Tier 1 · Model 4 of 4 · v0.2

AI Content Integrity Model

A risk classification framework for the ethical dimensions of AI-generated and AI-manipulated visual content. Applicable to any organisation publishing AI-assisted imagery, video, animation or synthetic personas.

Three axes, read in order. The first group of rules is decided by one axis alone, because a consent failure is not something a disclosure label can argue down.

Axes: 3
Classifications: 4
Code: AI · CS · DC
Scope: visual and mixed media

Published March 2026. The classification sits alongside the Transparency band on the Report Card, and the two answer different questions.

CI v0.2 · Published · March 2026

The classifier · v0.2

Select one from each axis

Axis 1

Intervention

Axis 2

Consent

Axis 3

Disclosure

Disclosure code

AI-ANIMCS-AIDC-CAP

Integrity classification

Clear

Ethical use confirmed. Consent and disclosure meet the standard for responsible practice.

Risk factors

None. The combination selected meets the standard on all three axes.

Recommended action: Proceed with publication.

The three overriding rules

These fire before anything else and cannot be argued down by the other two axes.

CS-NONENot recommended. Real, identifiable individuals depicted with no release of any kind.

CS-ESTNot recommended. Deceased or historical individuals without confirmed estate consent.

DC-NONENot recommended. No indication was given to the audience that AI was involved.

01.

About this model

This model applies to AI-generated imagery, AI-animated photography, synthetic video and AI personas. Text-only AI content is not in scope; consent and disclosure obligations for text are addressed by existing editorial standards and the AI Transparency Model.

The AI Transparency Model records how much AI was involved in producing a piece of work. That is a necessary disclosure, but it does not address a separate and equally important question: whether the AI involvement was ethically sound in relation to the people depicted and the audiences reached. Standard image licensing and photography rights frameworks address ownership and permission to use, but they were not designed for AI manipulation of real people's likenesses, and they do not require disclosure of AI involvement to the audience. This model fills that gap.

The model assesses three things: what AI did to the content, what authorisation exists for any real people depicted, and what the audience was told. From those three inputs it produces an Integrity classification (Clear, Caution, High Risk or Not Recommended) and a machine-readable disclosure code that can travel with the asset through production and delivery pipelines. The classification sits alongside the Transparency grade on the Report Card; the two scores answer different questions and are assessed independently.

Disclosure is necessary but not sufficient. An AI declaration does not resolve an underlying consent failure. A piece of content can carry a transparent AI label and still be ethically unacceptable if the people depicted did not consent to AI manipulation of their image.

02.

Why the categorical claim lives here

A fully AI-generated image cannot report as Extensive on the Transparency scale. Where a person wrote the concept, retouched the output and selected it, the score plateaus at Considerable, and the first weighting whose arithmetic would reach Extensive strands Considerable, which the framework refuses.

The claim being reached for is categorical rather than proportional: the picture is synthetic. That is AI-IMG on this model, not a band on that one. It follows, and it is binding, that a visual disclosure surfaces its Content Integrity code at equal prominence to its Transparency band.

Disclosure cannot fix a consent failure. A Not Recommended classification cannot be resolved by adding a label to the asset. Consent must be obtained, or the content must not be published.

Read the Transparency Model ›

01.

The three assessment axes

Every piece of AI visual content is assessed against three axes. The results combine to produce an Integrity classification and a machine-readable disclosure code. Each axis is scored independently using the codes below.

Axis 1

Intervention

What AI did to or created in the content. This determines baseline ethical exposure and is the first component of the disclosure code.

AI-TXT

Text only

AI generated or assisted with written content. No imagery, no real people manipulated.

Baseline risk: low

AI-IMG

Synthetic imagery

AI-generated visuals depicting landscapes, places or objects with no identifiable real individuals depicted.

Baseline risk: low

AI-ENH

Image enhancement

AI-applied animation or enhancement of real photography where any people present are unidentifiable — distant, backs turned or faces obscured.

Baseline risk: caution

AI-ANIM

Animation of real people

Real photographs or footage manipulated to show identifiable individuals moving, performing actions or appearing in altered settings. The VisitBritain/TripAdvisor scenario of animating a bagpipe player or a walking family falls under this code.

Baseline risk: high_risk

AI-CHAR

Character-based synthetic persona

A clearly fictional AI persona — creature, mascot or stylised character not intended to pass as human. Audiences can reasonably recognise the character as synthetic.

Baseline risk: caution

AI-PERS

Human-passing synthetic persona

A synthetic human-like persona designed to appear realistic. Includes hyper-realistic AI influencers and spokespeople used in campaigns. Audiences may not immediately recognise the persona as AI-generated without explicit disclosure.

Baseline risk: not_recommended

Axis 2

Consent

What authorisation exists for real subjects depicted or represented. This is the central ethical gate. A standard photography release does not cover AI manipulation of a subject's image.

CS-NA

Not applicable

No real people depicted. Content is fully synthetic or involves only landscapes, objects or unidentifiable figures.

Baseline risk: low

CS-ANON

Anonymous subjects

People visible in the content but cannot reasonably be identified — backs turned, distant, faces obscured or in crowd shots where individuals are indistinguishable.

Baseline risk: low

CS-AI

Explicit AI-use consent

Subjects provided specific written consent for their image or likeness to be used in AI-manipulated or AI-generated content. This must be explicit and separate from a standard photography release — a standard release does not qualify.

