AI Content Transparency and Marking Plan: [Organisation] [Product or service]
Plan for meeting transparency duties for AI interaction and
AI-generated or manipulated content: your role (provider or
deployer), the systems and outputs in scope, the four Article 50
obligations of the EU Artificial Intelligence Act (Regulation (EU)
2024/1689 as amended by Regulation (EU) 2026/1744), the technical
marking approach (C2PA Content Credentials 2.4, watermarking,
metadata, fingerprinting), robustness and interoperability testing,
user-facing labels, exceptions, China labelling rules for
China-facing services, and records. It is organised to be read with
the Commission's Code of Practice on Transparency of AI-generated
Content (final, June 2026) and the Commission guidelines on Article 50
(July 2026). Replace every [bracketed] value and every row marked
Example. Article 50 applies from 2 August 2026; for generative
systems placed on the market before that date, the Article 50(2)
marking duty applies from 2 December 2026 (Article 111(4), inserted
by Regulation (EU) 2026/1744); check the consolidated text. This is a documentation format, not
legal advice.
Document control
| Field | Value |
|---|
| Document ID | [TRN-001] |
| Organisation | [Legal name, address] |
| Product or service | [Name, version] |
| Role under Article 50 | [Provider / Deployer / Both, see section 1] |
| Markets | [EU / China / Korea / California / other] |
| Code of Practice signatory | [Yes, date signed / No, other adequate means] |
| Version | [0.1] |
| Status | [Draft / In review / Approved / Superseded] |
| Owner | [@transparency-owner] |
| Approvers/reviewers | [@legal], [@product-lead], [@ml-lead], [@security-lead], [@accessibility-lead] |
| Last updated | YYYY-MM-DD |
| Next review | YYYY-MM-DD |
| Classification | [Internal] |
| Related documents | [AI use case intake], [AI system card], [Launch review], [Applicability register] |
Approvals
| Role | Name | Decision | Date |
|---|
| Product owner | [@name] | [Approve / Approve with conditions / Reject] | YYYY-MM-DD |
| Legal counsel | [@name] | [Decision] | YYYY-MM-DD |
| ML or platform engineering lead | [@name] | [Decision] | YYYY-MM-DD |
| Trust and safety lead | [@name] | [Decision] | YYYY-MM-DD |
Revision history
| Version | Date | Author | Change summary |
|---|
| 0.1 | YYYY-MM-DD | [@name] | First draft |
| 1.0 | YYYY-MM-DD | [@name] | Example: Approved before the 2 August 2026 application date |
| 1.1 | YYYY-MM-DD | [@name] | Example: Added audio watermark after robustness test TST-AUD-003 failed for metadata only |
Contents
1. Role and scope
1.1 Role determination
| System | Who develops it | Under whose name is it offered | Who uses it on people or content | Our role |
|---|
| SYS-001 (Example) Support chat assistant | Built in-house on a third-party GPAI model | Ours | Our customers | Provider (Art. 50(1), 50(2)) |
| SYS-002 (Example) Marketing image generator | Third-party SaaS | Vendor's | Our marketing team publishes outputs | Deployer (Art. 50(4)) |
| SYS-003 (Example) In-app image generation feature | In-house pipeline on an open-weight model | Ours | Our users | Provider (Art. 50(2)); users may be deployers |
1.2 Scope statement
[State what this plan covers and what it does not.] Example: This plan
covers all generative and conversational features of [Product name]
offered to users in the EU and China, and all AI-generated content the
company publishes on its own channels. It does not cover internal-only
tools that produce no output shown to people outside the company, which
are tracked in the AI inventory.
1.3 Responsibilities
| Area | Responsible | Accountable | Consulted |
|---|
| Interaction disclosure (50(1)) | Product design | [@product-lead] | Legal, accessibility |
| Machine-readable marking (50(2)) | ML platform | [@ml-lead] | Security |
| Emotion recognition or biometric notice (50(3)) | Product owner of the feature | [@product-lead] | DPO |
| Deepfake and public-interest text labels (50(4)) | Content and communications | [@comms-lead] | Legal, editorial |
| China labelling | China operations | [@china-ops-lead] | Local counsel |
2. Systems and outputs in scope
| ID | System | Modality | Channel | Volume per month | Generative or manipulative? | Article 50 paragraphs |
|---|
| OUT-001 (Example) | SYS-001 support assistant | Text, chat | Web widget, mobile app | 1.2 million messages | Generative | 50(1); 50(2) text |
| OUT-002 (Example) | SYS-003 image feature | Image (PNG, JPEG, WebP) | App, download, API | 400,000 images | Generative | 50(2) |
| OUT-003 (Example) | SYS-003 image feature | Image edit of user photo | App | 150,000 edits | Manipulative | 50(2); 50(4) if realistic person |
| OUT-004 (Example) | SYS-004 voice agent | Synthetic speech | Phone line | 30,000 calls | Generative | 50(1); 50(2) audio |
| OUT-005 (Example) | SYS-002 vendor tool | Image and video | Social media posts | 60 posts | Generative | 50(4) deployer |
Outputs explicitly out of scope: [Example: code completions in the
internal IDE plug-in (internal only, source code is listed by the
Commission as not requiring marking); short numeric outputs of the
pricing model.]
