Can a rational decision have a bad outcome?

Yes. Uncertainty means a justified choice can still turn out unfavorably.

Must rational agents use machine learning?

No. Learning can improve their information, but the decision principle does not specify one implementation method.

Does rationality guarantee ethical behavior?

No. The objective, constraints, and assumptions require separate scrutiny.

AI Glossary / Graphs and symbolic AI / Intermediate

Rationality

Definition

In AI, rationality is the principle of choosing actions that maximize expected performance, given the agent's available evidence, prior knowledge, and possible actions. It evaluates decisions in relation to a specified task and performance measure. It does not require knowing the future or obtaining the best actual outcome on every occasion.

What it means

An agent observes something about its environment and acts. Calling its behavior rational means asking whether its choices make sense for the objective and information it has. A choice can be justified before the outcome is known even when an unlikely unfavorable event later occurs.
The objective matters. An assistant rewarded only for completing requests quickly may behave differently from one evaluated on correct completion, user preferences, and the consequences of mistakes. Rationality relative to a poorly designed objective is not a guarantee that people will approve of the resulting behavior. The technical term should therefore not be used as a general endorsement.

How it works

Specify the environment, permitted actions, observations, and measure of success. Estimate the possible consequences of each action using the information available. Then compare expected performance, accounting for uncertainty. New evidence can change which action is preferred.
Gathering information can itself be an action. An agent may inspect a situation or ask a useful question before committing, when the expected benefit exceeds the cost. Practical systems also face limited time and computation. Their implementations approximate the ideal rather than evaluating every conceivable future. The assumptions and resource limits should remain visible when judging their decisions.

Three practical examples

A scheduling assistant can compare appointment options using the user's stated preferences, travel constraints, and the likelihood of availability.
An inventory simulation can compare ordering policies under uncertain demand and an explicitly defined penalty for excess or missing stock.
A support assistant can request a product version before suggesting instructions when the answer could change which instructions are appropriate.

Think of it this way

Think of selecting a route using the traffic information available when you leave. A reasonable choice can still encounter an unexpected delay. The analogy is limited because AI requires an explicit representation of actions, evidence, and success; a person's everyday judgment may incorporate considerations that the model has omitted.

Worked example

Consider an original planning example for a small event. The objective is to minimize a simple disruption score; all numbers are invented. An outdoor setup scores 0 if the weather is dry and 20 if it rains. An indoor setup scores 6 in either case. The current probability of rain is 0.2.
The outdoor setup has expected score 0.8 × 0 + 0.2 × 20 = 4. Indoors has expected score 6. Under this model, outdoors is preferred. If rain actually occurs and produces score 20, that outcome alone does not show that choosing outdoors was irrational when the decision was made.
Suppose a revised forecast raises the rain probability to 0.4. The outdoor expectation becomes 8, so indoors is now preferred. The decision threshold is 20p = 6, giving p = 0.3; the two options tie at that probability.
For a further idealized calculation, imagine perfectly learning the weather before choosing. One would choose outdoors when dry and indoors when rainy, giving expected score 0.8 × 0 + 0.2 × 6 = 1.2. Relative to the original best score of 4, perfect information can improve the expected score by 2.8. An information source costing more than 2.8 score units cannot be justified by that benefit in this simplified model. Real forecasts are imperfect, so this is an upper benchmark rather than a purchase recommendation.

Illustrative case study

In a fictional case, Marigold Events evaluates a planning assistant using a disruption score similar to the example. Staff initially judge every unfavorable result as an assistant failure. They then separate incorrect reasoning from reasonable decisions followed by unfavorable outcomes.
The review also reveals that the score omitted accessibility needs. The team revises the task specification before evaluating further choices. This illustrates two different checks: whether the agent chooses well under its model, and whether that model expresses the actual priorities. The case describes no real business results.

Common misconceptions and mistakes

Rationality is not omniscience, infallibility, or a synonym for intelligence in every sense. It does not mean choosing the outcome with the highest probability while ignoring its value or cost. Nor does it require maximizing money; performance can represent other goals.
An agent cannot justify avoidable ignorance simply by pointing to missing evidence. If obtaining relevant information is feasible and worthwhile, doing so belongs in the decision problem. Conversely, endless investigation can be unreasonable when its cost exceeds its expected benefit.

Business use

For business use, document what the agent optimizes and which constraints restrict its actions. Review representative decisions using the evidence available at the time, while separately measuring whether probability estimates and outcomes are well calibrated.
Keep objective design distinct from optimization quality. A system can accurately optimize a measure that rewards the wrong behavior. Testing should include cases where asking, waiting, escalating, or declining an unsuitable action better serves the specified task than immediate completion.

Check your understanding

At a rain probability of 0.4, which option minimizes the example's expected disruption?

Answer: Indoors, with score 6, rather than the outdoor expectation of 8.

Related terms

Learn next

Change the indoor score from 6 to 10. The threshold becomes p = 0.5. Explain why the preferred action can change even when the weather evidence remains identical: the consequences, not just probabilities, enter the decision.

Sources

Sources: Russell and Norvig, Artificial Intelligence: A Modern Approach, author-hosted Intelligent Agents chapter, Section 2.2 on rationality, performance measures and information gathering: https://aima.cs.berkeley.edu/4th-ed/pdfs/newchap02.pdf .

Previous: Rational subgroups Next: Rauch-Tung-Striebel smoother