Baseline risk: low

CS-STD

Standard release only

A standard photography or model release is in place, but no specific consent was obtained for AI manipulation, animation or use in AI-generated contexts. Standard releases typically cover photography and traditional editing only.

Baseline risk: high_risk

CS-NONE

No consent

Real, identifiable individuals depicted with no model release or consent of any kind. Common in user-generated content or public photography used without rights.

Baseline risk: not_recommended

CS-EST

Estate or deceased

Deceased or historical individuals. Estate consent not confirmed. Even historical public figures carry significant reputational and legal risk when their likeness is used in AI-animated or AI-generated contexts.

Baseline risk: not_recommended

Axis 3

Disclosure

What the audience was told about the AI nature of the content. Disclosure is necessary but not sufficient — it cannot resolve a consent failure.

DC-PROM

Prominent in-content disclosure

A visible label, watermark, badge or on-screen notice embedded in or overlaid on the content itself. Visible to any viewer regardless of whether they read captions, descriptions or metadata.

Baseline risk: best_practice

DC-CAP

Caption or post-level disclosure

AI use acknowledged in accompanying text — social caption, post copy, article description or press release note. Visible on-platform to anyone who reads the accompanying copy.

Baseline risk: acceptable

DC-META

Metadata or regulatory minimum

AI disclosure recorded in file metadata (EXIF, XMP, IPTC) or provided to the platform for compliance purposes. Not visible to the general audience consuming the content.

Baseline risk: insufficient

DC-NONE

No disclosure

No indication given to the audience that AI was involved in creating or manipulating this content. Not recommended for any visual or persona content regardless of intervention type.

Baseline risk: not_recommended

01.

The disclosure code system

Every assessed piece of content produces a three-part machine-readable disclosure code combining one value from each axis. The code is designed to be embedded in asset metadata (EXIF, XMP, IPTC), included in agency delivery notes, referenced in procurement clauses and appended to social media captions. It is readable by humans, embeddable by systems and usable as a supply chain compliance check.

Disclosure code format

[AI-type] · [CS-consent] · [DC-disclosure]

Example

AI-ANIM · CS-AI · DC-CAP

animated real people, explicit AI consent obtained, caption-level disclosure

The code is designed to be attached to any asset: in file metadata (EXIF, XMP, IPTC), in contract deliverable notes, in social captions or in report footers. It is readable by humans, embeddable by systems and usable as a procurement compliance check.

Procurement use. A suggested contract clause for commissioning AI-generated or AI-manipulated content: "All AI-generated or AI-manipulated visual content delivered under this contract must carry a DTTT disclosure code in the format [AI-type] · [CS-consent] · [DC-disclosure]. Assets classified as High Risk or Not Recommended require written approval before delivery." Take legal advice before using this clause in live contracts.

02.

Integrity classification

The three axes combine to produce one of four classifications. This is a risk classifier, not a quality scale. It does not measure how much AI was used; it assesses whether the ethical conditions for responsible publication have been met.

Clear

Ethical use confirmed. Consent and disclosure meet the standard for responsible practice.

Proceed with publication.

Caution

One or more aspects of consent or disclosure fall below the recommended standard.

Editorial review advised before publishing. Document the review decision.

High risk

Significant ethical and potentially legal concerns identified.

Do not publish without formal legal and editorial review.

Not recommended

Fundamental ethical issue identified that cannot be resolved through disclosure alone.

Do not publish.

03.

How the classification is decided

The rules are read in three groups, and the first group that fires decides the outcome. Nothing in a later group can soften a classification an earlier one set.

Group 1 · decided by one axis

These three fire first and cannot be argued down by the other two axes. A consent failure is not a disclosure problem.

CS-NONE, any, anyNot recommended. Real, identifiable individuals depicted with no release of any kind.

CS-EST, any, anyNot recommended. Deceased or historical individuals without confirmed estate consent.

any except AI-TXT, any, DC-NONENot recommended. No indication was given to the audience that AI was involved.

Group 2 · high risk

Read only where nothing in group 1 fired. Each combines a consent gap or a passing persona with disclosure that does not carry the weight.

AI-ANIM, CS-STD, anyHigh risk. Identifiable individuals animated without specific AI-use consent.

AI-PERS, any, DC-METAHigh risk. A human-passing persona disclosed only in metadata.

any, CS-STD, DC-META or DC-NONEHigh risk. A consent gap combined with insufficient disclosure.

Group 3 · caution

Read last. Each is publishable with an editorial decision recorded, and none is resolved by the label alone.

AI-ENH, any, DC-NONECaution. Animation of real imagery with no disclosure, even where people are unidentifiable.

AI-CHAR, any, DC-NONECaution. A fictional persona still requires disclosure.

AI-PERS, any, anyCaution. A human-passing persona carries inherent representational risk.

any, CS-STD, DC-CAPCaution. Caption disclosure mitigates but does not close the consent gap.

any, any, DC-METACaution. Metadata alone is not audience-facing transparency.

04.

JSON schema

The canonical machine-readable definition of this model, including all intervention, consent and disclosure codes, classification logic and the disclosure code system, is published as a JSON schema file under CC BY 4.0.

dttt-integrity-v0.2.json

AI Content Integrity Model, current published version

Download JSON ›