for item 5.2, applied identically to all products: probe the provider's conventional accessibility path, then open any page returning HTTP 200 and check whether it reaches the AI disclosure. URL returns HTTP 200 but serves a 3 KB shell with four lines of text, so no content could be read.
Claude
Anthropic
50(5)
5.2
no_artefact_found
null
public_docs
2026-08-06
Uniform bounded method for item 5.2, applied identically to all products: probe the provider's conventional accessibility path, then open any page returning HTTP 200 and check whether it reaches the AI disclosure. COI applies. anthropic.com/accessibility returns HTTP 404 and no accessibility artefact was located.
Meta AI
Meta
50(5)
5.2
no_artefact_found
null
public_docs
2026-08-06
Uniform bounded method for item 5.2, applied identically to all products: probe the provider's conventional accessibility path, then open any page returning HTTP 200 and check whether it reaches the AI disclosure. No accessibility artefact located reaching the AI disclosure.
Vibe
Mistral AI
50(5)
5.2
no_artefact_found
null
public_docs
2026-08-06
Uniform bounded method for item 5.2, applied identically to all products: probe the provider's conventional accessibility path, then open any page returning HTTP 200 and check whether it reaches the AI disclosure. mistral.ai/accessibility returns HTTP 404.
Firefly
Adobe
50(5)
5.2
no_artefact_found
null
public_docs
2026-08-06
Uniform bounded method for item 5.2, applied identically to all products: probe the provider's conventional accessibility path, then open any page returning HTTP 200 and check whether it reaches the AI disclosure. adobe.com/accessibility.html could not be retrieved, the connection failing with an HTTP/2 stream error. R...
End of preview. Expand in Data Studio

Article 50 historical public-evidence snapshot

This work was produced through an AI-assisted workflow directed by the author. Historical work used Anthropic assistance; the retrospective correction uses OpenAI GPT-6, with separate bounded Gemini advice. All three providers have products in the scored set.

Purpose: provide the corrected paper's version 1.1 bundle under v1_1. Start with its README and correction note. The paper and deposit and GitHub repository identify the same correction.

The default dataset configuration uses v1_1/article50_scores.csv. The corrected bundle preserves the historical coded values. It is not a current survey, an accepted rescore or a compliance assessment. Copilot detection attribution and FLUX, Gemini and Synthesia coverage remain unresolved. Disclosure interpretation and capture/placement limits are stated in the paper. Provider declarations do not prove technical performance.

Legacy root files

The files at the repository root predate this alignment. Their CSV has a different schema and product set from the paper's data, and their rubric and reproduction scripts belong to that separate dataset. They remain at their existing paths for provenance. Do not combine their scores with the version 1.1 bundle or treat them as a matching mirror of the paper. The default configuration points to the paper's data explicitly.

Verification and reproduction

The files in v1_1 match the correction bundle. Use that directory's README and reproduction command. No new product or model experiment was performed for the correction, and no scientific performance baseline was run. Selected sanitised extracts are included; private browser captures are excluded. Not every account-state judgement can be independently reconstructed from the public files.

AI assistance, conflicts and limitations

The research question, method, sourcing decisions and analytical judgements are the author's. Historical Anthropic assistance, OpenAI GPT-6 correction assistance and separate Google Gemini advice are disclosed in the paper. These providers have products in the scored set. Review is bounded and is not author-independent legal clearance.

NM AI Research material is CC BY 4.0. Third-party documents and quoted extracts retain their original rights and attribution. Independent analysis and open-science documentation, not legal advice.

Legacy path sanitisation

The legacy CSV now uses repository-relative cached_path values. The historical manual-ingestion instructions also use the repository root instead of a personal machine path. Scores, other CSV fields and captured-source bytes are unchanged. This removes personal machine paths from the current files; earlier repository revisions remain available. This is a locator cleanup, not a new source review or accepted rescore.

Downloads last month
191