Generated images and video carry risks that a text-focused AI acceptable use policy rarely spells out: a real person's face or voice, an unlabelled synthetic photo in a press release, or a campaign visual the organisation cannot own. This guide is a policy annex for generated media at work. It assumes you already have an AI acceptable use policy and adds the rules that only apply to images and video: likeness consent, labelling and provenance, copyright and ownership, brand approval, and the controls that make those rules enforceable. It is written for security, legal, marketing and communications leads who have to approve, run or audit the use of image and video generators.
Key takeaways
- Treat generated media as its own risk class. Text rules on data input and output review do not cover likeness, synthetic realism or provenance, so add a short annex rather than stretching the general policy.
- Consent comes before any real likeness. Generating or editing an identifiable person, including executives and staff, needs documented consent and a stated purpose, and some uses should be prohibited outright.
- Label realistic synthetic media and keep provenance. EU AI Act Article 50 requires deployers to disclose deep fakes from 2 August 2026, and C2PA Content Credentials plus IPTC source codes give you a machine-readable way to do it.
- Do not assume you own the output. The US Copyright Office concluded in January 2025 that prompts alone do not make a user the author, so brand assets that must be protected need real human authorship.
- Enforce through tools, accounts and review, not just words. An approved tool list, enterprise accounts, a review gate for public content and an asset register make the policy auditable.
Why images and video need their own rules
The general AI Acceptable Use Policy Template already covers approved tools, data classification for inputs and human review of outputs, and it requires labelling of AI-generated images, audio and video that show real people or events. That single line is the right principle but not enough to run a marketing team, an internal comms function or a training video studio. Generated media differs from generated text in four ways that drive the rules below.
- Realism: a photorealistic image or a cloned voice can be mistaken for a record of something that happened, which turns an internal shortcut into a reputational or legal event once it leaves the building.
- Likeness: faces and voices identify people. Using them touches personality and image rights, data protection and, for intimate or defamatory content, criminal law in many jurisdictions. The Deepfake and Synthetic Media Laws Map tracks these laws by country.
- Ownership: copyright protection for generated media depends on human authorship, so an organisation can publish a visual it cannot stop others copying.
- Persistence: media is shared, cropped and re-encoded, which strips metadata. NIST AI 100-4 notes that metadata is often removed when files are shared and that most watermarks tested have proved vulnerable to removal and spoofing.
Use case tiers for generated media
Tier uses by audience and realism, then attach controls to each tier. This keeps low-risk internal work fast and concentrates review where the exposure is. The approach mirrors the wider AI Use Case Risk Tiering Guide, applied to media.
| Tier | Examples | Default rule | Required controls |
|---|---|---|---|
| 1. Internal, non-realistic | Mood boards, slide icons, storyboards, abstract illustrations | Allowed with approved tools | Approved tool and enterprise account; no confidential inputs above the tool's approved classification |
| 2. Internal, realistic or edited | Photoreal mock-ups, edited staff photos for an intranet, training scenes | Allowed with conditions | Tier 1 controls plus label on the asset; consent for any identifiable person |
| 3. External, non-realistic | Social graphics, blog illustrations, product concept art | Allowed after review | Brand review, copyright and trademark check, asset register entry, provenance kept |
| 4. External, realistic, or any real person | Photoreal adverts, synthetic spokesperson video, executive avatar, voice clone, edited news or event imagery | Exception only | Legal and communications approval, written consent, disclosure label, Content Credentials, retained prompts and source files |
| 5. Prohibited | Intimate or sexualised imagery of real people, depictions of real people saying or doing things they did not, fake evidence, documents or IDs, impersonation of customers, regulators or competitors | Never permitted | Block where tools allow, report as a policy breach, route to HR and legal |
Decision rule: if a reasonable viewer could take the output as a real photograph, recording or statement by a real person, it is at least Tier 4. When in doubt, move up a tier.
