Article or blog post
Typical: ModerateMaximum: Considerable
The argument, the research and the editorial judgement are ours. The drafting frequently is not, and we say so on each piece.
GS-DTTT@1.1
DTTT AI Governance Statement
Our stated positions on AI across the things the AI Transparency Framework measures, and how those positions are held.
These are limits rather than descriptions of practice. A ceiling states the most AI involvement we deem acceptable for a kind of work. It does not say how much we use, and most of our work sits below its ceiling. Our own website scored Considerable against a ceiling of Extensive.
We publish this because we publish the framework. An organisation that asks others to state their position should state its own.
01.
Stated, never calculatedOur positions
Ceilings9 media · TM v2.0 bands
Each carries what we typically do and the most we permit. Only the maximum is a position: the typical is a description, and it is stated here rather than counted. It becomes a counted figure once our own disclosures are held in one place.
TypicalPermittedNot permittedMaximum threshold
Report or strategy document
Typical: ModerateMaximum: Considerable
The research, the analysis and the conclusions are ours. Drafting, editing and production are frequently AI-assisted.
Still image
Typical: ModerateMaximum: Moderate
An image of a place should be a record of that place. We keep our photography substantially our own.
ExceptionIllustrated treatments applied to our own photography are permitted. A standard design is applied across case study imagery so that a set reads coherently. The result is visibly an illustration rather than a photograph, and it is labelled as one.Social post
Typical: LimitedMaximum: Considerable
The angle and what we choose to say are ours. Copy and visuals are frequently AI-produced.
ExceptionThe illustrated treatment above applies here too.Video
Typical: MinimalMaximum: Moderate
Video of a place should be a record of that place. Editing, captions and audio are frequently AI-assisted, and the footage is not.
ExceptionThe illustrated treatment above applies here too.Website, microsite or digital product
Typical: ConsiderableMaximum: Extensive
We build our own digital products with AI. The concept, the structure and the content are ours, and the code frequently is not. This site scored Considerable, and its disclosure is published on its home page.
Email or newsletter
Typical: ModerateMaximum: Considerable
What goes in a newsletter is an editorial decision and it is ours. The writing around it is frequently AI-assisted.
Presentation or deck
Typical: ConsiderableMaximum: Extensive
The argument is ours.
Slides, charts and layout are frequently produced with AI.
Bid or submission
Typical: ConsiderableMaximum: Considerable
The approach and the methodology are ours, and a buyer is entitled to know that. Drafting, compliance checking and production are frequently AI-assisted.
What AI may depictCI v0.2
One rule across still images, social posts and video, because the same picture appears in all three and a rule that changed by medium would be arbitrary.
Permitted
- Fully synthetic imagery where no real people appear.
- Clearly fictional characters and illustration.
- Illustrated treatments of our own photography, per the exception above.
Not permitted
- Enhancement of real photography that leaves it looking like a photograph.
- Animation of real imagery showing identifiable people.
- Human-passing synthetic personas.
The distinction is what a viewer is entitled to believe. An illustration announces itself. A wholly synthetic image is not claiming to be a photograph. An enhanced photograph is, and that is where the deception sits.
We do not publish AI photography of real places. Illustration and clearly synthetic imagery are fine, and we label them as illustration.
Consent
Explicit AI-use consent is required before any AI manipulation of an identifiable person.
A standard photography release does not cover it.
Disclosure
Caption or post-level disclosure is the minimum on anything AI-generated.
A viewer should be told in the place they are looking.
Review
Anything published under our name is reviewed by a named person before it goes out.
Review is a governance requirement and never a score adjustment. What produced a piece of work and who checked it are different facts, and the model measures one of them.
Ethics review
The AI Ethics Consideration Tool is run before work begins where the work involves identifiable people, client data, or anything published under a member's name rather than ours.
Suppliers
Vendors and agencies delivering work to us disclose under this framework.
We publish a disclosure framework. Asking our own suppliers to use it is the least we can do, and not asking would make the framework a thing we recommend rather than a thing we use.
EnvironmentalEM v0.6
We record environmental grades where a disclosure carries one. We do not currently set a limit, because the data to set a meaningful one does not exist.
We will revisit this when first-party disclosures under the EU AI Act become available from August 2026.
02.
No model applies hereHow these positions are held
What is fixed and what is operational
Fixed. The ceilings, the depiction rule, the consent and disclosure standards, and the supplier expectation. These change only through a version release, with a dated entry saying what changed and why.
Operational. Which instruments we run and when, our review cadence, and how disclosures are presented. These change without a release.
The split exists so that everything is neither untouchable nor negotiable. Both make a position ungovernable.
Version log
Every change to this statement is dated, with what changed and why.
The log records corrections of our own errors as well as changes of position. A log containing only improvements is a marketing document.
Exceptions
Work falling outside a stated position is approved by Nick Hall or Romy Cywie, and the exception is recorded with its reason.
A position with no exception route is one people breach silently, which is the unwritten rule returning in a second form.
Review
This statement is reviewed every six months, matching the validity window of a recurring disclosure so the two rhythms align.
What we have not decided
1.How far the illustrated treatment extends.
The exception covers a standard design applied across a set. It does not settle where a treatment stops being a treatment and becomes a new image, and that judgement is currently a person's rather than a rule's.
2.Whether synthetic voice is permitted in video.
Our video ceiling leaves room for an AI voiceover, but nobody decided that. It is a consequence of a band rather than a position. Anyone proposing one asks first.
3.Whether these ceilings cover work produced for members under advisory engagements.
When we produce a strategy document for a member, whose position applies, ours or theirs. Not decided, and it will come up.
4.What happens when a supplier declines to disclose.
We expect suppliers to disclose. We have not decided what follows when one will not.
What this statement does not cover
Deliberately, rather than by omission.
These are real governance questions and they are not what this object is for. This statement covers the things the framework measures. The rest is an organisation's own to write, and it is where we help members as advisors rather than as publishers.
- Approved and prohibited tool lists, which date within weeks.
- Data classification and what may be entered into which system.
- Roles, permissions and access.
- Training expectations.
- Contract clauses.
Full version log
1.1 · 26 August 2026
Each ceiling gains a typical value alongside its maximum, because a maximum alone reads as a description of practice and the reasons alone were not enough to prevent that. An exception is added to the three visual media covering illustrated treatments of our own photography.
The exception corrects an error. Version 1.0 stated that we do not manipulate real photography. We apply a standard illustration design across case study imagery, which is manipulation under that rule, so the statement was false about our own practice from the day it was published. A colleague found it by reading the statement against how we actually work, which is what publishing one is for.
1.0 · 25 August 2026
First published.
Governance statements
Two layers, in one document. A statement that published one and not the other would be half an object, and it is the second half that is usually missing.
