Understand the technology
From spotting a pattern to improving a workflow.
For May and November 2027. This lesson is for SL and HL students. Learn through practical studio examples, a worked audit and a mitigation plan that acknowledges limitations.
01 | WHEN AN IMAGE MEETS THE BRIEF BUT MISSES THE PEOPLE
Imagine Visionary Studios generates a set of advertising images showing engineers. The images are sharp and match the requested style, but almost everyone shown looks similar in age and background. Technical quality alone has not produced a suitable campaign.
Bias is a systematic tendency that can skew a model’s outputs. In this lesson, the concern is patterns that create misleading, exclusionary or unfair representations of people. One surprising image is a reason to investigate; repeated patterns across comparable requests provide stronger evidence.
This is a fictional classroom example, not a finding reported about Visionary Studios in the case-study booklet.
02 | LEARNED PATTERNS ARE NOT NEUTRAL FACTS
An image model learns statistical relationships between visual features and accompanying descriptions. If its examples repeatedly associate a profession with one type of person, that association can influence generation.
The model is not deciding what is morally right. It produces patterns shaped by its training and design. A convincing image can therefore reproduce a social stereotype rather than offer a balanced picture of the world.
For the studio, the important question is not simply whether an output looks realistic, but whether it is appropriate for the intended audience and message.
03 | REPRESENTATION BIAS: WHO APPEARS?
Representation bias can arise when relevant people or contexts are missing or poorly represented in the data. A collection dominated by one setting gives the model fewer examples of alternatives.
Suppose a campaign needs to show people using public transport in several countries. A dataset dominated by one city might produce familiar stations and clothing even when a different location is requested.
Adding more files from the same narrow sources may increase dataset size without addressing the gap. Relevant coverage matters as well as volume.
04 | STEREOTYPES: HOW ARE PEOPLE SHOWN?
Counting appearances is only a starting point. People may appear equally often yet repeatedly be assigned different roles: one group leads meetings while another always takes notes.
A stereotype is an oversimplified generalisation about a group. In images it can appear through occupation, authority, clothing, setting or behaviour.
For Visionary Studios, review both presence and portrayal. Ask who is central, who has agency and whether the same narrow assumptions recur.
05 | LABELS AND CAPTIONS CAN ADD BIAS
A training image is often paired with text. Captions help connect words to visual patterns, so inaccurate or uneven descriptions can shape those associations.
For example, one set of people might regularly be described by professional role while another is described mainly by appearance. The text supplies different signals even if both groups are visible.
Curation should therefore examine captions and labels as well as pixels. Correcting misleading descriptions can address a different problem from collecting additional images.
06 | BIAS CAN ENTER AT SEVERAL STAGES
Source choices shape coverage.
Labels, filtering and objectives shape learned relationships.
Prompts and human choices shape what reaches the client.
Bias is not only a training-data issue. A studio might generate varied candidates but repeatedly select the same narrow portrayal. A mitigation plan must follow the whole production process.
07 | WHAT DOES FAIRNESS MEAN HERE?
Fairness involves judging whether people are treated and represented appropriately in context. There is no single percentage that settles every creative brief.
A campaign depicting a particular historical event has different requirements from a general advertisement about modern workplaces. The studio should explain the purpose, audience and relevant representation criteria before evaluating results.
Fairness does not always mean identical numbers in every image. It does require defensible decisions about inclusion, portrayal and possible harm.
08 | LISTEN TO AFFECTED PEOPLE
The person commissioning an image may not notice every misleading or offensive detail. Reviewers with relevant cultural knowledge and people affected by a portrayal can identify issues a narrow team overlooks.
This is not a request for one person to speak for an entire community. Gather a range of perspectives, record disagreements and explain how decisions are reached.
For the studio, consultation can turn a vague instruction such as “be inclusive” into clearer review questions about roles, settings and audience interpretation.
09 | DESIGN A COMPARABLE TEST
Testing one favourite prompt is weak evidence. Build a set of realistic briefs, generate several outputs per brief and record the model version, settings and review criteria.
When investigating a particular effect, keep other conditions as comparable as practicable. Random generation means a single result may be unusual, so inspect repeated outputs rather than select only examples supporting an initial opinion.
Use an agreed rubric and record why reviewers flag an image. A repeatable method makes a later comparison more useful.