3. Article 50 obligations
3.1 Article 50(1): interaction disclosure (provider)
Providers must design systems that interact directly with natural
persons so that the people concerned are informed that they are dealing
with an AI system, unless that is obvious to a reasonably well-informed,
observant and circumspect person given the circumstances and context.
A narrow exception exists for systems authorised by law to detect,
prevent, investigate or prosecute criminal offences, subject to
safeguards, unless they are available for the public to report a
criminal offence.
| System | Applies? | Obvious from context? | Control | Evidence |
|---|
| SYS-001 (Example) | Yes | No: the chat window looks like human live chat | Persistent "AI assistant" header plus first message disclosure | Screenshot EV-001 |
| SYS-004 (Example) | Yes | No: synthetic voice is natural | Spoken disclosure in the first seconds of the call | Call script EV-002 |
3.2 Article 50(2): machine-readable marking of synthetic content (provider)
Providers of systems, including general-purpose AI systems, that
generate synthetic audio, image, video or text content must ensure the
outputs are marked in a machine-readable format and detectable as
artificially generated or manipulated. The technical solutions must be
effective, interoperable, robust and reliable as far as technically
feasible, taking into account the type of content, implementation cost
and the generally acknowledged state of the art. The duty does not
apply to assistive functions for standard editing, to systems that do
not substantially alter the input data or its semantics, or where use is
authorised by law for criminal law enforcement.
| Item | Value |
|---|
| Application date | 2 August 2026; systems placed on the market before that date: 2 December 2026 (Article 111(4), inserted by Regulation (EU) 2026/1744) |
| Modalities marked | [Text / Image / Audio / Video] |
| Techniques per modality | See section 4 |
| Detection interface | [Public verification page / API / partner tool] |
| Owner | [@ml-lead] |
3.3 Article 50(3): emotion recognition and biometric categorisation (deployer)
Deployers of an emotion recognition system or a biometric
categorisation system must inform the natural persons exposed to it of
its operation and must process personal data in line with the GDPR,
Regulation (EU) 2018/1725 and Directive (EU) 2016/680 as applicable. A
law enforcement exception applies where permitted by law. Note that
emotion recognition in the workplace and in education is a prohibited
practice under Article 5(1)(f), except for medical or safety reasons.
| System | Applies? | Notice method | Lawful basis reference | Evidence |
|---|
| [None in scope] (Example) | No | n/a | n/a | Use case intake record AI-UC-2026-014 confirms no such feature |
3.4 Article 50(4): deep fakes and AI-generated public-interest text (deployer)
Deployers of a system that generates or manipulates image, audio or
video content constituting a deep fake must disclose that the content
has been artificially generated or manipulated. Where the content is
part of an evidently artistic, creative, satirical, fictional or
analogous work, the disclosure is limited to disclosing the existence
of such content in a way that does not hamper the display or enjoyment
of the work. Deployers of a system that generates or manipulates text
published to inform the public on matters of public interest must
disclose that it was artificially generated or manipulated, unless the
text has undergone human review or editorial control and a natural or
legal person holds editorial responsibility for the publication (or the
use is authorised by law for criminal law enforcement).
