Bias, Fairness, and Responsible AI Use

The concept

An LLM, as Module 1 established, learns statistical patterns from enormous amounts of existing text and images — and existing text and images carry the biases of the world that produced them. A model trained on decades of job postings that skewed toward describing engineers as "he" will reproduce that pattern unless specifically corrected for; an image model trained on a training set where "CEO" disproportionately returns a narrow demographic will generate that narrow demographic by default when asked for one. This isn't a bug introduced by a careless engineer — it's a direct, mechanical consequence of learning patterns from real-world data, which means it shows up in every major AI tool to some degree, not just the ones with less investment in mitigating it.

This makes "the AI said so" a uniquely bad justification for any output touching people or decisions about people — hiring language, performance reviews, loan or eligibility summaries, generated portraits meant to represent "a typical" anything. The AI isn't a neutral arbiter; it's a mirror reflecting patterns in its training data, including the unfair ones, back at you with the same confident tone it uses for everything else. Treating its output as automatically fair because it came from a machine rather than a person is exactly backwards — a human writer at least has the chance to notice their own bias in the moment; a model reproduces a pattern without any equivalent moment of self-awareness.

Three concrete checks catch most of what matters in practice. Stereotyping: does the output describe a role, trait, or behavior in gendered, raced, aged, or otherwise group-coded language where the actual content doesn't call for it — a job description that uses "aggressive" and "dominant" for a sales role in ways that skew toward coding it as a particular gender, for instance? Framing asymmetry: does the output favor one group's perspective, interests, or default assumptions over another's when the prompt didn't ask it to — a summary of a policy debate that treats one side's claims as established fact and the other's as "claims," without a factual basis for the asymmetry? Swap-test sensitivity: if you ran the exact same prompt with one demographic detail changed — a name, a pronoun, an age, a role — would the output change in a way that isn't justified by anything actually in the prompt? A meaningfully different answer to an otherwise identical question is the clearest, most checkable signal that bias is present.

These checks don't require special tools — the swap test in particular is something anyone can run in under a minute by re-asking the same question with one detail changed and comparing. The habit to build is running it on anything that touches people or decisions about people, before using the output, not treating bias-checking as a specialist's job separate from your own.

Worked example

A learner generates a short job description for a "warehouse operations lead" and separately generates a set of four AI portrait images for "a professional in a leadership role" to use on an internal slide. Reviewing the job description, they notice the language leans heavily on words like "commanding," "assertive," and "able to take charge of a room" — phrasing that, while not naming a gender, tends in practice to code the role as masculine and may unintentionally discourage qualified candidates who don't see themselves reflected in that language. Reviewing the four portrait images, all four show a similar age range and a similar demographic profile, despite the prompt never specifying either — the model defaulted to whatever was statistically most common for "leadership role" in its training data.

Running the swap test on the job description, the learner re-prompts with "write this in a way that's equally welcoming to all qualified candidates regardless of background" and gets noticeably softer, trait-neutral language ("organized," "clear communicator," "comfortable coordinating a team") — confirming the first version wasn't a neutral default but a specific, correctable skew. For the images, they revise the prompt to explicitly request a demographically varied set, producing four images with a visibly different and more representative range on the second attempt.

Exercise

Take one AI output from earlier in this course — a job description, a summary, a drafted message, or a new one-line prompt you write now that touches a person or a group of people in some way. Run it once as written, then run the exact same prompt again with one demographic detail changed: a name, a pronoun, an age, a stated role or background.

Compare the two outputs side by side and note any difference that isn't justified by anything actually present in the prompt — a change in tone, assumed competence, assumed background, or visual representation. If you find a meaningful difference, note what a corrected version of the prompt would ask for instead (an explicit instruction for neutral or representative output, as in the worked example) to close the gap.

Recap

• Bias in AI output is a direct, mechanical consequence of learning patterns from real-world training data — it shows up in every major tool to some degree, not as a rare glitch.
• "The AI said so" is never a neutral justification for output touching people or decisions about people — check for stereotyping, framing asymmetry, and swap-test sensitivity before using it.
• The swap test — rerunning an identical prompt with one demographic detail changed — is a fast, checkable way anyone can catch unjustified differences, and takes under a minute to run.

Resources

• NIST: "AI Risk Management Framework" — https://www.nist.gov/itl/ai-risk-management-framework
• Partnership on AI: "Responsible AI practices" — https://partnershiponai.org/
• Anthropic: "Claude's constitution" — https://www.anthropic.com/news/claudes-constitution