AI Advice Reinforces Traditional Gender Roles
A · Strong evidence Harvard Kennedy School and MIT Sloan negotiation researchers, including Hannah Riley Bowles, ran an exploratory and a preregistered audit across four leading models with 86,000 queries and report effects with confidence intervals.
Summary based on the full text·Original: about 30 minutes to read
In short
An arXiv paper (1 October 2026) by Nadja Born, Jared Curhan and Hannah Riley Bowles. When asked for job-negotiation advice for identical candidates who differ only in gender cues, four leading LLMs recommend lower salary targets to women. They also set different agendas: flexibility, caregiving and well-being for women; equity, mobility and pay for men. These differences track how people rate each term as a breadwinner or caregiver concern, so the advice steers users toward traditional roles.
Key numbers
- 86,000independent queries across the two audit studies
- $5,820higher pooled salary target recommended for male-named vs female-named candidates (Study 2)
- $1,594 to $15,915range of the male-female salary gap across the four models
- OR = 2.32odds ratio for remote or hybrid work being prioritised for women
- ρ = −0.71correlation between AI gender gaps per term and human breadwinner-caregiver ratings
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- Study 1 (explicit gender, 6,000 queries): men got higher salary recommendations in all six occupation-by-age cells; average gap $2,408 (95% CI 2,097-2,720).
- Study 2 (preregistered, names only, 80,000 queries on GPT-5.4-mini, Claude Sonnet 4.6, Gemini 2.5 Flash, Llama 3.3 70B): every model recommended higher targets to men; pooled gap $5,820, model gaps $1,594 (Llama) to $15,915 (Claude Sonnet 4.6).
- Men also got wider negotiation ranges: pooled difference $2,961.
- Study 1 term mentions: childcare support 80% to women, parental leave 69%, client engagement 89% to men.
- Study 2 rankings of 25 terms: remote or hybrid work (OR 2.32) and parental leave (OR 2.26) favoured for women; stock options (OR 0.61) and relocation (OR 0.63) for men. 24 of 25 terms differed significantly.
- 500 US adults rated each term on a breadwinner-caregiver scale; AI gender gaps correlate with those ratings at ρ = −0.71, across all four model families.
How they got it
Audit design: qualifications, experience and context fixed, only gender cues varied. Study 1 used open-ended prompts modelled on executive-education participants' real use, with terms extracted by rules, NLP and clustering into a taxonomy. Study 2 used a standardised salary-plus-ranking task with multiple first names and pronouns. Human ratings were collected on Prolific as preregistered. Model IDs, dates and sampling settings are in the supplement.
Limits
- Only stereotypically White names; race and ethnicity interactions are not tested.
- Audit outputs from single prompts; no measure of how users act on the advice or of negotiation results.
- Gap sizes vary widely by model, so the pooled dollar figure depends on weighting.
- Models tested are mid-tier or older versions; current flagship models may behave differently.
Why it matters for companies
Staff already ask chatbots how to negotiate offers, raises and flexible work, and the advice differs by gender even when nothing else does. Firms deploying career coaches, internal mobility assistants or manager copilots should run the same kind of name-swap audit before launch, covering non-salary terms as well as pay. Pay transparency policies catch salary gaps but not agenda steering on flexibility, equity and relocation, which is where this bias sits. Vendor bias claims that test only salary outputs miss most of the effect.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org
- Tags
- Work and AI, Bias