AI Transparency Model
Seven bands describing how much of a finished piece of work came from AI. Five are graded and reached by calculation. Two are declared, and no calculation can reach them.
The boundaries are unchanged from v1.1. Only the labels, the input and the published score changed, which is what preserves every disclosure already published.
The published model, in force since March 2026. Five graded bands from a percentage estimate of AI involvement, with two declarations at the extremes.
Still valid, still rendering, and still the correct reading of any disclosure issued under it. v2.0 does not reach the historical record.
Version
- Bands
- 7
- Input
- per activitypercentage
- Score
- points of possiblepercentage
- Code
- TM-[band]@2.0TM-[band]
Boundaries identical to v1.1, so disclosures under both are comparable
Superseded by v2.0 for new disclosures. Historical disclosures remain valid
The scale · v2.0
Select a band
No AI Involvement
Declared
Minimal
0 < r ≤ 0.10
Limited
0.10 < r ≤ 0.36
Moderate
0.36 < r ≤ 0.63
Considerable
0.63 < r ≤ 0.89
Extensive
0.89 < r < 1
AI Generated
Declared
Declared · outside the graded scale
No AI Involvement
Declared, never calculated. The involvement gate was answered no, so no activity is scored and the work carries no band.
Graded band
Minimal
Graded. Points earned divided by points possible falls in 0 < r ≤ 0.10.
Graded band
Limited
Graded. Points earned divided by points possible falls in 0.10 < r ≤ 0.36.
Graded band
Moderate
Graded. Points earned divided by points possible falls in 0.36 < r ≤ 0.63.
Graded band
Considerable
Graded. Points earned divided by points possible falls in 0.63 < r ≤ 0.89.
Graded band
Extensive
Graded. Points earned divided by points possible falls in 0.89 < r < 1.
Declared · outside the graded scale
AI Generated
Declared, never calculated. Every activity in the medium was AI led, the ratio equals one, and the graded scale stops short of it.
The scale · v1.1
Select a band
Human Created
Declared
Human Led
1 to 10%
AI-Assisted
11 to 36%
Collaborative
37 to 63%
AI-Generated
64 to 89%
AI Led
90 to 99%
Fully AI
Declared
Declared · outside the graded scale
Human Created
Declared. No AI involvement at any stage of the work, recorded outside the graded scale.
Graded band
Human Led
Graded. An estimate of 1 to 10% AI involvement in the finished work.
Graded band
AI-Assisted
Graded. An estimate of 11 to 36% AI involvement in the finished work.
Graded band
Collaborative
Graded. An estimate of 37 to 63% AI involvement in the finished work.
Graded band
AI-Generated
Graded. An estimate of 64 to 89% AI involvement in the finished work.
Graded band
AI Led
Graded. An estimate of 90 to 99% AI involvement in the finished work.
Declared · outside the graded scale
Fully AI
Declared. No human authorship beyond the brief, recorded outside the graded scale.
01.
How a band is reached
Involvement is scored per activity, weighted by medium, and expressed as points earned of points possible. A percentage is never published, because involvement across a production process is a judgement and a percentage misrepresents it as a measurement.
- 1
Name the medium
Weights are per medium, so nothing can be scored until the medium is chosen. Eight are ratified, each with its own ordered activity list.
- 2
Answer per activity
Three values only: 0 the human did this, 1 AI helped, 2 AI did this. Coarse on purpose, because involvement is a judgement rather than a measurement.
- 3
Earn points against the possible
Earned is the sum of each answer multiplied by its activity weight. Possible is the sum of two multiplied by each weight. Both figures are published.
- 4
The ratio selects the band
Earned divided by possible falls in a band. At either endpoint it falls in none, and the flow asks for a declaration instead of choosing one.
02.
The medium registry
Weights are defined per medium, not globally. Nine media are ratified. A definition is frozen once ratified, and adding one requires an exhaustive reachability check proving every band is reachable using the three answer values. Weights that strand a band are invalid.
Article or blog post
5 activities / 28 points
Report or strategy document
6 activities / 38 points
Still image
4 activities / 24 points
Social post
4 activities / 26 points
Video
6 activities / 40 points
Website, microsite or digital product
5 activities / 38 points
Email or newsletter
5 activities / 26 points
Presentation or deck
5 activities / 32 points
Bid or submission
6 activities / 40 points
All nine definitions are ratified and frozen. Points possible runs from 24 to 40, so figures are read within a medium and never compared across media.
03.
Two known limits of the proportion reading
Both are limits rather than defects. Neither is fixed by adjusting weights, and stating them here is what stops someone later trying.