Likeness and consent
Most organisational harm from generated media involves a real person. Your annex should define likeness broadly: face, body, voice, name and distinctive mannerisms of any identifiable individual, whether an employee, executive, customer, celebrity or member of the public.
- Written, specific consent: consent should name the purpose, the channels, the period and whether the likeness may be used to train or fine-tune a model or build an avatar. Blanket clauses in employment contracts are a weak basis, because staff may not feel free to refuse. Make withdrawal possible and record what happens to existing assets when consent is withdrawn.
- Data protection: under UK and EU GDPR an image or voice of an identifiable person is personal data, and processing it to create a model of that person raises questions of lawful basis and possibly special category biometric data. Run a DPIA before building any avatar or voice model. The ICO maintains guidance on AI and data protection, currently under review following the Data (Use and Access) Act.
- Executives and public figures: leaders are the likeness most valuable to fraudsters. In the Arup deepfake fraud, Hong Kong police said the fake participants on a video call were built from publicly available footage. An official executive avatar programme normalises synthetic video of leaders, so pair it with a verification procedure staff can use. See the Executive Deepfake and CEO Fraud Guide.
- Third parties and celebrities: prohibit generating recognisable third parties for commercial use without a licence. Several jurisdictions protect likeness and voice through personality, publicity or passing off rules; check the laws map for each market you publish in.
- Customers: customer photos, ID images and call recordings must never be used as generation inputs unless a specific lawful purpose and approval exist.
Labelling, disclosure and provenance
Labelling has three layers: what the law requires, what a human viewer sees and what a machine can read. Your policy should set all three.
What the EU AI Act requires
Article 50 of the EU AI Act applies from 2 August 2026. Article 50(2) places the marking duty on providers of generative systems: outputs must be marked in a machine-readable format and detectable as artificially generated or manipulated, with exceptions for assistive standard editing that does not substantially alter the input. Article 50(4) places the disclosure duty on deployers, which includes an organisation using a generator in its own content: deployers of a system that generates or manipulates deep fakes must disclose that the content has been artificially generated or manipulated. For evidently artistic, creative, satirical or fictional works, disclosure is limited to acknowledging that such content exists in a way that does not spoil the work. Article 50(6) confirms these duties sit alongside other EU and national transparency rules. Detail is in the EU AI Act Article 50 Transparency Obligations Guide and the EU Code of Practice on Marking and Labelling AI-Generated Content Guide. China has its own explicit and implicit labelling rules, covered in the China Generative AI Measures and AI Labelling Rules Guide.
Machine-readable provenance
| Mechanism | What it is | Strength | Limit |
|---|---|---|---|
| C2PA Content Credentials | A signed manifest of assertions about the asset, bound to it by a claim and claim signature (C2PA specification 2.2, May 2025) | Tamper-evident; records actions such as create and edit; identifies the signer | Can be separated from the file when metadata is removed; a signature shows who made the claim, not that it is true |
| IPTC Digital Source Type | Controlled vocabulary in image metadata: trainedAlgorithmicMedia (created using generative AI), compositeWithTrainedAlgorithmicMedia (edited using generative AI), algorithmicallyEnhanced (algorithmic correction such as sharpening) | Widely supported by asset managers; simple to set | Unsigned metadata; easily stripped or edited |
| Invisible watermark (soft binding) | Identifier embedded in the pixels or audio; C2PA uses it to match an asset back to a stored manifest | Can survive some re-encoding and screenshots | NIST AI 100-4 reports most watermarks tested are vulnerable to removal and spoofing |
| Visible label or caption | On-image text, caption or voice-over stating the content is AI generated | Seen by every viewer; meets the human disclosure need | Can be cropped; must be designed into templates |
Policy position: for Tier 2 and above, require a visible label where realism could mislead, set the IPTC Digital Source Type in your asset manager, and keep Content Credentials where the tool or editing pipeline produces them. Do not rely on any one mechanism: NIST AI 100-4 states that none of the techniques offers a comprehensive solution on its own. The AI Content Provenance and C2PA Guide covers signing and verification in depth.
