Is equal impact the same as equal treatment?

Not necessarily. Applying the same procedure can lead to different modeled consequences.

Can equal selection rates prove equal impact?

No. Selection rates alone omit the utilities of subsequent outcomes.

Is there one accepted formula?

No. The phrase is used with different formalizations, so a source and operational definition are necessary.

AI Glossary / Graphs and symbolic AI / Intermediate

Equal impact

Definition

Equal impact is a fairness objective concerned with the benefits and costs that a decision system creates for affected people. In the AIMA usage relevant here, people with comparable likelihood of the relevant outcome should receive the same expected utility regardless of demographic class. This considers consequences of correct and incorrect predictions, rather than relying only on equal decision rates.

What it means

Two systems can approve the same proportion of requests while producing different consequences for the people involved. A missed opportunity may impose a larger burden in one situation, and a mistaken positive decision may create costs that an approval count does not capture.
Expected utility summarizes outcomes using stated values for benefits and harms, weighted by their probabilities. Choosing those values is a substantive modeling decision. Equal impact is therefore not one universally defined spreadsheet column. Research also uses the phrase for other formulations, including long-run effects across repeated interactions; the chosen meaning must be stated.

How it works

Begin by identifying whose outcomes matter and what counts as a benefit or cost. Specify the relevant comparison group, outcome likelihood, decision alternatives, and time horizon. Estimate outcome probabilities and assign an explicitly justified utility model, then compare expected utilities under the decision process.
For a simple finite example, multiply each possible outcome's utility by its probability and add the results. An operational study needs more than this arithmetic: uncertainty, missing outcomes, and the choice of comparison variables affect interpretation. A numerical equality can hold under one utility model and fail under another, so the assumptions belong beside the reported result.

Three practical examples

A fictional training-program study compares the consequences of incorrect acceptance and incorrect rejection, rather than checking only acceptance percentages.
An AI ethics workshop asks participants to specify whose gains and losses are represented in an automated allocation objective.
A service-design simulation compares one-time decision statistics with cumulative effects when users interact with the same system repeatedly.

Think of it this way

Think of two people receiving equally sized boxes whose contents have different value to them. Counting boxes captures a distribution of outputs but may miss their effects. The analogy is limited because human welfare cannot generally be reduced to an agreed single number, and a defensible fairness analysis needs more detail than the size of an individual benefit.

Worked example

Consider a purely fictional decision exercise with two otherwise comparable participants, A and B. Both have probability 0.8 of a successful outcome if offered a training place. For A, suppose the illustrative utility is +10 on success and -5 on failure. Its expected value is 0.8 × 10 + 0.2 × (-5) = 7.
For B, use the same success utility but an illustrative failure burden of -15. The expected value is 0.8 × 10 + 0.2 × (-15) = 5. Identical predicted success probabilities and identical offers therefore do not produce equal expected utility under this stipulated model.
This arithmetic does not justify treating B less favorably. It instead reveals a question about the modeled burden and possible support. If a hypothetical support measure reduced B's failure burden to -5 without other changes or costs, the expected value would also be 7. Real analysis would have to investigate whether that assumption is credible and whose costs the support introduces.

Illustrative case study

In a fictional design seminar, the Ash team presents an allocation tool with equal selection rates across two groups. A reviewer asks what happens after selection and what an erroneous rejection costs. The team realizes that its dashboard describes decisions but contains no model of their effects.
It adds a separate impact-analysis exercise with transparent assumptions and sensitivity checks. Participants examine several plausible benefit and burden values instead of declaring the tool fair from one favorable calculation. This invented case illustrates a broader question, not a validated fairness intervention or an actual organizational result.

Common misconceptions and mistakes

Equal impact does not mean that every person must receive exactly the same realized outcome. Expected values can match while individual outcomes differ. It is also distinct from demographic parity, which concerns selection rates, and from accuracy, which counts predictive correctness under a specified measure.
Do not treat a convenient utility scale as an objective measurement of human worth. Do not substitute a long-run equal-impact definition from one paper for AIMA's consequence-based comparison without saying so. No single metric establishes that a system satisfies every ethical concern or applicable requirement.

Business use

The concept helps business teams ask what an AI decision actually changes for customers, workers, or other affected people. It encourages discussion of false predictions, downstream consequences, and the people whose outcomes are absent from the objective.
A useful review records the fairness definition, comparison conditions, utility assumptions, and uncertainties. Involve people able to assess the real consequences represented by the numbers. The entry provides an analytical vocabulary for that discussion; it does not prescribe a legal test or a universal formula for deployment approval.

Check your understanding

In the toy example, both participants have success probability 0.8. Why are their expected utilities different?

Answer: The assumed cost of failure differs. Expected utility includes both success and failure outcomes, so matching the success probability does not by itself make the full weighted sum equal.

Related terms

Learn next

Learn next about fairness metrics and utility modeling. Write down the exact population, outcome, and time horizon for a fictional allocation task. Compare an output-rate measure with a consequence measure, then vary one utility assumption to see which conclusions remain stable.

Sources

Russell and Norvig, Artificial Intelligence: A Modern Approach, original Philosophy and Ethics of AI chapter, section 27.3.3, equal-impact passage (page 999), university-hosted copy consulted: https://zoo.cs.yale.edu/classes/cs470/lectures/newchap27.pdf Zhou and colleagues, original paper illustrating a distinct repeated-interaction formulation: https://arxiv.org/abs/2209.01410

Previous: Epsilon greedy policy Next: Equality of opportunity