Social Justice and Prompt Engineering: What We Know So Far

The talk opens with a short primer on prompt engineering, covering few-shot examples and chain-of-thought prompting. It then argues that large language models absorb the prejudices in their training data, so the way a request is worded can reduce harm for both users and builders.

Several recent studies are reviewed. One audit of a commercial model posed yes-or-no decisions about hypothetical applicants and found demographic skews; telling the model that discrimination is illegal and to ignore protected traits worked best. Another study found far less bias when models judged one candidate in isolation than when they compared two, and that instructions to treat groups equally cut bias. Supplying outside data, such as a market price, reduced lowball offers tied to a seller's name. Text-to-image studies showed skewed portraits, uneven quality and refusals on sensitive prompts, and multilingual prompts shifted results depending on grammatical gender.

The speaker also replicated a reference-letter study, where men were described by achievements and women by personality, and found that added context and fairness instructions moved the output toward parity. The takeaways: keep high-stakes decisions away from models without human review, ask for absolute rather than relative judgments, anchor prompts with external data, and remember that removing names is not enough because models infer identity from other cues. Expect outputs to shift with small wording changes, and avoid lock-in to a single vendor.