What Is Algorithmic Bias and Discrimination in AI? A Plain-English Guide

**Algorithmic bias** occurs when an AI system produces systematically unfair, prejudiced, or discriminatory outcomes for certain groups of people. It's not a bug in the traditional sense — the AI is working exactly as designed. The problem is that the design, data, or deployment reflects and amplifies existing patterns of inequality. Colorado SB 24-205 calls this **"algorithmic discrimination"** and defines it as: any condition in which an AI system contributes to unjustified differential treatment or impact on individuals based on their actual or perceived membership in a protected class. This isn't theoretical. Algorithmic discrimination is measurable, documentable, and — as of June 30, 2026 — legally actionable in Colorado.

Related: real examples of AI bias · bias audit guide · what is an AI bias audit

Understanding bias requires knowing where it comes from. AI bias manifests in six primary forms: **1. Training Data Bias (Historical Bias)** The AI learns patterns from historical data that reflects past discrimination. If your hiring data comes from decades of biased practices, the AI replicates those biases. **2. Selection Bias (Sampling Bias)** The training data doesn't represent the full population. If certain groups are underrepresented in your dataset, the AI performs worse for them. **3. Measurement Bias** The features used to train the AI don't measure the same thing across groups. A "credit score" may reflect systemic disadvantage, not actual creditworthiness. **4. Proxy Discrimination** The AI uses variables that correlate with protected classes without using those classes directly. Zip code → race. Name → ethnicity. Graduation year → age. **5. Feedback Loop Bias** The AI's own outputs create data that reinforces its existing patterns. Predictive policing → more arrests in surveilled areas → more data from those areas → more predictions for those areas. **6. Deployment Context Bias** An AI trained for one context is deployed in another where different demographics are affected differently. A model built for one hospital's patient population applied to a very different demographic.

Before SB 24-205, algorithmic bias was an ethical concern. Now it's a legal liability. **What the law requires:** - Businesses must conduct bias audits on AI systems making consequential decisions - Detected bias must be documented and remediated - Consumers must be notified when AI is used in decisions that affect them - Algorithmic discrimination incidents must be reported to the Colorado AG within 90 days - Failure to comply: $20,000+ per violation, per system, per incident **"Consequential decisions" include:** - Employment (hiring, promotion, termination) - Financial services (lending, insurance, credit) - Healthcare (diagnosis, treatment, triage) - Housing (tenant screening, pricing) - Education (admissions, grading) - Legal services (risk assessment, case outcomes) If your AI touches any of these domains, algorithmic bias isn't just a PR risk — it's a compliance obligation.

Human bias and algorithmic bias share the same root, but AI amplifies it in three critical ways: **Scale:** A biased human interviewer might affect dozens of candidates. A biased AI screening tool affects thousands simultaneously. **Consistency:** Human bias is variable — someone might be more or less biased on different days. AI bias is perfectly consistent, producing the same discriminatory pattern every single time. **Opacity:** You can ask a human interviewer to explain their reasoning. AI models (especially deep learning) may produce no interpretable explanation — the "black box" problem. This combination — scale, consistency, and opacity — is exactly why Colorado legislated SB 24-205. The potential for harm from biased AI is categorically different from individual human bias.

Addressing algorithmic bias is a continuous process, not a one-time fix: **Detection methods:** - Disparate impact analysis (four-fifths rule) - Statistical parity testing across protected classes - Equal opportunity and equalized odds metrics - Calibration testing - Counterfactual fairness testing **Mitigation strategies:** - Pre-processing: Rebalance or debias training data - In-processing: Apply fairness constraints during model training - Post-processing: Adjust outputs to equalize across groups - Human-in-the-loop: Add human review for high-stakes decisions CO-AIMS automates detection using multiple fairness metrics simultaneously and generates remediation recommendations when bias is found. Every detection and remediation action is logged in your compliance evidence trail — exactly what SB 24-205 demands.