| Output | Deep fake or public-interest text? | Editorial control? | Label | Evidence |
|---|
| OUT-005 (Example) Campaign video with a synthetic spokesperson | Deep fake (realistic person who does not exist saying things) | n/a | On-screen "AI-generated" label for the full duration and in the caption | EV-010 |
| OUT-006 (Example) AI-drafted market update on the company blog | Public-interest text (financial information to the public) | Yes: named editor reviews and approves every post | No AI label required; editorial record kept | Editorial log EV-011 |
3.5 Article 50(5): timing, form and accessibility
| Requirement | Our design | Evidence |
|---|
| Clear and distinguishable | Example: Label uses text plus icon, minimum 4.5:1 contrast | Design spec EV-020 |
| At the latest at first interaction or exposure | Example: Disclosure appears before the first user message is sent | EV-001 |
| Accessibility | Example: Label exposed to screen readers via accessible name; audio disclosure for voice | Accessibility test EV-021 |
4. Technical marking approach
4.1 Technique matrix
| Modality | Signed metadata (C2PA) | Imperceptible watermark | Fingerprint or hash registry | Visible label | Notes |
|---|
| Image (Example) | Yes: manifest with c2pa.actions (c2pa.created), digital source type for trained algorithmic media, c2pa.ai-disclosure assertion | Yes: [vendor or in-house] watermark, linked as a soft binding | Yes: perceptual hash stored 24 months | Optional in app, required on deep fakes | Manifest also published to a manifest repository for recovery |
| Video (Example) | Yes: manifest per file | Yes: frame-level watermark | Yes | Required on deep fakes | Test re-encoding at three bitrates |
| Audio (Example) | Yes where the container supports it | Yes: audio watermark | Yes | Spoken disclosure for calls | Metadata alone fails on telephony |
| Text (Example) | Not applicable to plain text in chat; manifest for exported documents | [Statistical text watermark / not used, with justification] | Yes: output log with hash for verification requests | Interface label | Text marking is technically immature; document state of the art |
4.2 C2PA Content Credentials profile
| Item | Value |
|---|
| Specification version | C2PA Technical Specification 2.4 (April 2026); check c2pa.org for later versions |
| SDK and version | [c2pa-rs / c2pa-node / other, version] |
| Signing certificate | [X.509 certificate, issuing CA, conformance status] |
| Assertions written | Example: c2pa.actions with c2pa.created, digital source type, c2pa.ai-disclosure, c2pa.soft-binding |
| Ingredients | Example: user-uploaded source image recorded as ingredient for edits |
| Key storage | [HSM / cloud KMS], rotation every [12] months |
| Manifest storage | [Embedded / embedded plus remote repository] |
4.3 Watermarking and fingerprinting
| Item | Value |
|---|
| Watermark technique | [Name, vendor, version] |
| Payload | Example: 64-bit identifier linking to the provider and a generation record |
| Detector access | Example: internal API; public verification page; detector provided to partners on request |
| Fingerprint method | [Perceptual hash / embedding fingerprint] |
| Retention of fingerprints | [Period, legal basis] |
| Tamper prohibition | Example: acceptable use policy forbids removing or altering marks |
4.4 Pipeline placement
Marks are applied server-side before the output leaves the trust
boundary, so that no client path can deliver an unmarked file. Bulk and
API outputs pass through the same pipeline.
5. Robustness, interoperability and detection testing
5.1 Test plan
| ID | Modality | Transformation | Mark checked | Target | Result | Date |
|---|
| TST-IMG-001 (Example) | Image | JPEG recompression at quality 70 | Watermark | Detected in at least 95% of 1,000 samples | 98.4% | YYYY-MM-DD |
| TST-IMG-002 (Example) | Image | Screenshot and crop to 60% | Watermark, fingerprint | At least 90% | 91.0% | YYYY-MM-DD |
| TST-IMG-003 (Example) | Image | Upload to [social platform] and download | C2PA manifest | Recorded (stripping expected) | Stripped; recovered from repository via soft binding | YYYY-MM-DD |
| TST-AUD-003 (Example) | Audio | Telephony codec (narrowband) | Watermark | At least 90% | 93.2% | YYYY-MM-DD |
| TST-VID-001 (Example) | Video | Transcode to 720p at low bitrate | Watermark | At least 90% | 89.1% (below target, ticket opened) | YYYY-MM-DD |
| TST-INT-001 (Example) | Image | Validation with an independent C2PA validator | Manifest | Valid signature, assertions read | Pass | YYYY-MM-DD |
5.2 False positives and reliability
| Measure | Target | Result |
|---|
| False positive rate on human-made reference set | Example: below 0.1% on 10,000 photographs | [Result] |
| Detector availability | Example: 99.5% monthly | [Result] |
| Adversarial removal test | Example: red team attempts with common removal tools | [Findings reference] |
5.3 Re-test triggers
Re-run the suite when the generator model, watermark library, C2PA SDK
or output encoder changes, when a platform changes its upload pipeline,
and at least every [six] months.
