AI in Employment Decisions: When the Algorithm Discriminates
An employer uses an automated system to screen resumes, score candidates, evaluate performance, or flag employees for termination. The system is presented as objective — free from the biases that affect human decision-making. But algorithms are built on data, and data reflects the world that produced it. When the algorithm discriminates, the employer is responsible.
The promise of artificial intelligence in employment decisions is neutrality. A machine does not have biases. It does not prefer one race over another, one gender over another, one age group over another. It processes data and produces outputs. The outputs are objective because the inputs are mathematical.
The promise is wrong. AI systems learn from historical data, and historical data reflects historical discrimination. A resume-screening algorithm trained on a decade of hiring decisions at a company that disproportionately hired young white men will learn to prefer resumes that look like those of young white men. A performance evaluation algorithm trained on manager ratings that reflect racial or gender bias will reproduce that bias at scale. The machine does not intend to discriminate. It does not need to. It inherits the discrimination embedded in the data, and it applies it with consistency and speed that no human decision-maker could match.
California employment law does not exempt algorithmic discrimination from its reach. An employer that uses an AI system to make employment decisions is responsible for the outcomes those systems produce — and if those outcomes constitute discrimination, the employer is liable.
The legal framework
FEHA prohibits discrimination in employment on the basis of protected characteristics. The prohibition is not limited to intentional discrimination by human actors. It extends to any employment practice — including the use of automated decision-making systems — that results in discrimination, whether through disparate treatment or disparate impact.
Disparate treatment occurs when the algorithm uses a protected characteristic as a factor in the decision — directly (the system considers the applicant’s age) or indirectly (the system uses proxies that correlate with protected characteristics, such as graduation year as a proxy for age, or zip code as a proxy for race). Disparate treatment through an algorithm is no less unlawful than disparate treatment by a human supervisor.
Disparate impact occurs when a facially neutral algorithm produces outcomes that disproportionately disadvantage members of a protected class, and the employer cannot demonstrate that the practice is justified by business necessity. A resume-screening tool that filters out a disproportionate number of applicants over fifty — not because it considers age, but because its criteria correlate with age — may produce a disparate impact that violates FEHA, even if the employer did not intend the result.
The employer cannot escape liability by pointing to the algorithm. “The AI made the decision” is not a defense. The employer chose to use the system, chose the data on which it was trained, and chose to rely on its outputs. The employer is responsible for the outcomes, just as it would be responsible for the decisions of a human manager it hired and empowered to act.
Where AI is being used
AI and automated systems are increasingly deployed at every stage of the employment relationship.
Hiring. Resume-screening tools that filter applications based on keywords, qualifications, or pattern-matching. Video interview platforms that evaluate candidates’ facial expressions, tone of voice, and word choice. Chatbots that conduct initial screening conversations and advance or reject candidates based on their responses. Assessment tools that score candidates on cognitive ability, personality traits, or “cultural fit.”
Performance management. Algorithms that score employees based on productivity metrics, customer ratings, or peer evaluations. Systems that identify “low performers” for termination or “high potentials” for promotion based on data patterns. Automated monitoring tools that track keystrokes, mouse movements, email volume, and other proxies for productivity.
Termination. Algorithms that select employees for layoff based on performance scores, efficiency metrics, or cost models. Systems that flag employees for disciplinary action based on attendance patterns, policy compliance, or behavioral indicators.
In each of these applications, the risk of discriminatory impact is real. The system’s outputs are only as fair as the data and the design — and when neither has been audited for bias, the outputs may systematically disadvantage employees and applicants in protected classes.
California’s regulatory direction
California has been among the most active states in addressing AI in employment. While the regulatory landscape continues to develop, several legal frameworks are already relevant.
FEHA’s existing framework applies to AI-driven employment decisions without modification. The anti-discrimination provisions reach any employment practice that produces discriminatory outcomes, regardless of the mechanism. An employer that uses an AI tool that disproportionately screens out disabled applicants, or that rates women lower than men on performance metrics, or that selects older workers for layoff at higher rates has violated FEHA — whether or not the employer intended the result and whether or not the employer understood how the algorithm worked.
The disparate impact framework is particularly well-suited to algorithmic discrimination claims. The employee establishes a prima facie case by showing a statistically significant disparity in outcomes between protected and non-protected groups. The employer must then demonstrate that the practice is justified by business necessity. Even if the employer meets that burden, the employee can prevail by showing that an alternative practice with less discriminatory impact would serve the employer’s needs.
Transparency and notice. California has considered legislation requiring employers to disclose their use of AI in employment decisions and to provide affected individuals with notice and an opportunity to challenge adverse determinations. While the specific legislative landscape may continue to evolve, the direction is toward greater accountability and transparency.
What employees can do
An employee or applicant who believes they have been harmed by an AI-driven employment decision should consider several factors.
The outcome. Were you screened out of a hiring process, rated poorly on a performance metric, selected for layoff, or denied a promotion? And does the outcome correlate with a protected characteristic — your age, race, gender, disability?
The process. Was the decision made or influenced by an automated system? Were you evaluated by an algorithm, scored by a tool, or filtered by a screening platform? If so, what data did the system use, and how was it designed?
The pattern. Do the outcomes affect a disproportionate number of people in your protected class? If the company’s AI-driven hiring process consistently screens out applicants over fifty, or if its performance algorithm consistently rates women lower than men, the pattern is evidence of disparate impact.
Access to the algorithm’s design and training data is typically available only through discovery in litigation. But the employee need not understand the algorithm’s inner workings to establish a claim. The statistical disparity in outcomes — combined with the employer’s use of an automated system — is sufficient to state a case and to trigger the employer’s obligation to justify the practice.
If an employer’s use of AI or automated decision-making has produced outcomes that disadvantage you because of your age, race, gender, disability, or other protected characteristic, the technology does not shield the employer from liability. California law holds employers responsible for the employment decisions they make — regardless of whether a human or a machine made them.
This article is general information about California law and is not legal advice. Every situation turns on its particular facts.