Audit Gender Representation in AI-Generated Images

Auditing gender representation in AI-generated images helps reveal who appears, how they are portrayed, and whether patterns change across prompts or tools. You will need a defined sample, a consistent coding plan, and careful handling of uncertain or sensitive identity; the audit describes observable output, not the gender identity of real people.

What You’ll Need

• A set of images or media outputs collected under documented conditions, including the prompts and generation settings where available.
• A spreadsheet or other table for recording samples, coding decisions, and notes.
• A written codebook defining categories, such as perceived presentation, role, activity, clothing, and visibility. Include “unclear,” “not applicable,” and “not represented” where appropriate.
• At least one reviewer; a second independent reviewer is useful for checking consistency.
• A plan to store files securely, avoid unnecessary personal data, and follow the tool’s terms and any applicable research or organizational requirements.

Step-by-Step

1. State the audit question. Make the goal specific before collecting anything. For example: “Across prompts for a project manager, what gender presentations appear, and who is shown leading?” Decide whether you are assessing image frequency, visual roles, stereotypes, or differences among systems. Avoid framing the task as discovering a model’s intent.
2. Define the scope and unit of analysis. Choose the systems or media types, the date range, and the population of prompts you will test. Decide whether one generated image, one person within an image, or one prompt-output set is the unit. For scenes with multiple people, record each person separately and keep a shared image ID so people are not mistaken for independent image samples.
3. Create a prompt sampling plan. Use a balanced set of prompts that varies relevant factors while keeping other wording consistent. Include neutral prompts and, if relevant, prompts that specify a role or activity. Record the exact text, generation date, tool or model label, settings, number of outputs, and any edits. Do not quietly discard awkward results; define exclusion rules in advance, such as corrupted or blank files.
4. Generate or collect outputs consistently. Keep settings as similar as the system allows, including aspect ratio, style options, and number of outputs. If a tool changes over time, note that; do not assume results from different dates are directly comparable. Save original outputs and a copy of the prompt log. Avoid using real people’s faces or sensitive personal material unless you have a clear, appropriate basis to do so.
5. Write a codebook before coding the full sample. Define each field and how to use it. For instance, “visual gender presentation” should mean a reviewer’s perception from visible cues, not a claim about identity. Provide categories that fit the audit and allow uncertainty rather than forcing every person into a binary. Add separate fields for role, action, camera prominence, clothing, and context if they address your question. Describe concrete cues and examples, while noting that cues vary by culture and do not reliably establish identity.
6. Pilot the codebook on a small subset. Have reviewers code the same images independently, then compare disagreements. Clarify ambiguous definitions, add missing options, and repeat the pilot if necessary. Do not resolve uncertainty by guessing. If the media does not show enough to code a field, use “unclear” or “not visible,” as defined in the codebook.
7. Code the full set and keep an audit trail. Reviewers should apply the same rules and record codes without changing them to fit expected results. For each decision, retain the image ID and, where needed, a short note explaining ambiguity. If one person codes everything, have another reviewer check a meaningful subset. Keep disagreements visible and document how they were resolved.
8. Summarize counts and patterns carefully. Report the sample size and the number of images or people in each category, including unclear cases. Calculate percentages using an explicitly stated denominator; for example, people coded for visual presentation, not all generated files. Compare groups only when prompts and conditions are sufficiently similar. Separate counts from interpretation: a higher share in one sample does not by itself establish why the difference occurred.
9. Review portrayals, not just presence. Inspect who is shown as an expert, leader, caregiver, or background figure; who is active or passive; and whether clothing, age, body type, or setting repeatedly signal narrow stereotypes. Use fields that were defined in advance, and include representative examples only when sharing them is appropriate. A count alone can miss meaningful differences in agency or context.
10. Write up the method and limitations. State the prompts, sample size, tools and dates, settings, codebook, reviewer process, exclusions, and calculation method. Explain that appearance-based coding is subjective and cannot determine a depicted person’s gender identity. Note limits such as a small or nonrepresentative prompt set, system updates, cultural assumptions, and variation between generations. Share only what is needed, with access controls for sensitive files.

