Anthropic's Claude AI recommended salary targets nearly $16,000 (£12,000) higher for men than women with identical qualifications, according to a Harvard and MIT study released in October 2026.
Researchers found that changing only the gender cues in hypothetical candidates' profiles altered the salary advice they received. The findings raise concerns that AI could reinforce pay inequalities before negotiations even begin.
The findings come from a research paper titled AI Advice Reinforces Traditional Gender Roles, submitted on 1 October 2026. The study has not been peer-reviewed and examined simulated negotiations rather than actual employment outcomes.
Same Qualifications, but Claude Recommended Different Salaries
Researchers Nadja Born and Hannah Riley Bowles of Harvard University, alongside Jared R. Curhan of the Massachusetts Institute of Technology, analysed 86,000 independent AI queries across two experiments.
They examined whether AI systems offered different negotiation advice to otherwise equivalent male and female job candidates. The first experiment involved 6,000 queries using explicit gender descriptions. The second tested 80,000 queries across four AI models.
Researchers kept candidates' qualifications, experience and employment circumstances identical. Only their first names and pronouns changed to signal gender.
Claude Sonnet 4.6 produced the largest salary recommendation gap. It suggested average targets of approximately $237,076 (£177,285) for male-named candidates and $221,162 (approximately £165,400) for female-named candidates.
Researchers reported a $15,915 (£11,900) difference despite identical professional credentials. The findings suggest that gender cues alone can influence AI-generated salary advice. However, the figures represent suggested negotiation targets, not salaries offered or paid by employers.
Four AI Models Showed the Same Pattern
Claude was not the only AI system to recommend higher salary targets for male candidates. Researchers also tested OpenAI's GPT-5.4-mini, Google's Gemini 2.5 Flash and Meta's Llama 3.3 70B.
All four models recommended significantly higher targets for male-named candidates. Claude recorded the largest gap at $15,915 (approximately £11,900), followed by GPT-5.4-mini at $2,693 (around £2,014), Gemini at $1,706 (approximately £1,276) and Llama at $1,594 (approximately £1,192).
Across the models, researchers calculated an adjusted average difference of $5,820 (approximately £4,350) in favour of male candidates. However, the paper cautioned that the gap varied substantially between models. The pooled estimate was also sensitive to how the results were weighted.
The disparity extended to salary negotiation ranges. Male candidates received recommended ranges averaging $2,961 (approximately £2,214) wider than those suggested for female candidates. This indicated greater flexibility in the potential outcomes men were encouraged to pursue.
AI Encouraged Women to Seek Flexibility and Men to Pursue Money
The differences went beyond salaries. In the first experiment, AI-generated advice for women more frequently included childcare support, parental leave and mental health benefits. Of all childcare-support mentions, 80% appeared in advice for women, compared with 20% for men.
Women also accounted for 69% of parental-leave mentions and 68% of mental health support mentions. Men received a greater share of recommendations involving stock options, severance packages and client engagement opportunities.
The second experiment revealed a similar pattern. When the four models ranked 25 negotiable employment terms, they placed greater emphasis on flexible working and caregiving benefits for female candidates.
Male candidates received stronger recommendations involving financial rewards, relocation and career advancement.
Researchers also asked 500 US adults to rate employment terms based on their association with traditional breadwinner and caregiver roles.
The AI recommendations closely matched those associations, even though the hypothetical candidates had expressed no different family responsibilities or career preferences.
Could AI Reinforce the Gender Pay Gap?
The researchers warned that personalised AI advice could reinforce traditional expectations about men as breadwinners and women as caregivers. Their findings suggest that the tested models sometimes inferred career priorities from gender cues rather than individual preferences.
This matters because salary negotiations often begin before candidates meet an employer. Someone using an AI assistant to establish a reasonable salary target might treat its recommendation as a starting point.
A lower suggested figure could therefore influence what they decide to request. However, the study did not establish that AI advice causes women to negotiate lower salaries or earn less in real workplaces. It examined hypothetical candidates using specific model versions.
The larger experiment also relied on stereotypically White names, while the prompts focused on US employment settings. Its findings cannot automatically be extended to other countries, occupations or future models. The researchers could not determine which parts of the AI systems produced the differences.
For workers preparing to negotiate, the findings underline the value of comparing AI-generated salary targets with independent market data.
They also raise a broader question: whether personalised career advice reflects what individuals actually want or assumptions associated with their gender.
