AI Glossary / Governance, ethics and safety / Beginner

Outcome

Definition

An outcome is a possible or observed result of a process, action or experiment. In probability, an outcome is one element of a defined sample space. In AI evaluation, the word may instead refer to a measured result or a downstream consequence, so its scope should be stated.

What it means

An AI system can produce a recommendation, a user can act on it, and something can happen afterwards. Each stage has a result, but those results are not interchangeable. The model's output might be a delivery-time prediction; the observed outcome might be the actual arrival time; the business consequence might be a missed appointment.

In a probability model, an outcome describes one possible result at the level of detail chosen for the experiment. An event is a set of outcomes meeting some condition. This distinction helps analysts calculate probabilities without mixing the possible results with the questions asked about them.

For practical work, always ask: the outcome of what, for whom and over which period? A successful software execution is not necessarily a beneficial outcome for the person affected.

How it works

Define the process and the unit of observation before collecting results. A trial could be one delivery, one user session or a sequence of actions. Specify the possible results and how they will be observed, including missing or delayed observations.

A random variable assigns a value to each outcome. For a two-item shipment, an outcome might record whether each item arrived on time, while a random variable counts the number arriving late. Several different outcomes can receive the same count. A probability distribution describes how likely the values or outcomes are under the model.

In deployment evaluation, connect outputs to later observations carefully. A prediction can change an action, and that action can influence the eventual result. Comparing an outcome before and after deployment does not automatically establish that AI caused the change. Context, selection and other simultaneous changes may matter. Measurement definitions and causal conclusions therefore require separate attention.

Three practical examples

• A forecasting team compares predicted demand with observed demand during a defined week, keeping the prediction separate from the later result.
• A simulation records whether each of two machines completes its assigned job, then calculates the event that at least one job is unfinished.
• A support team distinguishes an assistant generating a reply from the customer actually completing the intended task.

Think of it this way

Think of planning a journey. The route suggestion is an output from the planner, arrival is an observed result, and reaching an important meeting on time is a consequence that matters to the traveler.

The analogy does not imply that every consequence can be attributed to the planner. Traffic, departure time and the traveler's decisions also influence what happens.

Worked example

Outcomes, events and a random variable

Consider an invented experiment with two deliveries, each classified as on time or late. At this level of detail, there are four outcomes: both on time; first on time and second late; first late and second on time; and both late.

Suppose the model assigns probabilities of 0.64, 0.16, 0.16 and 0.04 respectively. These add to 1. The event “at least one delivery is late” contains the last three outcomes, so its probability is 0.36.

Let X count late deliveries. X equals 0 for the first outcome, 1 for either middle outcome and 2 for the last. Its expected value is 0 × 0.64 + 1 × 0.32 + 2 × 0.04 = 0.40. That average is not a fifth outcome: no trial produces four-tenths of a late delivery.

Illustrative case study

Fictional case: Rowan Support reports that its new assistant successfully answers nearly every incoming request. A review discovers that “success” means an answer was generated without a technical error. It does not indicate whether the information helped the customer.

The team keeps the execution metric but names it accurately. It adds an appropriately designed task-resolution measure and records cases where the later outcome is unknown. It also examines other service changes before attributing improvements to the assistant.

The lesson is to preserve the chain from system output to observed result to claimed benefit. The case supplies no measured improvement and does not assume that every outcome can be observed directly.

Common misconceptions and mistakes

• An output and an outcome always mean the same thing. Some contexts use them similarly, but evaluation often needs to distinguish generated material from downstream results.
• An expected value must be a possible individual result. An average can lie between discrete possibilities.
• A better observed outcome proves an AI intervention caused it. Causal attribution requires a suitable design and assumptions.

Business use

A useful business evaluation names the outcome, the population, the observation period and the collection method. It also specifies how missing observations and unintended consequences are handled. Otherwise teams may compare numbers that describe different things.

Include measures of both intended benefits and relevant harms, rather than treating completion of a software action as sufficient evidence of success. When a decision depends on causal impact, ask what comparison supports that claim. Outcome monitoring can reveal changes that deserve investigation without independently proving their cause.

People also ask

Must an outcome be numerical?

No. A result can be a category, a sequence or another object; a random variable can map it to a useful value.

Is an event one outcome?

It can contain one outcome or several, depending on the condition being considered.

Can an outcome remain unknown?

Yes. Missing, delayed or unobservable results should be represented honestly rather than silently treated as success or failure.

Check your understanding

Two different shipment outcomes each contain exactly one late delivery. Are they the same outcome because the late-delivery count is equal?

Answer: No. They are distinct outcomes mapped to the same value of the counting random variable.

Related terms

• Sample space
• Event
• Random variable
• Expected value
• Causal inference
• output

Learn next

Read sample space and random variable. For one AI workflow, list the model output, the observed later result and the business consequence, then identify which of those your current dashboard actually measures.

• Sample space
• Random variable

More in Governance, ethics and safety

• Opt-in - Opt-in is a choice mechanism in which a person actively agrees to participate in a specified activity or enable a specified use.
• Opt-out - Opt-out is a mechanism through which a person declines, stops or excludes themselves from a specified activity or use.
• Out-group homogeneity bias - Out-group homogeneity bias is the tendency to perceive members of a group outside one's own as more alike than members of one's own group.
• Personal data - Personal data is information relating to an identified or identifiable person.
• Popularity Bias - Popularity bias is a tendency for already popular items to receive disproportionate exposure while less popular items receive little representation.
• Population Bias - Population bias is a systematic mismatch between the people represented in a dataset and the population about which conclusions or predictions are intended.

Sources

Jung and colleagues, The Aalto Dictionary of Machine Learning: https://aaltodictionaryofml.github.io/ADictML_Main.pdf. The outcome entry and random-experiment discussion were reviewed in returned text. NIST AI RMF Core: https://airc.nist.gov/airmf-resources/airmf/5-sec-core/. Its evaluation and context framework supports the broader governance discussion. Calculations and business scenarios are original teaching examples.

Previous: Outage probability Next: Outer Join