Copyright, ownership and brand
The US Copyright Office published Part 2 of its Copyright and Artificial Intelligence report, on copyrightability, on 29 January 2025. Its conclusions shape what a brand team can rely on.
- Prompts alone: with current generally available technology, prompts alone do not give the user enough control to be the author of the output, however many times the prompt is revised.
- Human inputs: where a person supplies their own copyrightable work as input and it remains perceptible in the output, they are the author of at least that portion.
- Selection, arrangement and modification: human creative selection, arrangement and modification of generated material can be protected, but the protection covers the human contribution, not the generated material on its own.
- Registration: applicants should disclose more than de minimis AI-generated material and describe the human contribution.
Other jurisdictions take different positions, so legal should set the rule per market. The practical policy consequences are consistent:
- Logos, mascots, product designs and other assets the organisation must own and enforce should be created or substantially reworked by humans, with records of that work.
- Generated visuals may be used for supporting content where exclusivity does not matter.
- Prompts must not ask for a named artist's style, a competitor's trade dress, a trademarked character or a copy of a specific photograph.
- Check vendor terms for output ownership, indemnity, permitted commercial use and whether inputs train the vendor's models, and record the result in the approved tool list. The AI Vendor Security Questionnaire Template includes data use questions.
- Run a reverse image or similarity check on Tier 3 and 4 assets before publication to catch near copies of existing works.
Policy annex clauses
These clauses extend an existing AI acceptable use policy. Replace bracketed text.
- Scope: this annex applies to any image, video, animation, avatar or synthetic voice created or materially edited with an AI tool by or for [Organisation], including by agencies and contractors.
- Approved tools: only tools on the [Approved AI Tools Register] may be used, through [Organisation]-managed accounts. Personal accounts and free tiers are not permitted for work content.
- Inputs: do not upload confidential or restricted information, customer images, identity documents or unreleased product designs unless the tool is approved for that classification.
- Real people: do not generate or edit the likeness or voice of any identifiable person without written consent recorded in [Consent Register] and approval from [Legal]. Prohibited uses in Tier 5 are never permitted.
- Labelling: realistic generated or manipulated media published externally must carry a visible disclosure and machine-readable provenance as set out in [Labelling Standard].
- Review: Tier 3 and 4 assets require [Brand] review and, for Tier 4, [Legal] and [Communications] approval before publication.
- Records: keep the prompt, tool and version, source inputs, consent reference and approver for every Tier 3 and 4 asset for [retention period].
- Agencies: contracts with agencies and freelancers must require disclosure of AI use, compliance with this annex and transfer of all records.
- Reporting: report suspected deepfakes of [Organisation] staff or brand to [Security Team] at [contact].
Implementing the annex
- Discover current use: review proxy, CASB or SSE logs and expense claims for image, video and voice generation services, and survey marketing and comms teams and agencies. The Shadow AI Discovery Guide covers discovery methods.
- Pick approved tools: assess terms, data use, account controls, provenance output and content filters; move users onto enterprise accounts with SSO.
- Publish the annex and labelling standard: include label templates for slides, video lower thirds and social posts so labelling is the easy path.
- Stand up registers: a consent register for likeness and an asset register in the digital asset manager with fields for source type, tool, approver and consent reference.
- Pilot with one team: run the review gate with marketing for a month, measure turnaround and adjust tiers before extending to all teams and agencies.
- Train: short, role-based sessions for creators and approvers, plus deepfake awareness for finance and executive assistants. See the Spotting Deepfake Scams Guide.
- Roll back safely: keep a takedown procedure for any asset found to breach the annex: remove, notify affected people, correct the record publicly where needed, and log the incident.
