Suppose a bank’s credit model approves 60% of applicants from one neighborhood and 35% from another. Is that unfair? It depends on why. If the gap reflects actual repayment ability, some would call it fair; if it reflects historical lending discrimination baked into the training data, or a feature that stands in for race, most would not. Algorithmic fairness is the work of measuring gaps like this, deciding which ones are unjustified, and changing the data, model or decision process to reduce them.
The field has many formal definitions, and they pull in different directions. Demographic parity asks for equal approval rates across groups. Equal opportunity asks that qualified people in each group be wrongly rejected at the same rate; equalized odds also requires that unqualified people be wrongly approved at the same rate. Calibration asks that a given score mean the same risk in every group. These can conflict: Kleinberg and colleagues proved that, outside narrow special cases, a risk score cannot be calibrated within each group and also balanced in the average scores it gives to people who do and don’t turn out positive in each group. Choosing a fairness metric is therefore a value judgment about which mistakes matter most, not a purely technical step.
That is why toolkits such as Fairlearn and standards such as NIST’s bias guidance treat fairness as sociotechnical: it depends on who is affected, what the decision is, and what the law requires, not only on model statistics. In practice, teams measure performance separately for each group, look for selection bias and proxy features in the data, involve the people affected, and keep a human able to challenge the result. It sits within responsible AI and AI governance, and laws such as the EU AI Act treat hiring and credit as high-risk uses.
In episode 58 at 41:09, Daniel Chait argues that AI interviews could help hiring if built on “a solid structured hiring process with good science and good thinking behind fairness.” That is his view on what AI hiring needs, not a measured result. Galileo sponsored the episode.
Hear it from the guest
“Good science and good thinking behind fairness”
Quotes lightly edited to remove filler words.
Sources
- Barocas, Hardt and Narayanan: Fairness and Machine Learning: Limitations and Opportunities — Free textbook covering classification fairness criteria, their relationships and limitations, relative notions of fairness and causality; checked October 24, 2026.
- Kleinberg, Mullainathan and Raghavan: Inherent Trade-Offs in the Fair Determination of Risk Scores (2016) — Proves that, except in highly constrained cases, no method can satisfy three common fairness conditions at the same time; checked October 24, 2026.
- Fairlearn user guide: Fairness in machine learning — Open-source toolkit documentation that treats fairness as sociotechnical and describes measurements such as demographic parity; checked October 24, 2026.
- NIST SP 1270: Towards a Standard for Identifying and Managing Bias in Artificial Intelligence (2022) — NIST's guidance on statistical, human and systemic bias in AI and approaches to managing it; checked October 24, 2026.
