Can utility be negative?

Yes. Its origin is a modeling choice, and negative values can conveniently represent unfavorable outcomes.

Is utility the same as a reward?

A reward is a signal in a learning setup; its relationship to the intended preferences must be designed and assessed.

Must an AI system feel satisfaction to use utility?

No. Utility can simply be a formal representation guiding its decisions.

AI Glossary / Graphs and symbolic AI / Advanced

Utility

Definition

Utility is a numerical representation of how desirable an outcome is to a particular decision-maker or agent. Higher utility represents a more preferred outcome within the specified model. In expected-utility decision-making, outcome utilities are combined with probabilities to compare actions whose results are uncertain.

What it means

An AI system needs some account of what counts as a good result. Predicting what will happen does not answer what should be preferred. Utility supplies the preference side of that distinction: it might reflect service quality, resource consumption, reliability, inconvenience, or several consequences together.
Utility is not an objective measure of goodness shared by everyone. It depends on whose preferences are represented and what the outcome description includes. A value of 80 is not an 80% probability, and a value of zero need not mean an outcome has no practical consequences. The numerical scale must be interpreted through the model that defines it.

How it works

First describe the relevant outcomes and establish preferences among them. A utility function assigns numbers that represent those preferences. When uncertainty matters, the assignments must also support the intended comparisons between risky alternatives; ranking certain outcomes alone does not settle those comparisons.
For an action a with possible outcomes o, expected utility is the sum of P(o given a) multiplied by U(o). The agent compares these expectations across available actions, subject to any separate constraints. The highest expectation identifies the preferred action under those modeled probabilities, utilities, and constraints.
This is a decision rule, not a promise about the next realized outcome. An action with higher expected utility can still produce an unfavorable result. Appropriate preferences and accurate probability estimates remain essential inputs.

Three practical examples

A scheduling assistant compares appointments using the user's preferences for travel, waiting time, and disruption. A convenient appointment can have high utility even if another appointment is easier for the system to book.
A delivery planner evaluates service outcomes that include arrival timing, damage, and cost. The most likely delivery outcome is not necessarily enough to choose a carrier; less frequent but serious failures can affect the decision.
A game-playing agent assigns utilities to terminal results such as wins, draws, and losses. Its search uses those preferences to compare possible continuations, while probabilities or opponent models describe how those results might occur.

Think of it this way

Probability answers how likely a destination is to be reached. Utility answers how much the traveler values arriving there under the specified conditions. Choosing a route requires both when destinations or conditions are uncertain. Knowing that a destination is likely says nothing by itself about whether the traveler wants it.

Worked example

Consider a fictional delivery decision with two feasible options. All quoted utilities already include the relevant service and cost consequences; they are illustrative preference values, not currency amounts or measured performance.
Option A produces its routine on-time outcome with probability 0.95 and utility 70. With probability 0.05 it fails, giving utility -100. Its expected utility is 0.95 × 70 + 0.05 × (-100) = 61.5.
Option B produces an excellent outcome with probability 0.80 and utility 100. Otherwise it produces an acceptable delayed outcome, with probability 0.20 and utility 50. Its expected utility is 0.80 × 100 + 0.20 × 50 = 90.
The expected-utility model therefore prefers B. Looking only at A's 95% chance of its routine outcome would miss differences in the consequences. B can still realize the delayed outcome with utility 50, below A's routine outcome of 70; the comparison does not guarantee B wins on every occasion.
If every utility is replaced by twice its value plus 10, the expectations become 133 for A and 190 for B. The preference remains unchanged. Positive changes of scale and origin preserve this expected-utility comparison; arbitrary nonlinear changes need not.

Illustrative case study

In a fictional customer-support operation, an automation project initially treats every quickly closed ticket as equally valuable. The resulting system closes straightforward cases rapidly but also encourages premature closure of unresolved cases.
The team reviews the outcome definition with service staff and customers. It distinguishes verified resolution, repeat contact, unresolved closure, and the effort required from the customer. These distinctions inform a revised preference model, while mandatory escalation rules remain explicit constraints.
The team then evaluates decisions against those outcomes rather than celebrating a higher closure count alone. This does not prove that its numerical values perfectly represent everyone's interests. It makes the assumptions inspectable and allows reviewers to identify missing consequences before the system is trusted with broader decisions.

Common misconceptions and mistakes

“Utility means profit.” Money can matter, but the model may include nonfinancial consequences.
“A high utility is a high probability.” Utility represents desirability; probability represents uncertainty.
“Higher expected utility guarantees a better outcome.” It describes a preference over uncertain alternatives, not a guarantee for an individual trial.
“An arbitrary rating can always be averaged with probabilities.” Expected-utility use requires a scale appropriate to preferences over risky outcomes.
“Utilities from different people are automatically comparable.” Their origins, scales, and underlying preferences may differ.

Business use

Utility provides a way to make decision criteria explicit, especially when an AI system must trade off several consequences. Begin by stating whose preferences count, the relevant time horizon, and which requirements cannot be traded away.
Inspect the outcomes rather than only adjusting numerical weights. If a model omits repeat work, accessibility, or effects on another team, optimizing its utility can still produce poor decisions. Use sensitivity analysis to see whether modest changes in preferences or probabilities reverse the recommendation. Keep a record of how the values were obtained and who is responsible for revising them when the operating context changes.

Check your understanding

Why does the example prefer B?

Answer: Its expected utility is 90, compared with A's 61.5.

Does B's selection make delay impossible?

Answer: No; the model assigns delay probability 0.20.

Why cannot probability replace utility?

Answer: Knowing what is likely does not establish which consequences are preferred.

Related terms

Utility function
Expected utility
Preference elicitation
Decision theory
Rational agent
risk attitude
reward function
multiattribute utility

Learn next

Compare rankings of certain outcomes with preferences over lotteries. Then examine how probability errors and omitted consequences can change a decision even when the expected-utility calculation is arithmetically correct.

Expected utility

Sources

Poole and Mackworth, Preferences and Utility: https://artint.info/3e/html/ArtInt3e.Ch12.S1.html University of California, Berkeley, CS 188, Utilities: https://inst.eecs.berkeley.edu/~cs188/textbook/vpis/utilities.html

Previous: UTF-8 Next: Utility function