Marcus Andrey Study
Ler em português Library

Framework · Imagery · 3 min read

Representation audit

A counting method for the people in your product's images: lay them all out, record who appears in which role, compare with who uses the product, and fix the brief, so stereotypes are found by evidence instead of by luck.

  1. 1Lay out

    Every image with people in the product, side by side.

    Include emails and empty states
  2. 2Record

    The role each image gives, and only what it plainly shows.

    Expert, helper, served
  3. 3Count

    How often each visible group appears in each role.

    "Who is always the expert?"
  4. 4Compare

    Put the counts next to who uses the product.

    "Who never appears?"
  5. 5Fix brief

    Change the brief, prompts and guideline, then the images.

    Worst cases first

Then ask someone outside the team, and repeat once a year

The pattern is in the set, not in any one photo, so the audit counts roles across the whole product.

Use it when

  • Before a new image library, shoot or generated set goes into the product.
  • Someone notices "all our experts look the same" and the team needs facts, not opinions.
  • Once a year, as part of the image audit of the product.

How to use it

  1. Lay out every image with people. Export or screenshot every image in the product that shows a person, side by side, including empty states, onboarding and emails. Output: one board with every person in the product.
  2. Record the role, not the identity. For each person, write the role the image gives them: expert, helper, customer, worker, served, decorative. Add only what the image shows plainly, such as an age band, a visible disability or assistive device, or hands-on work. Output: one row per person in a simple table.
  3. Count roles by group. Count how often each visible group appears in each role. Look for patterns: who is always the expert, who is always served, who never appears. Output: a small count table.
  4. Compare with your users. Put the counts next to who actually uses the product, from research or analytics, or who you want to reach. Output: the gaps, written as sentences ("No older customers appear; a third of orders come from customers over 60").
  5. Fix the brief, not just the image. Write who you want to see, in which roles, into the image brief, the shot list, the generator prompts and the image guideline. Replace the worst cases first. Output: an updated brief and a list of images to replace.
  6. Ask someone else. Show the board and the counts to people outside the team, ideally people the gaps affect. They notice exclusions faster. Output: their notes, added to the next brief.

Worked example

The bakery ordering app audits its 46 images with people. The table shows the bakers (the experts) are all young men, the customers are all young women with children, and nobody older or with a visible disability appears anywhere, while the team's own research shows many regular customers are retired neighbours who order for the week. The fix: a new shot list for the next shoot (an older baker at the oven, a regular customer with a walking stick collecting an order, hands-only shots for the ordering steps), and the image guideline now names the roles to vary. Six images are replaced in the first release, the rest at the shoot.

Watch out for

  • Guessing identity. You can't see someone's ethnicity, gender identity or disability from a photo. Record what the image shows and the role it gives, and say so in the report.
  • Auditing one photo at a time. Every single photo looks fine. The pattern is in the set.
  • Fixing images, not the brief. Replace the photos without changing the brief, prompt and guideline, and the same pattern comes back with the next batch.
  • Tokenism. One diverse photo on the home page while every expert stays the same is not a fix. Look at the roles.

Template

One row per person, then the summary.

Image | Where in the product | Role the image gives | What the image shows plainly | Generated? (y/n)
...   | ...                  | expert / helper / customer / worker / served | ... | ...

Summary
Images with people: ...   People counted: ...
Pattern 1: ... appear as ... in ... of ... cases
Pattern 2: ... never appear in the role of ...
Compared with our users (source: ...): ...
Changes to the brief / shot list / prompts / guideline: ...
Images to replace first: ...
Reviewed by (outside the team): ...
Source
Counting roles in a set: the method in the course's Bias in image generators lesson. Stereotypes in generated images: Bianchi et al., "Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale" (FAccT, 2023). Design led by the people it affects: Sasha Costanza-Chock, Design Justice (2020)

"See on Amazon" links are affiliate links. As an Amazon Associate I earn from qualifying purchases, at no extra cost to you.

Saved only on this device.

↑ ↓ to move · Enter to open · Esc to close