AI-Generated Performance Reviews & Bias Risks

Yes. Under Title VII of the Civil Rights Act of 1964 and established agency principles, employers are strictly liable for the discriminatory tangible employment actions taken by supervisors. If a manager uses a commercial LLM to generate narrative reviews, and the model incorporates gender-coded, racially disparate, or age-biased evaluations that impact ratings, bonuses, or promotions, the employer cannot disclaim responsibility by blaming third-party generative AI software.

In the landmark Supreme Court decision Price Waterhouse v. Hopkins (490 U.S. 228), evaluating female employees negatively for perceived interpersonal traits (e.g., being 'too abrasive,' 'lacking warmth,' or 'needing a course at charm school') while praising male employees for identical assertiveness constitutes unlawful sex stereotyping. Large language models (LLMs) trained on historical corporate datasets systematically mirror this bias, generating feedback that describes women with emotional personality descriptors and men with objective technical competencies.

Coded language refers to facially neutral adjectives and phrasings that reflect underlying demographic stereotypes. For female employees, LLM prompts often output critiques regarding 'tone,' 'approachability,' 'emotional maturity,' or 'collaborative warmth.' For older workers, models frequently introduce tropes regarding 'adaptability to change,' 'technological enthusiasm,' or 'speed of execution.' In employment discrimination litigation, statistical patterns of coded language serve as potent circumstantial evidence of disparate treatment.

Colorado SB 24-205 classifies artificial intelligence systems used to evaluate employee performance, promotions, or compensation as 'high-risk AI systems.' Deployers must: (1) complete annual algorithmic impact assessments to detect potential algorithmic discrimination; (2) provide advance written disclosure to employees detailing the AI system's purpose and inputs; (3) maintain an active risk management policy; and (4) notify the Colorado Attorney General within 90 days if algorithmic discrimination is uncovered.

Yes. Pasting unredacted employee performance notes, peer reviews, medical leave absences, or disciplinary records into public consumer generative AI tools (such as free-tier ChatGPT) violates workplace privacy policies, state data protection statutes (e.g., CCPA/CPRA employee privacy mandates), and confidentiality covenants. Many commercial LLMs utilize user prompt inputs to train public foundation models, risking public disclosure of sensitive personnel data.

Employers should conduct natural language processing (NLP) text audits across all written review narratives to analyze word frequency distributions across gender, racial, and age cohorts. Calculating sentiment polarity scores, tracking the prevalence of communal vs. agentic language, and cross-referencing narrative scores with objective quantitative performance metrics (e.g., sales quota attainment or code commit volume) will uncover hidden rating discrepancies.

A compliant human-in-the-loop workflow requires that generative AI is restricted to grammar editing or summarizing manager-provided bullet points, rather than generating substantive ratings or critiques from scratch. People Ops must mandate that managers document the specific objective work outputs supporting every rating and subject all AI-assisted reviews to an independent calibration committee review prior to delivery to the employee.

Yes. The EEOC's Artificial Intelligence and Algorithmic Fairness Initiative actively accepts and investigates charges where automated tools or generative models contributed to adverse employment outcomes—including denied bonuses, lower merit increases, PIP placements, or demotions. If the review text reflects stereotypical tropes, the EEOC can issue a cause finding under Title VII or the ADEA.

Employers must institute a formal Workplace Generative AI Policy that: (1) strictly prohibits pasting identifiable employee data into public, non-enterprise LLMs; (2) forbids using AI to generate substantive performance ratings or developmental criticisms; (3) requires explicit managerial disclosure if enterprise AI was used to format or polish manager-written notes; and (4) mandates objective, behavioral calibration reviews.

Under Desert Palace, Inc. v. Costa and 42 U.S.C. § 2000e-2(m), an employee only needs to demonstrate by direct or circumstantial evidence that a protected characteristic (such as sex or race) was a 'motivating factor' for an adverse employment decision, even if other legitimate performance factors existed. If AI-generated narrative feedback contains gender stereotypes that influenced the manager's final score, the employer faces Title VII liability and fee shifting.