Small organisations: a single approved tool, a one-page annex, a shared consent folder and one named approver are enough. Large organisations: integrate the review gate into the digital asset manager workflow, set provenance at export, and audit agency output quarterly.
Detection and measurement
- Monitor: network and SaaS logs for unapproved generation services; DAM uploads missing a source type field; public brand monitoring for synthetic content using your logos or executives.
- Metric 1: share of generation traffic going to approved tools through managed accounts.
- Metric 2: percentage of published Tier 3 and 4 assets with a complete register entry (tool, prompt, approver, consent where needed).
- Metric 3: percentage of external realistic assets carrying both a visible label and machine-readable provenance at audit.
- Metric 4: median review turnaround for Tier 3 assets, which shows whether the gate is pushing people to workarounds.
- Metric 5: number of takedowns or policy breaches per quarter, and time from report to removal for impersonation of staff or brand.
Framework mapping
| Annex element | NIST SP 800-53 Rev 5 | NIST CSF 2.0 | ISO/IEC 27001:2022 Annex A |
|---|---|---|---|
| Rules of use and prohibited content | PL-4 Rules of Behavior | GV.OC-03 legal and regulatory requirements | A.5.10 Acceptable use of information |
| Training for creators and approvers | AT-2 Literacy Training and Awareness | PR.AT-01 awareness and training | A.6.3 Awareness, education and training |
| Provenance and integrity of published media | SI-7 Software, Firmware, and Information Integrity | PR.DS-01 data at rest protection | A.5.12 Classification and A.5.13 Labelling of information |
| Records and retention | SI-12 Information Management and Retention | ID.AM-07 data inventories | A.5.33 Protection of records |
| Copyright and likeness | None specific | GV.OC-03 | A.5.32 Intellectual property rights; A.5.34 Privacy and protection of PII |
For threat modelling, MITRE ATLAS lists AML.T0088 Generate Deepfakes. See AML.T0088 Generate Deepfakes.
Common failure patterns
- One-line policy: a general AUP that says "label AI images" with no tiers, templates or owner, so nobody labels anything.
- Consent by contract clause: relying on a broad employment clause to build staff avatars, then facing withdrawal requests with no process to retire the assets.
- Metadata-only labelling: trusting embedded provenance that social platforms and messaging apps strip, leaving no visible disclosure.
- Brand assets nobody owns: a generated logo or mascot adopted as a core brand element that the organisation cannot register or enforce.
- Agency blind spot: strict internal rules while agencies use any tool they like, with no contractual disclosure or records.
- Executive likeness normalised: synthetic CEO videos used internally with no verification procedure, so a fraudulent one looks routine.
Practitioner checklist
- Add a generated media annex to the existing AI acceptable use policy and name an owner.
- Define the five tiers and the decision rule for realism.
- Publish the approved tool list with output ownership, data use and indemnity terms recorded.
- Create a consent register covering purpose, channels, period, model training and withdrawal.
- Issue visible label templates and set IPTC Digital Source Type in the asset manager.
- Keep Content Credentials through the editing and export pipeline where tools produce them.
- Require human authorship for logos and other assets the organisation must own.
- Put the annex into agency and freelancer contracts.
- Run a takedown and verification procedure for deepfakes of staff and brand.
- Report the five metrics to the AI governance forum quarterly.
Standards and references
- EU AI Act: Article 50 Transparency Obligations for Providers and Deployers
- US Copyright Office: Copyright and Artificial Intelligence, Part 2 Copyrightability
- NIST: AI 100-4 Reducing Risks Posed by Synthetic Content
- C2PA: Content Credentials Specification 2.2
- IPTC: Digital Source Type NewsCodes
- ICO: Guidance on AI and Data Protection
- US Copyright Office: Copyright and Artificial Intelligence
Related NHI Mgmt Group resources: AI Acceptable Use Policy Template, Deepfake and Synthetic Media Laws Map, Shadow AI Policy Template, Deepfake and AI Impersonation Guide.