Implicit Bias in the Justice Delivery System

Assessments of credibility. A judge or a magistrate decides whom to believe, and confidence, manner, language and accent affect that judgment.

Sentencing, where discretion is widest and where perceptions of remorse, of dangerousness and of prospects of reform are made.

Remission, parole and prison classification.

Why it matters more than individual prejudice

Two reasons, and both are worth marks.

It survives good intentions. A system can be staffed entirely by people who genuinely intend to be fair and still produce unequal outcomes, because the bias operates on the perception and not on the intention. The usual remedies for prejudice, exhortation and codes of conduct, do not reach it.

It compounds. Each decision point produces a small effect, and the points are sequential. A slightly higher probability of being stopped, multiplied by a slightly higher probability of arrest rather than warning, of bail being opposed, of being unrepresented, and of a longer sentence, produces an outcome at the end far more unequal than any single step.

How it is measured, and the honest limits

An answer that reports the measurement critically is a much better answer than one that treats the science as settled.

Implicit association tests measure how quickly a person pairs categories with attributes, on the reasoning that faster pairing reveals a stronger association.

Audit and correspondence studies send otherwise identical applications or complaints differing only in a marker of group identity and compare the responses. These are the most persuasive design because everything else is held constant.

Outcome studies compare decisions across groups while controlling for legally relevant factors, and attribute the residue to something other than the facts.

The limits, stated plainly. The link between scores on association tests and actual behaviour is contested and appears to be weak at the individual level. Outcome studies can never control for everything, so a residue is evidence and not proof. And a finding about a population does not tell you anything about a particular decision. What survives all of this is the audit design, which is why it carries most of the weight.

What can be done about it

Reduce discretion where it is not needed. Structured decisions, checklists and stated criteria narrow the space in which automatic associations operate.

Require reasons. A decision that must be explained is a decision the maker has to examine, which is the practical value of the requirement in Sunil Batra that reasons for bar fetters be recorded in the prisoner's history ticket.

Blind the decision where possible. Removing identifying information at the point of decision is the most effective single intervention where it can be done.

Change the conditions. Bias operates hardest under time pressure and cognitive load, so reducing caseloads and delay is a bias measure and not only an efficiency measure.

Collect and publish disaggregated data. Nothing can be corrected that is not measured, and the absence of data is itself a policy choice.

Diversify the decision makers, which changes both the associations in the room and the willingness of others to challenge a judgment.

Training, with a caution. Awareness training alone has a poor record and can produce complacency. It is worth doing as part of a package and is not a substitute for structural change.

A worked example

Two men are arrested on the same night in the same city for the same offence, an assault outside a bar. Both are twenty six, both were drinking, neither has a previous conviction, and the injuries are identical.

The first speaks the language of the officers, is dressed in office clothes, gives an address in a housing society, and produces a colleague who vouches for him.

The second speaks a different first language, works on a construction site, gives an address at a labour camp, and has nobody to call at two in the morning.

Trace the sequence. The first is more likely to be released on station bail or to have his arrest treated as a formality. The second is more likely to be produced before a magistrate. At the bail hearing the first has an address that is verifiable, employment that is documented and a surety; the second has none of these, and every one of the criteria applied at that hearing correlates with social position rather than with the risk the criteria are meant to measure. The first engages a lawyer of his choice; the second waits for legal aid. The first is at liberty and can gather evidence; the second cannot. At trial the first appears from the body of the court, composed, and the second from custody.

Notice what has not happened. No officer, magistrate or judge has held any prejudice or done anything improper. Every decision is defensible in isolation, and each was made on a criterion that looks neutral. The inequality was produced by the accumulation, and by criteria that are proxies for social position.

And notice the remedy this suggests. It is not to exhort the officials. It is to change the criteria: to accept alternative proofs of identity and residence, to fix bail conditions that do not require a surety with property, to provide representation at the first hearing rather than later, and to publish the figures on who is released and who is not.

Distinctions that carry marks

What implicit bias is NOT

It is not an accusation of prejudice. The entire point of the concept is that it operates in people who are not prejudiced, and treating it as an accusation guarantees that the discussion will go nowhere.

It is not an excuse. That a bias was unconscious does not make the outcome acceptable; it changes the remedy from blame to redesign.

It is not proved by an unequal outcome alone. An unequal outcome is consistent with bias and with other explanations, and the work of an argument is to eliminate the others.

It is not the whole of the problem. Chapter 530 deals with the part that is structural and needs no bias at all: money, delay, representation, and who can afford to defend themselves.

Quick revision

1. Implicit bias is an unconscious association operating on perception, memory and judgment; explicit bias is consciously held and reportable.
2. Sincerity is neither a defence nor evidence, which is why asking officials whether they are biased tests only explicit bias.
3. It operates hardest where information is thin, time is short and discretion is wide, which describes most criminal justice decisions.
4. Decision points: whether an incident is treated as a crime, registration, stop and search, arrest, identification, bail, representation, charge, credibility, sentencing, remission and parole.
5. Two reasons it matters: it survives good intentions, and it compounds across sequential decisions.
6. Measurement: association tests (contested), audit studies (strongest), outcome studies (evidence, not proof).
7. Remedies: reduce unnecessary discretion, require reasons, blind the decision, reduce time pressure, publish disaggregated data, diversify decision makers, and treat training as a supplement rather than a solution.

Test yourself

1. Define implicit bias and distinguish it from explicit bias. Implicit bias is an unconscious association between a social group and an attribute, acquired from the surrounding culture, which operates automatically on perception, memory and judgment. Explicit bias is a belief or attitude the person holds consciously and can report. The crucial difference is awareness: a person affected by implicit bias would sincerely deny holding it and may consciously believe the opposite, which is why sincerity is neither a defence nor evidence, and why asking officials whether they are biased tests only the explicit kind.

2. Where can implicit bias enter a criminal process? At every point where discretion exists: whether an incident is treated as a crime at all; whether a complaint is registered; who is stopped, questioned and searched; whether a person is arrested or released and how force is used; how witnesses perceive and identify; whether bail is opposed and granted; the quality of representation; which offence is charged; assessments of credibility; sentencing, where discretion is widest; and remission, parole and prison classification.

3. Why does implicit bias matter more than individual prejudice? Because it survives good intentions, so a system staffed entirely by people who genuinely mean to be fair can still produce unequal outcomes, and the usual remedies for prejudice, exhortation and codes of conduct, do not reach it. And because it compounds: each decision point contributes a small effect and the points are sequential, so slightly higher probabilities of being stopped, arrested, refused bail, left unrepresented and sentenced longer multiply into an outcome far more unequal than any single step would suggest.

4. How is implicit bias measured, and what are the limits of the evidence? By implicit association tests, which time how quickly categories are paired with attributes; by audit or correspondence studies, which send otherwise identical applications differing only in a marker of group identity; and by outcome studies, which compare decisions across groups while controlling for legally relevant factors. The limits are real: the link between association test scores and individual behaviour is contested and weak, outcome studies cannot control for everything so a residue is evidence rather than proof, and population findings say nothing about a particular decision. The audit design carries most of the weight because everything else is held constant.

5. What can be done about it? Reduce discretion where it is not needed by structuring decisions and stating criteria. Require reasons, since a decision that must be explained is one the maker has to examine. Blind the decision to identifying information wherever that is possible, which is the single most effective intervention. Reduce time pressure and caseload, since bias operates hardest under load. Collect and publish disaggregated data, because nothing unmeasured can be corrected. Diversify decision makers. And treat awareness training as a supplement, since alone it has a poor record and can produce complacency.