Limit, not a defect
Proportion does not measure originality
A report built on an organisation’s own data and one built on AI-summarised desk research can score identically. The model measures how much of the artefact came from AI and cannot distinguish the provenance of what it says. Weighting cannot address it, and weights must not be adjusted in the attempt.
Limit, not a defect
A generated image cannot report as Extensive
Where a person wrote the concept, retouched the output and selected it, the score plateaus at Considerable. It moves once, from Moderate at six of ten to Considerable at seven of eleven, and holds. The first weighting whose arithmetic would reach Extensive strands Considerable, which makes it invalid, so the registry refuses it rather than scoring with it.
It follows, and is binding, that a visual disclosure surfaces its Content Integrity code at equal prominence to its Transparency band, and that any grading flow prompts for Content Integrity whenever the medium is visual. The distinction between a photograph, an illustration and a graphic is Content Integrity's axis and never Transparency's.
01.
How a band is reached
Involvement is estimated as a single percentage of the finished work, and the percentage selects the band directly.
- 1
Choose an entry
Human Created, AI Involvement, or Fully AI. The two declarations sit outside the graded scale.
- 2
Set the percentage
A slider from 1 to 99 estimates how much of the work AI contributed.
- 3
The percentage selects the band
The band follows directly from the number, and the number is published alongside it.
- 4
Record the card
The Report Card carries the grade, the percentage and a written contribution summary.
02.
Two things v1.1 could not do
Neither was a defect in the published model. Both are the reasons v2.0 exists.
Known reading
The percentage reads as a measurement
A single number invites the reader to treat involvement as measured rather than judged. This is the reason v2.0 publishes points and stops publishing percentages.
Known reading
One scale for every kind of work
A generated image and a drafted report were scored on the same undifferentiated scale, so the same number described very different production processes. v2.0 answers this with per-medium weights.
01.
Seven bands, five graded and two declared
The graded scale is open at both ends. Minimal requires a ratio above zero and Extensive stops short of one, so neither endpoint belongs to any graded band. Both return a declaration instead.
No AI Involvement
TM-NONE@2.0
Declared. Gate answered no.
Minimal
TM-A@2.0
0 < r ≤ 0.10 of points possible
Limited
TM-B@2.0
0.10 < r ≤ 0.36 of points possible
Moderate
TM-C@2.0
0.36 < r ≤ 0.63 of points possible
Considerable
TM-D@2.0
0.63 < r ≤ 0.89 of points possible
Extensive
TM-E@2.0
0.89 < r < 1 of points possible
AI Generated
TM-GEN@2.0
Declared. Ratio equals 1, every activity AI led.
Under no circumstance is either endpoint assigned by calculation, and under no circumstance is a ratio rounded into an adjacent band to avoid the signal. Declaring is a human act and remains one.
01.
Five graded bands and two declarations
The scale a reader meets on any disclosure issued before v2.0. The band widths are the same widths v2.0 uses, which is why the two are comparable.
Human Created
TM-NONE
Declared. No AI involvement at any stage.
Human Led
TM-A
1 to 10% AI involvement
AI-Assisted
TM-B
11 to 36% AI involvement
Collaborative
TM-C
37 to 63% AI involvement
AI-Generated
TM-D
64 to 89% AI involvement
AI Led
TM-E
90 to 99% AI involvement
Fully AI
TM-GEN
Declared. No human authorship beyond the initial brief.
Grades A and E are deliberately narrow. Both are strong claims that require genuine confidence. Grade C is the widest band because collaborative working is the most common pattern across professional knowledge work.
01.
What changed at version 2.0
Version 2.0 replaces a percentage with a judgement about a finished piece of work. Every disclosure already published remains valid, and no card anyone has embedded needs changing.
If you already have a disclosure
Nothing to do.
Every code issued under version 1.1 remains valid and continues to mean exactly what it meant when it was issued. Version 1.1 and version 2.0 share the same band boundaries, so a disclosure made under either is describing the same thing on the same scale. Cards embedded on other websites keep rendering with the labels of the version they were issued under, and will do so permanently.
If you want to re-score a piece of work under version 2.0 you can, and you may find it arrives at a different band. That is because the way the answer is reached has changed, not because the scale moved.
01
They stay valid
A v1.1 disclosure carrying a percentage remains a true statement about work assessed under that model. v2.0 does not reach the historical record, and nothing is restated.
02
They aggregate together
Because the boundaries are identical, v1.1 and v2.0 disclosures may sit in one aggregate. It displays v2.0 labels and states the version span, because of the collision above.
