Large Language Models Systematically Recommend Lower Salaries for Women, Minorities, and Refugees in Negotiation Advice

Two independent studies have demonstrated that large language models systematically recommend lower salaries for women, ethnic minorities, and refugees when asked for salary negotiation advice.

A peer-reviewed study by Geiger et al. (PLOS ONE, February 2025) submitted 395,200 prompts to four ChatGPT versions, systematically varying gender, university, and major. All four models showed statistically significant bias advantaging men over women, with an average gender gap of approximately $1,060 — about 1% of recommended salary. Gender was the smallest disparity found; model version and prompt framing produced larger variation.

A separate study by Sorokovikova et al. (presented at the GeBNLP workshop at ACL 2025) tested five LLMs — GPT-4o Mini, Claude 3.5 Haiku, Llama 3.1 8B, Qwen 2.5 Plus, and Mixtral 8x22B — and found consistent gender-based salary bias across all models. In the most extreme example, ChatGPT's o3 model suggested a male medical specialist in Denver ask for $400,000 while suggesting $280,000 for an identical female persona — a $120,000 gap. Pay gaps were most pronounced in law and medicine. The biases were found to be compounding: individuals at the intersection of multiple marginalized identities (e.g., a female refugee from an ethnic minority) received the lowest salary recommendations.

The Ontario Human Rights Commission formally cited these findings in its submission to Canada's renewed AI Strategy consultations, as evidence that AI systems perpetuate and amplify existing patterns of discrimination. On January 21, 2026, the OHRC and the Information and Privacy Commissioner of Ontario jointly published AI principles addressing algorithmic discrimination in employment contexts.

Ontario's Working for Workers Five Act, which took effect January 1, 2026, requires employers to disclose AI use in hiring — but does not cover AI systems used by job seekers for salary negotiation advice, leaving this discrimination pathway unaddressed by current law.

The hazard is structural: as millions of workers increasingly rely on AI chatbots for career and salary advice, systematically biased recommendations create a pathway to widening pay gaps at population scale — even without any employer deploying a biased system.

The findings are based on two studies: one peer-reviewed (Geiger et al., ChatGPT only) and one workshop paper (Sorokovikova et al., multi-model). AI models are regularly updated, and bias patterns documented in one model version may not persist in subsequent versions. Some AI developers have implemented bias testing and mitigation procedures for their models. The OHRC's citation of the findings in policy consultations reflects the study's relevance to ongoing regulatory discussions, though broader replication would strengthen the evidence base.