Common Mistakes

• Treating appearance as identity: A visual audit cannot verify whether a person is a woman, man, nonbinary, or any other identity. Label the variable as perceived presentation or visible cues and state its limits.
• Forcing binary categories: This erases ambiguity and can distort findings. Include nonbinary or mixed presentation where appropriate, plus unclear and not-visible options; do not assume every sample supports every category.
• Changing prompts between comparisons: A result may reflect different wording, settings, or output counts rather than a system difference. Keep comparisons controlled and document unavoidable changes.
• Counting images and people as if they were the same: One image can contain several people. State which unit your percentages use and avoid mixing image-level and person-level totals.
• Reporting percentages without denominators: Always show counts and the relevant total. A percentage based on a small sample can look more conclusive than it is.
• Assuming a pattern proves intent or cause: The audit can document an output pattern, not establish why it happened. Avoid claims about training data, model motives, or real-world impact unless separately supported.

Tips

• Use prompt pairs that differ in only one factor when testing that factor, and repeat generations when output variation matters.
• Keep raw files, prompts, settings, coding sheets, and later revisions linked by stable IDs.
• Ask reviewers to code independently before discussing disagreements; this helps distinguish unclear rules from genuine differences in interpretation.
• Report uncertainty plainly. “Not enough evidence to classify” is more accurate than a forced answer.
• When publishing examples, consider whether the image could identify a real person, expose sensitive information, or reinforce the stereotype being discussed. Crop or omit examples when appropriate, and explain selection criteria.

Troubleshooting

Reviewers disagree often: Pause full coding, inspect the disputed cases, and refine the definitions or examples. Recode the affected subset using the revised codebook, and disclose that change. Do not hide disagreements by averaging incompatible judgments.

Too many outputs are unclear: Check whether the prompt or image format provides enough visible information for the field. Mark cases unclear rather than inferring identity from a name, job, or stereotype. You may revise the research question to focus on roles or visibility instead.

Results shift between runs: Increase repetitions if feasible, record random seeds or settings when exposed, and report variation. If the system or settings changed, split the analysis into separate conditions rather than pooling them without explanation.

The sample is unbalanced: Add outputs for underrepresented prompt conditions using the same procedure, or report the imbalance and limit comparisons. Do not silently remove excess cases to make the chart look even.

You cannot reproduce a result: Preserve exact prompts, dates, settings, output counts, and available model identifiers. Some systems change or do not expose all settings, so describe the result as a snapshot rather than a repeatable guarantee.

Frequently Asked Questions

Can an image audit determine a person’s gender identity?
No. It can record how a person may be perceived from visible presentation, but identity cannot be reliably established from an image.

How many images should I review?
There is no universal minimum. Choose a sample large enough for your question and resources, report its size, and treat small samples as exploratory rather than representative.

Should I use only neutral prompts?
Not necessarily. Neutral prompts can show default patterns, while controlled prompt pairs can test how a specified role or wording changes outputs. Keep these analyses distinct.

Should I count multiple people in one image?
Yes, if people are your unit of analysis. Record each person separately and link them to the same image ID; otherwise, state that you count images rather than people.

Can I compare two image generators?
Yes, if you use comparable prompts, settings, dates, and output counts. Document differences in controls or system versions, and avoid treating the comparison as proof of cause.

What should I do with ambiguous cases?
Use a predefined unclear or not-visible category, keep the case in the reported totals, and explain how it affects the denominator for each measure.

Related

• Run GPT-OSS Locally on Mac With LM Studio
• On-Device LLMs: How Hybrid AI Balances Privacy and Power
• Deno 2.5, Pragmatic Turn, Bytes and 423 Explained