03
Cards in the wild keep working
The embed contract accepts a percentage parameter for ever. It is accepted, never emitted, never a headline figure. An unversioned band resolves to v1.1 and renders v1.1 labels.
02.
Why the model changed
The first version asked one question and expected a percentage. How much of this work was AI, on a slider from one to ninety-nine.
Two things were wrong with that.
A percentage implies a measurement, and nobody was measuring anything. A person moved a slider to where their judgement told them, and the framework published the result as though it had been counted. That is a claim of precision the model could not support, on a framework about honest disclosure.
And one scale could not describe work as different as a strategy document and a generated image. Writing a report and producing a photograph involve different activities in different proportions, and a single question could not tell the difference. The same figure meant different things depending on what was being described, and nothing on the card said so.
03.
What version 2.0 does instead
You choose the medium first. An article, a report, an image, a social post, a video, a website, an email, a presentation or a bid. Each has its own list of the activities that go into producing it.
Then you answer once per activity, choosing between a person did this, AI helped, and AI did this.
Those answers produce points earned of points possible, and that ratio selects a band. The bands are Minimal, Limited, Moderate, Considerable and Extensive.
Two further declarations sit outside that scale. No AI Involvement and AI Generated are stated rather than calculated, because they are assertions about the work rather than measurements of it. No combination of answers reaches them: a person declares them, and puts their name to that.
04.
What that changes in practice
The score is a fraction rather than a percentage
Twenty-two of forty tells you what was asked and what was answered. Sixty per cent tells you a number.
Scores are not comparable across media
A report or strategy document has 38 points possible and an email or newsletter has 26, so the same band means the same proportion of a different thing. That is why a disclosure always states its medium, and why a set of disclosures is shown as a distribution rather than an average.
Band names replace letter grades
What was Collaborative is now Moderate. The boundaries have not moved.
Codes now carry their version
TM-C@2.0 says which band and which ruler. A code without a version is incomplete, because the same band name has not always meant the same thing.
The change, axis by axis
| Axis | v1.1 | v2.0 |
|---|---|---|
| Bands§1.1 | Seven. Two declared endpoints outside a five-band graded scale. | Seven. Same shape, and the boundaries are byte-identical. |
| Band labels§1.1 | Human Led, AI-Assisted, Collaborative, AI-Generated, AI Led. | Minimal, Limited, Moderate, Considerable, Extensive. |
| Endpoint labels§1.1 | Human Created at 0%, Fully AI at 100%. | No AI Involvement, AI Generated. Both declared, never calculated. |
| Input§1.3 | One percentage, chosen on a slider. | One answer per activity in the chosen medium: the human did this, AI helped, AI did this. |
| Score published§1.3 | A percentage. | Points earned of points possible. A percentage is never published, displayed or stored as the headline. |
| Weights§1.8 | None. The scale was medium-agnostic. | Per medium. Eight ratified definitions, each with an exhaustive reachability check. |
| Code form§1.5 | A band letter, often unversioned in the wild. | TM-[band]@[version], canonical uppercase. A code without a version is invalid. |
| Both endpoints§1.1 | Reachable by moving the slider to either end. | Neither is reachable by calculation. Both return a declaration-required signal. |
05.
The label collision
At version 1.1, the fourth band was labelled AI-Generated. At version 2.0, AI Generated is the declaration above the top of the scale. The same words, in two versions, mean different things.
At v1.1, band D was labelled AI-Generated. At v2.0, AI Generated is the declared endpoint above band E. The same words now mean two different things, and one of them is not even a graded band.
Any surface rendering a historical v1.1 disclosure renders it with its v1.1 label and its version, never with the v2.0 label. This is a permanent obligation rather than a transitional one, which is why the version pin is never optional on a rendered band.
This is the reason the version switch at the top of this page is a permanent control and not a preview toggle. A reader looking at a disclosure has to know which ruler measured it.
Rendered as a v1.1 disclosure
AI-Generated
TM-D · graded band · 64 to 89%
Rendered as a v2.0 disclosure
AI Generated
TM-GEN@2.0 · declared · declared
06.
How this was decided
The nine media, their activity lists and their weights were ratified by the product owner, eight of them on 10 August 2026 and the registry as it now stands on 25 August. Each was checked exhaustively to confirm that every band is reachable, because a weighting that strands a band is a scale with a rung missing.
The framework is published by the Digital Tourism Think Tank under Creative Commons Attribution 4.0. Decision authority sits with DTTT as product owner. External input is sought and recorded and informs decisions without constituting approval. There is no standing governance committee, and the version log is the accountability.
The full definitions, the weights and the reachability results are published alongside the schema.