Does every AI system learn while it is being used?

No. Many systems use a fixed trained model during operation and change only through a separate update process.

Does learning require labeled examples?

No. Labeled examples support supervised learning, but other approaches use unlabeled data or feedback from actions and outcomes.

Can learning produce undesirable behavior?

Yes. A system can fit biased, misleading, or irrelevant patterns, or optimize a measure that fails to capture the intended goal.

AI Glossary / Governance, ethics and safety / Beginner

Learning

Definition

Learning, in AI, is the process through which a system uses experience or data to develop or improve its ability to perform a task. A useful operational account specifies the task, the experience available, and the measure of performance. Learning does not imply human understanding, and running a learning procedure does not guarantee useful improvement.

What it means

Instead of specifying every response directly, developers give a learning system a way to adjust from examples, observations, or feedback. What changes may be numerical parameters, a decision structure, stored examples, or another representation used to perform the task.

The word improvement needs a reference point. A system might become better at matching historical labels while remaining poor at handling a new setting. It might also improve one measure while worsening another that matters to users.

In everyday use, people may say that an assistant learns when it follows information supplied during a conversation. That behavior does not, by itself, show that the underlying model has undergone a lasting training update. The mechanism and persistence of the change should be distinguished.

How it works

A learning process combines experience, a model or representation, and a rule for changing that representation. In many supervised systems, examples include desired outputs, and training adjusts the model to reduce an error measure. Other approaches learn patterns without supplied target labels or use feedback from actions.

Developers choose what information is available, what objective guides changes, and how progress is evaluated. Those choices influence what the system can learn and which mistakes it may repeat.

Evaluation on suitable data not used to fit the model helps assess generalization. Good performance on the training examples alone can reflect memorization or exploitation of accidental patterns. A deployed system may remain fixed between approved updates, or it may use an explicit continuing-learning process. New user interactions do not automatically mean that the model updates from every interaction.

Three practical examples

1. A document classifier learns from examples assigned to categories, then attempts to classify new documents whose categories have not been supplied to it.

2. A grouping system learns patterns of similarity in product records without being given a target category for each record, helping analysts explore the catalog.

3. A simulated control agent adjusts its behavior using feedback from actions and their outcomes, with evaluation checking whether the resulting behavior achieves the intended task under relevant conditions.

Think of it this way

Think of practicing a skill using exercises and feedback. Repeated experience can improve performance, but practice with misleading feedback or unrepresentative exercises can also reinforce poor habits.

The analogy is limited. Machine learning does not establish that the system experiences practice, understands its purpose, or learns in the same way as a person. Its changes follow computational procedures and the information provided. Human-like language about learning should not substitute for knowing which part of the system actually changes.

Worked example

A fictional team trains a classifier to sort maintenance requests into two categories. Before training, a baseline correctly classifies 120 of 200 held-out requests, giving 60% accuracy. After training on a separate dataset, a candidate correctly classifies 156 of the same 200 held-out requests, giving 78% accuracy.

The observed increase is 18 percentage points. Errors fall from 80 to 44, a reduction of 36 / 80 = 45% on that evaluation set.

These figures describe the comparison, not a guarantee of future performance. The team must check that the held-out labels are reliable, that information did not leak into training, and that the requests resemble the intended use.

If the team repeatedly tunes the model against those same 200 cases, they cease to provide a clean final assessment. An additional held-out evaluation can then be needed to assess the selected model.

Illustrative case study

Fictional case: Orchard Records develops a document sorter using historical examples. Training accuracy becomes very high, but performance drops on recent documents. Investigation shows that an old template marker strongly predicted the historical category and is absent from the new documents.

The team revises the data and evaluation design to examine whether the sorter captures relevant document content rather than relying mainly on that marker. It also records the time period and template coverage of its training material.

The case is invented. It illustrates that learning can capture an available shortcut rather than the relationship developers hoped for. More training on the same narrow evidence would not necessarily resolve the underlying problem.

Common misconceptions and mistakes

Learning is not the same as storing every input, and it is not synonymous with becoming intelligent in every domain. A system can improve a narrow task without gaining broad reasoning ability.

More data or longer training does not always produce a better deployed model. Data quality, relevance, objectives, and evaluation all matter. A lower training error also does not establish better generalization. Finally, saving a user's preference in a database and retraining model parameters are different mechanisms, even if both change later behavior.

Business use

Businesses should define the intended learning problem before selecting a tool: what task should improve, what experience can support it, and how success will be measured. Include the costs and consequences of different errors, not just an aggregate score.

Document data sources, update procedures, evaluation boundaries, and responsibility for accepting a new model. If learning continues after deployment, assess how changes are monitored and controlled. Benefit comes from validated improvement in the intended workflow, not merely from describing a product as self-learning.

Check your understanding

Training accuracy rises, but performance on new representative examples falls. Has the evidence established useful improvement for deployment?

Answer: No. The model may be overfitting or learning an unsuitable pattern. Improvement should be assessed against the intended task and appropriate evaluation evidence, not training accuracy alone.

Related terms

Learn next

Next, distinguish training from inference. For one proposed AI project, write the task, available experience, and performance measure, then explain how you would check whether improvement transfers to new examples.

Sources

Tom Mitchell, primary machine-learning course description: https://www.cs.cmu.edu/~tom/10601_sp09/about.html Google Machine Learning Glossary, training and generalisation: https://developers.google.com/machine-learning/glossary These primary passages were read; the task-experience-performance formulation was also checked in an indexed excerpt of Mitchell's textbook. The complete textbook PDF exceeded retrieval limits. Examples and calculations are original.

Previous: Learned sparse retrieval Next: Learning automaton