Suppose 80% of homicides were committed by people with red hair

Following is the response from ChatGPT AI to the above question.

If we take a hypothetical where 80% of homicides are committed by a clearly identifiable subgroup, and membership in that subgroup is known before the homicide occurs, the key question is whether the characteristic is merely correlated with homicide or whether it identifies people who are individually at high risk of committing one.

If we use “red hair” to make the distinction clear: targeting people because they have red hair would not logically follow from the statistic. The useful intervention would have to address whatever causes the subgroup’s elevated homicide rate.

Several approaches could reduce the homicide rate:

1. Identify the causal risk factors within the subgroup.
   Determine why the subgroup has the elevated rate: previous violent offending, access to firearms, involvement in particular conflicts, substance abuse, victimization, socioeconomic conditions, etc.
2. Target behavior rather than the characteristic.
   If prior violent offending accounts for much of the concentration, interventions could focus on people with documented violent behavior rather than everyone possessing the identifying characteristic.
3. Intervene before violence occurs.
   Examples include credible-threat intervention, domestic-violence interventions, conflict mediation, intensive supervision of people with violent histories, and treatment for relevant substance-use problems.
4. Restrict access to means where evidence supports it.
   If firearms or another particular means account for a substantial fraction of the homicides, policies that prevent demonstrably high-risk individuals from accessing those means can potentially reduce homicide without treating the entire subgroup as dangerous.
5. Concentrate resources geographically or socially if violence is concentrated there.
   Focused policing, violence-interruption programs, environmental improvements, and services in high-violence locations can address the actual concentration without assuming that every member of the broader subgroup is responsible.
6. Use the subgroup statistically, but not as the sole criterion for individual action.
   A characteristic can be extremely useful for research and still be a poor basis for treating an individual as dangerous. The distinction is between: “80% of offenders belong to X” and “A person belonging to X has a sufficiently high probability of committing homicide that intervention against that person is justified.”

Those are very different propositions.

There is also an important mathematical point. Suppose subgroup X is 10% of the population but commits 80% of homicides. X would have a homicide-offending rate roughly 36 times that of the other 90%:

$$\frac{0.80/0.10}{0.20/0.90}=36$$

That would be an enormous concentration and would warrant serious investigation into why it exists. But even then, most people in X would presumably not commit homicide. That matters enormously when deciding what interventions are proportionate.

So the strongest general principle is:

Use the concentration of homicide to identify where to investigate and concentrate prevention, but use individual evidence of risk or behavior to decide whom to intervene against.

We can also take this hypothetical one step further and examine what changes if the identifying characteristic is immutable and genuinely predicts homicide at a very high rate. That gets into the difficult distinction between statistical discrimination, individual prediction, and effective crime-prevention policy.

The harder case is where the characteristic is immutable and genuinely predictive of homicide—not merely correlated through some third factor.

Suppose, hypothetically:

• X is 10% of the population.
• X commits 80% of homicides.
• The characteristic is known with certainty.
• The disparity persists after controlling for other measurable factors.
• There is no obvious causal mechanism that can simply be changed.

That creates a genuine policy problem, because statistical information can be highly predictive while still being poor justification for treating every individual member of X as dangerous.

1. Start with the base rates

Even the dramatic 80% figure can conceal an important fact.

If there are 100 homicides in a population of one million:

• 80 are committed by X’s 100,000 members.
• 20 are committed by the other 900,000.

The annual homicide-offending rate would be:

• X: 80 per 100,000
• non-X: 2.2 per 100,000

So X is about 36 times more likely to produce a homicide offender.

But simultaneously:

99.92% of X does not commit a homicide in that year.

That is the fundamental problem with using group membership alone.

2. Prediction and prevention aren’t the same thing

Imagine that we had a perfect statistical model saying:

“People in X are 36 times as likely to commit homicide.”

That doesn’t tell us what to do to an individual member of X.

There are at least three possible strategies:

Strategy	Potential benefit	Problem
Treat everyone in X as dangerous	Captures most offenders	Enormous number of innocent people affected
Ignore X entirely	Avoids discrimination	Ignores useful predictive information
Use X plus individual risk factors	Concentrates intervention	Requires additional information and safeguards

The third approach is generally much more informative because the objective isn’t to identify a group that contains offenders; it’s to identify individuals who are likely to become offenders.

