COMPAS

Gold definitionUpdated Apr 2, 2026

COMPAS, an acronym for Correctional Offender Management Profiling for Alternative Sanctions, refers to a proprietary algorithm and, more commonly in AI research, the associated dataset used for assessing recidivism risk in criminal justice. Developed by Northpointe (now Equivant), the algorithm predicts the likelihood of a defendant re-offending. In the context of AI fairness, the COMPAS dataset has become a critical benchmark. It contains demographic information, criminal history, and COMPAS scores for defendants, enabling researchers to analyze and identify attribute-specific biases in predictive models. Its significance stems from its role in highlighting and quantifying algorithmic bias in high-stakes decision-making, prompting extensive research into fair AI systems. Researchers and ML engineers in areas like algorithmic fairness, explainable AI, and responsible AI frequently use this dataset to validate new fairness assessment tools and develop bias mitigation strategies.

The COMPAS Dataset as a Fairness Benchmark

Origin and Purpose of COMPAS: The COMPAS algorithm, developed by Northpointe, assesses a defendant's likelihood of re-offending. The associated dataset, containing demographic and criminal history data alongside COMPAS scores, has become a standard for evaluating fairness in predictive models within the criminal justice system.
Validation of Fairness Tools with COMPAS: The COMPAS dataset is frequently employed for experimental validation of new fairness evaluation tools, such as Nishpaksh. This demonstrates the effectiveness of such methods in identifying attribute-specific bias within AI systems, as noted in recent research. [2601.16926v1]

At a glance

Executive summary

COMPAS refers to a dataset derived from a risk assessment tool used in the criminal justice system, which has become a crucial benchmark for studying algorithmic bias. Researchers use it to test new methods for detecting and addressing unfairness in AI models, especially in critical decision-making contexts.

TL;DR

COMPAS is a dataset from a criminal risk assessment tool that AI researchers use to find and fix biases in algorithms.

Key points

A dataset derived from a proprietary recidivism risk assessment tool used in the criminal justice system.
Serves as a benchmark to identify and quantify algorithmic bias in high-stakes predictive systems.
Used by AI fairness researchers, ML engineers, ethicists, and legal scholars to validate fairness tools.
Unlike synthetic datasets, COMPAS is based on real-world data, making its biases more representative and impactful for fairness research.
Continues to be a foundational dataset for developing and validating new fairness metrics, bias mitigation techniques, and transparent AI frameworks.

Use cases

Validating Fairness Tools: Testing new fairness assessment frameworks, like Nishpaksh, to ensure they can effectively identify attribute-specific biases in predictive models.
Developing Bias Mitigation Strategies: Training and evaluating algorithms designed to reduce or eliminate racial, gender, or other demographic biases in risk prediction.
Educational and Policy Research: Used in academic courses and policy discussions to illustrate the real-world implications of algorithmic bias and the challenges of achieving fairness.
Benchmarking Fair ML Algorithms: Comparing the performance of different fair machine learning algorithms on a widely recognized, real-world dataset.

Also known as

Correctional Offender Management Profiling for Alternative Sanctions, COMPAS dataset

COMPAS

The COMPAS Dataset as a Fairness Benchmark

At a glance

Executive summary

TL;DR

Key points

Use cases

Also known as

Related topics

Key Controversies and Research on COMPAS

Implications of COMPAS for Responsible AI

Sources