Understanding and reducing bias in AI outputs

0:00 Bias in AI outputs

Type the words successful entrepreneur into an image generator and look at the first eight results. Who do you see? And more importantly, who is missing? In this lecture you'll learn where bias in AI outputs comes from, how it shows up in everyday marketing work, and a five-step review routine you can run on your next campaign in under an hour. You'll also see me run a quick bias audit, live, on a real kind of brief.

0:34 Why it matters

Why should a marketer care? Three reasons. First, money. Your audience is more varied than a model's defaults, and people notice when they are left out. Second, harm. Stereotypes in ads shape how people are seen and treated. Third, rules. In the UK, the advertising code restricts harmful gender stereotypes. Equality and anti-discrimination laws cover areas like hiring, housing and credit in many countries. And ad platforms restrict discriminatory targeting in sensitive categories. So bias isn't only an ethics question. It is a performance question and a compliance question at the same time.

1:14 Where bias comes from

Here's the key idea. A model is like a very well read intern who has only read the internet's old newspapers. It has seen millions of pictures of executives, nurses and families, and it learned the most common patterns. When your prompt is vague, it falls back on those patterns. That's why bias usually isn't programmed in on purpose. It comes from four places: imbalanced training data, historical patterns, design and tuning choices made by the developer, and your own prompt. You control only the last one directly. But the last one matters more than most people think.

1:57 Where bias shows up

Where does it show up? In images, when a family is always one structure, or beauty defaults to lighter skin, which is a sensitive issue in South Asian and Gulf markets. In copy, when personas lean on clichés like busy moms who love shopping. In translation, with wrong honorifics or misgendering in Arabic and Urdu. In targeting and scoring, when a lead-scoring model quietly penalizes names or postcodes that correlate with ethnicity or class. And in moderation, when tools flag Roman Urdu or Arabizi comments as spam. Notice that the last two are about decisions, not pictures. Those are the higher-stakes ones.

2:41 Five-step routine

Now the routine. Five steps. One, prompt deliberately. Describe the real audience: ages, skin tones, dress, settings, roles. Two, generate variety and look at outputs as a set, because patterns only appear across many images. Three, review with a checklist. Who is shown? Who is missing? Is anyone shown in a lower-status role or through a cliché? Four, get lived-experience feedback from people in the communities you portray. And five, keep humans responsible for decisions about people. AI can organize information about creators, leads or applicants. A person with clear, fair criteria makes the call.

3:22 Example 1: Lahore bakery

First example, a simple one. A Lahore bakery wants a social post showing happy customers. The vague prompt returns eight images of young couples in western dress. The owner knows her real customers: families, grandparents buying for Eid, office workers at lunch, students. She rewrites the prompt with those groups and settings, generates eight more, and picks three that look like her shop on a Friday afternoon. Total extra time? About ten minutes. The post looks like her customers because she told the model who her customers are.

4:00 Example 2 (illustrative): creator ranking

Second example, a realistic business scenario with illustrative details. An agency in Karachi builds a creator-selection tool for a beauty brand. It ranks creators using engagement and audience data, and a language model summarizes each profile. After a month, the account lead notices something. Creators with darker skin, and creators writing in Roman Urdu, keep landing in the bottom third. Nobody designed that. It came from the historical campaign data the ranking learned from. The agency responds by removing proxy features, adding a human review of every shortlist, and checking the mix of the shortlist against the brand's audience each month. Illustratively, the shortlist became far more representative within two cycles, and campaign engagement held steady.

4:50 Watch me do it: mini audit

Watch me do it. I'm reviewing eight AI images for a small-business banking app. I lay them out as a grid and code each one in a sheet: perceived gender, age band, skin tone band, role shown. Here's what I see. Seven of eight are men. All eight are under forty. The one woman is behind the counter serving a man in a suit. That's the finding. Now the fix. I rewrite the prompt to vary gender, age, skin tone and shop type, and I tell the model to show every owner as the expert. I regenerate, recode, and compare. Then I send the shortlist to two customer-facing colleagues and ask one question: do these look like our customers? I record their answer in the campaign folder.

5:45 Common mistakes

Common mistakes. Tokenism, where one diverse face is added while the story stays stereotyped. Over-correcting in ways that feel fake or historically wrong, like a medieval scene that ignores the period. Treating a tool as neutral because it's a machine. Reviewing one image at a time, so patterns never show up. And the big one: letting a model rank or filter people without a human step and without ever checking outcomes. If you use AI to shortlist creators, leads or applicants, bias there costs real people real opportunities.

6:23 Measure it

How do you know it's working? Measure it. Track a representation gap: your audience mix compared with the mix in the creative you publish. Track review coverage: how many public campaigns went through the checklist before launch. Aim for all of them. Count representation-related complaints and comments per campaign. And for any system that ranks people, keep a decision log and review outcomes by group periodically, where you can lawfully measure it. What gets measured gets fixed.

6:56 AI as a second reviewer

Can you use AI to review AI? Yes, as a second pair of eyes, not as the judge. Here's how. Describe each image or paste each persona into a chatbot, give it your real audience profile, and ask it to flag stereotyped roles, missing groups and patronizing depictions in a table with a suggested prompt change for each. Tell it not to invent details. It will catch things you miss, especially when you've been staring at the same creative all afternoon. But models carry the same blind spots they learned from, so the final review still belongs to people, ideally people from the communities in the images. Think of the AI reviewer as a spell-checker for representation: useful, fast and never the last word.

7:50 Recap + try this now

Let's recap. Bias mostly comes from data and defaults, and your prompt is the lever you control. It shows up in images, copy, translation, and in decisions about people, which carry the highest stakes. Prompt deliberately, review as a set with a checklist, ask people with lived experience, and keep humans in charge of decisions. Try this now: run the vague versus deliberate prompt test from the lesson, eight images each, and code them in the sheet. Then read the lesson text for the full audit template. Next, we tackle misinformation.