3. The crucial question is the incremental predictive value

Suppose X initially predicts homicide extremely well.

Now add:

• previous violent offenses
• threats against specific people
• domestic violence
• weapon access
• involvement in violent conflicts
• age
• recent victimization
• credible reports of planned violence

Perhaps the resulting model identifies a much smaller population containing a large proportion of future offenders.

Then X may still be statistically relevant, but it doesn’t need to be the basis for intervention.

For example:

X + documented violent behavior + credible threat

is a very different proposition from:

X

The first identifies a particular risk profile. The second identifies a population.

4. There is another possibility: X is merely a proxy

This is perhaps the most important possibility to investigate.

Suppose X is associated with homicide at 36× the rate of the rest of the population.

Researchers discover that X is also disproportionately exposed to:

• a particular environment,
• a particular social network,
• childhood violence,
• a particular illicit market,
• poverty,
• a particular age distribution,
• or some other factor.

Once those variables are controlled for, the apparent effect of X disappears.

Then X isn’t really the underlying cause. It’s functioning as a proxy.

In that situation, the rational intervention is aimed at the underlying causes rather than X.

5. But suppose X remains highly predictive

This is where the hypothetical becomes genuinely difficult.

Imagine researchers establish that X remains an extraordinarily strong predictor even after controlling for everything else.

There are still different types of intervention.

Universal interventions could be applied to everybody:

• better homicide investigation,
• violence interruption,
• environmental crime prevention,
• improved emergency response,
• reducing access to commonly used lethal means.

Risk-based interventions could focus on individuals exhibiting additional warning signs.

Group-based interventions would specifically impose restrictions or surveillance on members of X.

The last category has the greatest danger of producing large numbers of false positives.

6. False positives become enormous when the base rate is low

This is a classic problem in statistical decision theory.

Suppose membership in X makes someone 36 times more likely to commit homicide. That sounds extraordinarily predictive.

But if homicide is still extremely rare, most people identified as “high risk” will nevertheless never commit homicide.

That’s why a policy can simultaneously be:

• statistically accurate at the population level, and
• terribly inaccurate when applied to individuals.

This distinction is often lost in discussions of crime statistics.

7. The most effective use of the information may therefore be indirect

If X is genuinely associated with 80% of homicides, knowing that fact could still be enormously useful without imposing restrictions on X itself.

Researchers could ask:

What is different about the 80% of homicides involving X?

For example, perhaps 70% involve interpersonal disputes, 60% involve firearms, 50% involve prior victim-offender relationships, and 80% involve people with previous violent encounters.

Those findings produce actionable intervention points.

The original characteristic then becomes a clue about where to look rather than the basis for punishment.

8. There’s also a distinction between prevention and punishment

This is particularly important.

If someone has already committed a violent offense, the state has evidence about that individual.

If someone has threatened to kill somebody, there is evidence about that individual.

If someone possesses a weapon while making a credible threat, there is substantially more evidence.

But:

“This person belongs to a demographic group that produces most homicide offenders”

is evidence about a population, not evidence that the particular individual has committed or intends to commit a crime.

That distinction is fundamental to any system based on individual rights.

The deeper question

This hypothetical ultimately leads to a more interesting question than “should we target the subgroup?”

It is:

How much predictive power should group membership have before it becomes legitimate to use it in an individual risk assessment?

There isn’t a simple mathematical answer, because that requires deciding what false positives and false negatives society is willing to tolerate.

And there is an important asymmetry:

• False negative: you fail to intervene against someone who subsequently commits homicide.
• False positive: you restrict, surveil, stigmatize, or otherwise interfere with someone who would never have committed homicide.

A homicide-prevention system necessarily has to balance those errors.

If we want to push the hypothetical to its logical extreme, the next interesting case is X = 50% of the population but responsible for 99% of homicides, or conversely X = 1% of the population responsible for 80%. Those produce radically different policy implications even though the headline statistic “X commits 80% of homicides” is identical.