6. User-facing disclosure and label design
6.1 Label catalogue
| ID | Context | Wording | Placement | Persistence | Languages |
|---|
| LBL-001 (Example) | Chat assistant | "You are chatting with an AI assistant." | Header and first message | Whole session | All UI languages |
| LBL-002 (Example) | Voice agent | "This is an automated AI voice assistant." | First sentence of call | Repeated on transfer | Spoken languages offered |
| LBL-003 (Example) | Deep fake video | "AI-generated" plus icon | Top left corner | Whole duration | Caption language |
| LBL-004 (Example) | Artistic or satirical work | "Contains AI-generated scenes" | Opening or closing credits | Once | Work language |
| LBL-005 (Example) | Downloaded image | Content Credentials icon with details on hover | Viewer overlay; metadata | File lifetime | n/a |
6.2 Design rules
- Use plain words such as "AI-generated" or "made with AI"; avoid
jargon such as "synthetic media".
- Pair text with an icon; never rely on colour alone.
- Keep the label readable at the smallest supported display size.
- Provide an accessible name for screen readers and a spoken form for
audio-only channels.
- Translate labels with the interface, and review each translation
with a native speaker.
- Do not let users switch off required labels; optional labels may be
user-controlled.
6.3 User testing
| Test | Participants | Finding | Change made |
|---|
| UT-001 (Example) | 12 users, two languages | 3 of 12 missed the header label on mobile | Label added to first message bubble |
7. Exceptions and their justification
| ID | Paragraph | Exception relied on | Output | Justification | Approved by | Review date |
|---|
| EXC-001 (Example) | 50(1) | Obvious from context | Voice command on a smart speaker that announces itself as an assistant | Product is sold and presented as an AI assistant | [@legal] | YYYY-MM-DD |
| EXC-002 (Example) | 50(2) | Assistive function for standard editing | Grammar correction in the editor | Does not substantially alter semantics | [@legal] | YYYY-MM-DD |
| EXC-003 (Example) | 50(4) | Human review and editorial responsibility | Blog market updates | Named editor, documented review | [@legal] | YYYY-MM-DD |
Rejected exception requests: [Example: EXC-R01, spell-check only on an
AI-written press release; rejected because a spell-check is not
editorial control.]
8. Code of Practice and guidelines mapping
8.1 Status
| Instrument | Status as recorded | Our position |
|---|
| Code of Practice on Transparency of AI-generated Content | Final version published after the closing plenary of 10 June 2026; open for signature | [Signed on YYYY-MM-DD / Not signed] |
| Commission Guidelines on the transparency obligations under Article 50 | Published 20 July 2026 (library page updated 31 July 2026) | [Read and applied, version] |
| Commission FAQ on Article 50 | Updated 24 July 2026 | [Read] |
8.2 Commitment mapping
| Code section | Topic | Our measure | Section of this plan |
|---|
| Providers (Example) | Multi-layered machine-readable marking per modality | Technique matrix | 4.1 |
| Providers (Example) | Detection interfaces for marks | Verification API and public page | 4.3 |
| Providers (Example) | Testing robustness and reliability | Test plan | 5 |
| Providers (Example) | Prohibiting removal of marks in terms of use | Acceptable use policy clause | 4.3 |
| Deployers (Example) | Labelling deep fakes and public-interest text | Label catalogue | 6.1 |
| Deployers (Example) | Documenting human review and editorial control | Editorial log | 7 |
Use the commitment numbers from the official text of the Code when
completing this table; this template does not reproduce them.
9. China-facing services
| Requirement (summary) | Source | Our measure | Evidence |
|---|
| Explicit labels on text, audio, image, video and virtual scene outputs, at the positions the rules describe | Measures Art. 4; GB 45438-2025 | Example: visible "AI-generated" text label on images; spoken cue at start of audio | EV-CN-01 |
| Explicit label kept in downloaded or exported files | Measures Art. 4 | Example: label burned into exported image | EV-CN-02 |
| Implicit labels in metadata: content attributes, provider name or code, content ID | Measures Art. 5; GB 45438-2025 | Example: metadata fields per the standard; watermark encouraged | EV-CN-03 |
| User agreement explains labelling methods and rules | Measures Art. 8 | Terms section [x] | EV-CN-04 |
| Users who request output without explicit labels: agreement sets out their labelling duties; recipient logs kept for at least six months | Measures Art. 9 | Example: not offered | n/a |
| No deletion, alteration, forgery or concealment of labels | Measures Art. 10 | Acceptable use clause | EV-CN-05 |
| Platform duties if we distribute third-party content | Measures Art. 6 | Example: not a dissemination platform | n/a |
| Labelling materials submitted with algorithm filing or security assessment | Measures Art. 12 | [Filing reference] | EV-CN-07 |