10 | A WORKED AUDIT: COUNTS AND PORTRAYAL
| Measure | Initial workflow | Revised workflow |
|---|---|---|
| Images flagged for stereotyped roles | 12 of 40 | 5 of 40 |
| Flagged proportion | 30% | 12.5% |
| Difference | 17.5 percentage points lower | |
The initial rate is 12 ÷ 40 × 100 = 30%. The revised rate is 5 ÷ 40 × 100 = 12.5%. Subtracting gives a reduction of 17.5 percentage points , not 17.5% relative reduction.
This suggests improvement on this rubric and sample. It does not prove the model is fair for all audiences, prompts or groups. Review the remaining issues and repeat testing on further representative briefs.
11 | NUMBERS NEED CAREFUL INTERPRETATION
A total can hide uneven outcomes. A workflow might improve overall while still producing poor portrayals in one setting or for people with intersecting characteristics.
Small samples also limit conclusions. A category appearing twice in a test provides little evidence about the full range of outputs.
Do not casually infer a person’s identity from their appearance or build sensitive labels without a suitable basis. Use an appropriate, respectful review process and state what the measurements can actually establish.
12 | MITIGATION THROUGH DATA CURATION
Bias mitigation means taking steps to reduce identified bias and its effects. Start with the diagnosed cause: fill relevant coverage gaps, review misleading labels and remove inappropriate repetition.
More balanced data may give a model a wider range of useful examples. However, equal counts alone do not ensure varied roles, accurate captions or fair outputs.
Keep provenance and permission checks alongside representation review. Fixing a fairness concern does not remove other responsibilities.
13 | MITIGATION THROUGH PROMPTS AND GENERATION
A clear prompt can specify relevant roles, settings and inclusive creative requirements. Generating a wider candidate set can help the studio find appropriate outputs.
These are workflow interventions, not proof that the underlying model has been fixed. A model may still handle some requests poorly or follow a stereotype despite the wording.
Evaluate prompt changes using comparable tests. Avoid adding identity details mechanically when they do not fit the brief or when they reinforce the assumptions being challenged.
14 | MITIGATION THROUGH HUMAN REVIEW
Before publication, reviewers can flag stereotyped roles, inaccurate cultural details or exclusionary patterns across the campaign. They can request regeneration, editing or a different production method.
A checklist helps make the process consistent, but people can also introduce bias. Use clear criteria, a way to resolve disagreements and escalation for difficult cases.
Review the complete set of chosen assets. Individually acceptable images can still create an unbalanced pattern when displayed together.
15 | CHOOSING A PROVIDER OR DIFFERENT WORKFLOW
If the studio uses an external model, it may be unable to change the training dataset. It can still examine provider disclosures, run its own tests, choose settings and decide whether the tool is suitable.
A provider’s fairness claim is a starting point for questions, not a substitute for evaluating the studio’s actual briefs. Compare relevant evidence and note what is unknown.
Where important requirements remain unmet, use another model, commissioned artwork or a mixed workflow. Responsible evaluation includes deciding not to use a tool for a particular task.
16 | TRADE-OFFS AND LIMITATIONS
Additional review takes staff time, consultation needs planning and collecting suitable data can be costly. Those costs should be weighed against poor campaign quality, exclusion and damage to trust.
Some interventions may also create new distortions if applied without context. The aim is to improve justified outcomes, not merely optimise one count while ignoring portrayal.
A strong evaluation explains the benefit, cost and remaining uncertainty of each proposed measure. No single safeguard guarantees that every output will be appropriate.
17 | MONITOR AFTER THE FIRST IMPROVEMENT
A model update, new client brief or different audience can change results. Keep test records and repeat relevant checks when the workflow changes.
Provide a route for clients and affected people to report concerns, then investigate and revise the workflow where justified. Feedback should lead to documented decisions rather than disappear into an inbox.
18 | BUILD A JUSTIFIED STUDIO RECOMMENDATION
Suppose Visionary Studios finds repeated stereotypes in workplace images. Recommend a response that includes identifying the pattern, testing it, applying mitigation and checking the result.
Reveal a possible response
Audit several relevant briefs using an agreed portrayal rubric and record repeated outcomes. Review data coverage and captions if the studio controls training, and test clearer creative requirements. Use informed human review before publication and compare the revised workflow with the initial one.
Keep evidence of remaining weaknesses and consider another production method where requirements are not met. A lower flagged rate supports a limited improvement claim, not a guarantee of universal fairness.
Use these independent teaching notes alongside the official 2027 case-study booklet. The numerical examples and studio scenarios on this page are classroom illustrations.
CHECK YOUR UNDERSTANDING | CLASSIFY THE FAIRNESS CHECK