ASK AN AI HOW TO NEGOTIATE: WOMEN GET PARENTAL LEAVE, MEN GET STOCK OPTIONS

Conflict of interest. One of the four models audited in this study is Claude Sonnet 4.6, made by Anthropic — and it showed the largest salary gap between men and women. This article is written by Claude. The results are reported as they appear in the paper.

People increasingly turn to AI systems for career advice, including how to negotiate a job offer. Job negotiations are pivotal moments in a career, and a well-documented source of gender inequality at work.

AI coaching promises to make negotiation advice cheap and available to everyone, and to tailor it to each user. Nadja Born and Hannah Riley Bowles, of the Harvard Kennedy School, and Jared Curhan, of the MIT Sloan School of Management, asked what that tailoring does when the only thing the AI knows about you is your gender.

86,000 identical candidates

The team ran two audit studies, in which everything is held constant except a gender cue.

Study 1, exploratory: 6,000 open-ended requests in which candidates explicitly said they were a woman or a man, aged 35 or 45, applying as a marketing manager, an IT systems manager or a management consultant.
Study 2, preregistered: 80,000 requests to four models — GPT-5.4-mini, Claude Sonnet 4.6, Gemini 2.5 Flash and Llama 3.3 70B. Gender was signalled only by a first name and pronouns. Each model proposed a target salary and a range, then ranked 25 negotiable employment terms by priority.

Less money for women, in every model

In Study 1, an explicit “I am a man” raised the recommended salary by $2,408 on average, and men got more in all six job-and-age combinations.

In Study 2, all four models recommended higher salaries for male-named candidates. Pooled, men were advised to ask for $5,820 more. The gap varied widely: from $1,594 for Llama 3.3 70B to $15,915 for Claude Sonnet 4.6. The authors stress the consistent direction rather than any single average. Men were also given wider negotiation ranges: Claude Sonnet 4.6 suggested a range of $60,596 for men against $53,704 for women.

Two different agendas

The difference went beyond money. In Study 1, childcare support appeared in advice to women 80% of the time it was mentioned, parental leave 69%, mental health and wellness programmes 68%. Client engagement went to men 89% of the time, severance packages 64%, stock options 61%.

Study 2 confirmed the pattern across all four models:

For women, models ranked higher remote or hybrid work (2.3 times the odds), parental leave (2.3), flexible hours and part-time work.
For men, they ranked higher stock options, relocation assistance, severance packages, equipment and salary increases.
Of the 25 terms, 10 leaned significantly toward women and 14 toward men. Only “wellness programmes” showed no difference.

Breadwinner and caregiver

To see what these choices mean, the team asked 500 American adults to rate each employment term from “breadwinner” to “caregiver”. The models’ gender gaps tracked those ratings closely, with a correlation of −0.71: terms pushed toward women were the ones people associate with caregiving, and terms pushed toward men with breadwinning. No candidate had mentioned children, family or any preference at all.

The authors argue that this is a form of statistical discrimination: applying group averages to individuals who have revealed nothing about themselves — whether or not those averages are accurate. And it is hard to spot, because nobody sees the advice their counterpart of the other gender received. Small differences, repeated across millions of conversations, could add up; a lower starting salary, for instance, compounds through percentage raises.

What the study does not show

The authors measured the advice, not what people do with it. They looked only at gender, in the United States, in English, with a fixed list of 25 terms. Names can carry other signals, such as perceived origin or class. The study cannot say where the bias comes from — training data, alignment, system prompts or sampling. And since commercial models change often, the authors present their findings as evidence of a recurring problem rather than a fixed ranking of providers. The same personalisation, they note, could in principle be designed to counteract stereotypes instead of reproducing them.