Y12W38VC Machines making decisions about people

Relating to or produced by an algorithm; following a prescribed step-by-step computational procedure or set of rules. Algorithmic systems make decisions by applying predetermined logical operations.

Word Breakdown: algorithm + -ic (relating to)

Word family: algorithm (n.), algorithmically (adv.)

Synonyms: computational, procedural, systematic

Collocations: algorithmic bias, algorithmic fairness, algorithmic decision

Example: Algorithmic systems in criminal justice make predictions about defendants based on historical data and patterns.

Out of proper proportion; in a way that is not fair, balanced, or equal. Disproportionate effects occur when harm or benefit falls unequally on groups.

Word Breakdown: dis- (not) + proportionately (fairly)

Word family: disproportionate (n.)

Synonyms: unequally, unfairly, unbalancedly

Collocations: disproportionately affect, disproportionately impact, disproportionate error rates

Example: The algorithm's errors fell disproportionately on certain racial groups, meaning false predictions were more common for some groups than others.

Impartial and just treatment; equity in decision-making and distribution of outcomes. In algorithmic contexts, fairness means equal treatment or equal error rates across groups.

Word Breakdown: fair + -ness (quality of)

Word family: fair (n.), fairly (adv.)

Synonyms: equity, justice, impartiality

Collocations: algorithmic fairness, fairness criterion, fairness across groups

Example: Algorithmic fairness requires that error rates be equivalent across racial and demographic groups, not just that the algorithm seems impartial.

A systematic distortion or prejudice in favor of or against a particular group, outcome, or interpretation. Bias can be intentional or unintentional, and it skews decisions or data analysis.

Word family: biased (n.), biasing (n.)

Synonyms: prejudice, distortion, skew

Collocations: algorithmic bias, racial bias, gender bias

Example: The algorithm exhibited bias because it made systematically different error rates for different racial groups.

Adjusted or tuned to precise standards; in algorithmic contexts, calibrated means that predictions are accurate across different groups. Calibration ensures that a score of '60' means the same probability regardless of group membership.

Word Breakdown: calibr- (adjust) + -ated (made to be)

Word family: calibrate (n.), calibration (n.)

Synonyms: adjusted, tuned, precise

Collocations: calibrated to accuracy, calibrated algorithm, properly calibrated

Example: A calibrated algorithm would predict the same actual recidivism probability for all groups receiving the same score.

Operated or controlled by machine without human intervention; performed automatically according to a program or system. Automated systems replace human decision-making with algorithmic procedures.

Word Breakdown: auto- (self) + mated (made)

Word family: automate (n.), automation (n.)

Synonyms: mechanical, programmed, automatic

Collocations: automated decision, automated system, automated bias

Example: Automated systems in criminal justice raise concerns because once implemented, they apply decisions consistently without human discretion.

Open to inspection and understanding; allowing visibility of how something works. Transparent algorithms reveal their logic and reasoning, not operating as black boxes.

Word Breakdown: trans- (through) + parent (visible)

Word family: transparency (n.), transparently (adv.)

Synonyms: clear, open, understandable

Collocations: algorithmic transparency, transparent system, lack of transparency

Example: A transparent algorithm would explain which factors influenced its prediction and why.

A statement or account of how or why something happened or works. In algorithmic contexts, explanations provide understanding of decision logic and the factors influencing outcomes.

Word Breakdown: ex- (out) + -planation (flattening/making clear)

Word family: explain (n.), explanatory (n.)

Synonyms: account, reasoning, justification

Collocations: explanation for bias, lack of explanation, provide explanation

Example: Without explanation, defendants cannot challenge algorithmic predictions or understand why they received a particular score.

The field of study focused on how to design and evaluate automated decision systems that treat groups equitably. Algorithmic fairness research grapples with competing definitions of what makes an algorithm fair.

Synonyms: fair algorithms, equitable automation

Collocations: algorithmic fairness research, algorithmic fairness criterion, achieve algorithmic fairness

Example: Algorithmic fairness researchers debate whether fairness means equal error rates or equal accuracy predictions across groups.

A controversial risk-assessment algorithm used in American courts to predict defendants' likelihood of future criminal behaviour. Journalists revealed that COMPAS made different types of errors for different racial groups.

Synonyms: risk assessment algorithm, recidivism prediction tool

Collocations: COMPAS algorithm, COMPAS bias, COMPAS scores

Example: COMPAS became the subject of intense scrutiny when research revealed its differential error rates across racial groups.

A fairness criterion requiring that error rates be equal across groups; if an algorithm has a 10% false positive rate for one group, it should have a 10% false positive rate for all groups. Equalised odds ensure that algorithmic mistakes affect all groups equally.

Synonyms: equal error rates, equitable error distribution

Collocations: equalised odds criterion, achieve equalised odds

Example: Equalised odds is one definition of fairness: an algorithm is fair if it makes mistakes at the same rate for all racial groups.

A fairness criterion requiring that a prediction score means the same probability of the outcome across all groups; a score of '60' indicates a 60% probability of recidivism regardless of the defendant's race. Calibration ensures predictive accuracy but may produce different error rates.

Word family: calibrated, calibrate

Synonyms: prediction accuracy, probability alignment

Collocations: calibration criterion, achieve calibration

Example: Calibration requires that a risk score accurately predict the actual probability of reoffending, but may result in unequal error rates across groups.

A mathematical result showing that different fairness criteria cannot all be satisfied simultaneously. If an algorithm achieves equalised odds (equal error rates), it cannot achieve calibration (equal prediction accuracy), and vice versa.

Synonyms: fairness trade-off, incompatible fairness criteria

Collocations: impossibility theorem shows, fairness impossibility theorem

Example: The impossibility theorem reveals a fundamental tension: no single algorithm can satisfy all competing definitions of fairness at once.