Framework · Imagery · 3 min read
Generated image policy template
A one-page policy with seven clauses (allowed, never, labelled, recorded, checked, tools, owner) that tells every designer where AI-generated images help, where they mislead, and how they are marked and tracked.
Mood, concepts and illustrations of ideas, in named places.
Marketing, empty statesReal products, places, results, or people presented as real.
No generated customersOn or right under the image, saying how it was made.
"AI-generated"Tool, date, prompt and source type in the library.
Keep Content CredentialsGenerated people checked as a set, with counts.
See the representation auditOnly approved tools, with terms and training data reviewed.
Commercial use allowedOne named person decides new cases and reviews the policy.
Every six monthsUse it when
- Designers already use image generators and each makes their own call.
- Someone wants a generated "customer" for a testimonial or a case study.
- Legal asks how the team meets the transparency rules for generated content.
How to use it
- Collect real cases. In one session with the team, list the generated images you have used or been asked for, and sort each into mood, record, person or promise. Output: 10 to 20 real cases on a board.
- Write Allowed and Never. Allowed: mood, concepts, illustrations of ideas, in named places. Never: real products people will receive, real places, events or results, people presented as real, anything a reader could take as a record. Output: the first two clauses.
- Write Labelled and Recorded. Say where the label goes (on or right under the image), its words ("AI-generated", how it was made, not "fake") and the component that shows it. Say what the library records: tool, date, prompt, and the Digital Source Type set to generated, keeping any C2PA Content Credentials the tool adds. Output: clauses three and four.
- Write Checked and Tools. Generated people are checked as a set, with counts, before use. Only approved tools, whose terms allow commercial use and whose training data and provenance marks you have reviewed. Output: clauses five and six.
- Name the Owner and the review date. One person decides cases the policy doesn't cover and reviews it on a set date. Output: clause seven, with a name and a date.
- Put it where work happens. Link it from the image style guide, the library and the review checklist. After a month, review the real uses and change it openly, with a date and a reason. Output: a living page, not a PDF.
Worked example
The bakery ordering app has three requests in one week: a dreamy wheat-field image for the "Our flour" page, a generated smiling customer for a review quote, and a generated photo of the new rye loaf before the bakers have baked it. With the policy: the wheat field is mood, allowed, with an "AI-generated" label in the caption component and the record in the library. The customer is a person presented as real: never, they use a real review with permission or none. The rye loaf is a promise of what people will get: never, the team waits for the first batch and shoots it. The owner, the design lead, adds the rye case to the policy's examples.
Watch out for
- A policy that only repeats the law. In the EU, the transparency rules are the floor. Label what could mislead, not only what the law names.
- A label on everything. If every image says "AI-generated", the label means nothing. Label generated images, and only those.
- Losing the record. Exports and edits can strip metadata and Content Credentials. Keep the master with its record in the library.
- No owner. Without a named person, the edge cases go back to whoever is on deadline.
Template
Adapt the brackets, keep the shape.
GENERATED IMAGES: OUR POLICY (version ..., ... date)
Allowed for mood, concepts and illustrations of ideas, in [marketing, editorial, empty states].
Never for real products people will receive; real places, events or results;
people presented as real (customers, employees, testimonials);
anything a reader could take as a record.
Labelled on or right under the image: "AI-generated", in the [label component].
Recorded in the image library: tool, date, prompt, Digital Source Type = generated;
keep any Content Credentials (C2PA) the tool adds.
Checked for bias as a set, with counts, before any generated people are used.
Tools only [approved tools]: commercial use allowed, training data and provenance reviewed.
Owner [name] decides cases not covered here and reviews this policy every [six months].
- Atlas of AIKate CrawfordBook · 2021 (Amazon)
- C2PA (external site)
- Content Credentials, verify (external site)
- Digital Source TypeIPTC (external site)
- EU AI Act, Article 50 (external site)
"See on Amazon" links are affiliate links. As an Amazon Associate I earn from qualifying purchases, at no extra cost to you.