| Generative AI service duties (security assessment, algorithm filing where required, content management) | Interim Measures for the Management of Generative AI Services (in force 15 August 2023) | [Filing reference] | EV-CN-06 |
10. Other jurisdictions
| Jurisdiction | Instrument | Duty in brief | Applies? | Measure |
|---|
| Republic of Korea | Framework Act on the Development of Artificial Intelligence and the Establishment of a Foundation for Trust (in force 22 January 2026) | Advance notice of AI use; labels on generative AI outputs, with a clear notice for realistic synthetic sound, image or video | [Yes / No] | [Measure] |
| California | AI Transparency Act (SB 942, as amended by AB 853; operative date reported as 2 August 2026) | Covered providers offer a free detection tool and include latent and optional manifest disclosures in image, video and audio | [Yes / No] | [Measure] |
| California | Companion chatbot law (SB 243, in force 1 January 2026) | Disclosure that a companion chatbot is not human and safety protocols | [Yes / No] | [Measure] |
| [Other] | [Instrument] | [Duty] | [Yes / No] | [Measure] |
11. Records, monitoring and audit
11.1 Evidence register
| ID | Evidence | Location | Retention |
|---|
| EV-001 (Example) | Chat disclosure screenshots per release | [Evidence store path] | 10 years |
| EV-030 (Example) | Robustness test reports | [Path] | 10 years |
| EV-031 (Example) | Signing certificate history | [Path] | 10 years |
| EV-032 (Example) | Exception approvals | [Path] | 10 years |
11.2 Monitoring indicators
| Indicator | Target | Frequency | Owner |
|---|
| Share of outputs with a valid manifest at egress | Example: 100% | Daily | [@ml-lead] |
| Watermark detection rate on sampled published content | Example: at least 90% | Monthly | [@ml-lead] |
| Complaints about unlabelled AI content | Example: each reviewed within 5 working days | Continuous | [@trust-safety] |
| Label regression in UI tests | Example: zero | Each release | [@qa-lead] |
11.3 Audit and review
Review this plan at least every [12] months, after any change to the
Code of Practice, the guidelines or the Regulation, and after a serious
labelling failure. Internal audit checks a sample of [50] outputs per
modality each year against sections 3 to 6.
12. Glossary and references
| Term | Meaning |
|---|
| Provider | Develops an AI system, or has it developed, and places it on the market or puts it into service under its own name |
| Deployer | Uses an AI system under its authority in a professional activity |
| Deep fake | AI-generated or manipulated image, audio or video that resembles existing persons, objects, places, entities or events and would falsely appear authentic or truthful |
| Content Credentials | Signed provenance information defined by the C2PA specification |
| Soft binding | A watermark or fingerprint that links content to its manifest when metadata is stripped |
| Explicit label | Under the China measures, a label people can perceive directly |
| Implicit label | Under the China measures, a label embedded in file data |
- Regulation (EU) 2024/1689 (Artificial Intelligence Act) as amended by
Regulation (EU) 2026/1744, Articles 3, 5, 50, 96 and 99
(consolidated text on EUR-Lex).
- European Commission, Code of Practice on Transparency of AI-generated
Content (final, June 2026), digital-strategy.ec.europa.eu.
- European Commission, Guidelines on transparency obligations for
providers and deployers of AI systems under Article 50 (July 2026).
- European Commission, FAQ on transparency obligations under Article 50.
- C2PA Technical Specification 2.4 (April 2026), spec.c2pa.org.
- Measures for Labeling AI-Generated and Synthetic Content (2025) and
GB 45438-2025, effective 1 September 2025 (cac.gov.cn).
- Interim Measures for the Management of Generative AI Services (2023).
- Related templates:
ai-data/ai-use-case-intake,
ai-data/ai-system-card, ai-data/ai-launch-review,
compliance/ai-regulatory-applicability-register,
ai-data/guardrails-spec.
Standard and licence note
Based on Regulation (EU) 2024/1689 (the Artificial Intelligence Act) as
amended by Regulation (EU) 2026/1744, Article 50, as published in the
Official Journal of the European Union, read with the Commission's Code
of Practice on Transparency of AI-generated Content (June 2026), its
Article 50 guidelines (July 2026), the C2PA Technical Specification 2.4
and the Chinese Measures for Labeling AI-Generated and Synthetic Content
with GB 45438-2025. EU legislative texts are free to reuse; obligations
are summarised in original wording and no text of the Regulation, the
Code, the guidelines, the C2PA specification or the Chinese measures is
reproduced. Check the consolidated text on EUR-Lex, the current
Commission pages and c2pa.org for the binding wording, later
amendments and newer versions. This is a documentation format, not
legal advice. Template licence: MIT.