Machine learning is often described as a way for computers to learn from experience rather than follow only instructions explicitly written for each task. For industrial automation engineers and decision-makers, that description is useful—but it can also raise a broader question: does a system that improves its performance actually learn, or does it merely produce better results? The answer depends on what is meant by learning.
Why the meaning of machine learning matters
There are two distinct ways to interpret “real learning.” In the broad sense, it means learning that genuinely occurs, rather than a process that is only labelled learning. In a narrower sense, it means a high form of human learning involving understanding and the acquisition of knowledge supported by adequate justification.
These definitions lead to different conclusions. If learning means a change in performance resulting from experience, many machine learning systems qualify. If it requires an agent to understand what it has learned and explain why its conclusions are justified, current machine learning may not qualify. The disagreement is therefore not necessarily about whether a system changes its behaviour. It is about what additional conditions make that change “real learning.”
Performance improvement is one definition of learning
A widely used way to describe machine learning is to ask whether a program improves at a task through experience. The task, experience and measure of performance are specified: a program is said to learn when its performance on a task improves as it gains experience, according to a chosen measure.
For example, a checkers program could gain experience by playing games and improve its ability to win. That improvement can be called learning under a performance-based definition, even if the program does not understand the game in the human sense. The definition focuses on observable change in performance, not on the system’s awareness or comprehension.
This broad use of learning also covers different kinds of machine learning work. Systems may be trained with labelled examples, seek patterns in data that has not been labelled, or use trial and error to find actions that produce greater cumulative rewards. In each case, the method differs, but the central question remains: does experience change performance on the task?
Learning, knowledge and understanding are not the same thing
Human learning is often discussed as the acquisition of knowledge and understanding. One useful distinction is between knowing-that—knowing a fact or proposition—and knowing-how—being able to perform a skill or task. Knowing how to do something in practice is not necessarily the same as knowing facts about how it is done.
This distinction helps clarify why a machine can display a form of task competence without necessarily demonstrating human-like understanding. A program may improve at classifying information or controlling a task, yet that improvement does not by itself show that it understands the relevant facts, can justify its conclusions, or knows how to perform in the same sense as a person.
Some demanding accounts of real learning require more than a correct output. In formal mathematics, for instance, one proposed standard requires an agent to understand a function’s formal definition, produce its values, and provide a proof that the results are correct. A related account treats genuine learning as acquiring knowledge that is true, believed, and supported by a sufficiently strong justification. These standards make understanding and reasons central, not just successful performance.
What this distinction means for automation teams
For technical decision-makers, the debate is best treated as a question of evaluation rather than terminology. Calling a system “learning” does not, on its own, establish what it can do or how its outputs should be interpreted. A performance-based definition tells you that a model’s results improve under specified conditions. It does not automatically show that the model understands the task or can justify its output.
When assessing a machine learning application, keep the claims tied to the evidence and criteria being used. Consider questions such as:
- What is the task? Identify the activity the system is expected to perform.
- What counts as improvement? Specify the performance measure used to judge results.
- What experience informs the system? Clarify whether it uses labelled examples, unlabelled data, or trial-and-error feedback.
- What does “learning” mean in the claim? Distinguish better task performance from understanding or justified knowledge.
These questions do not diminish the practical value of machine learning. They help prevent a performance result from being mistaken for a broader claim about understanding, general capability or human-like reasoning. They also make discussions among engineering, operations and business teams more precise.
Why study machine learning beyond its results?
Machine learning is worth examining for several reasons beyond building systems that improve at tasks. Studying computational models of learning can contribute to understanding learning processes. Developing machine learning also aims to give computers the ability to learn and extends the range of problems to which computers can be applied. Finally, the spread of artificial intelligence raises wider questions about its effects on society, education, employment and ethical debate.
For industrial organizations, the conceptual distinction is practical: describe a system in terms of the task it performs, the experience it uses and the outcomes by which it is evaluated. Avoid assuming that the word “learning” means the same thing in every technical or philosophical discussion.
Conclusion: define learning before judging it
Machine learning can be considered learning when the term refers to improved performance through experience. Under a stricter definition—one requiring understanding and justified knowledge—it may not count as real learning. Neither conclusion follows from the label alone. For automation professionals, the useful approach is to define the capability being discussed and assess it against explicit evidence. That keeps expectations grounded and makes decisions about AI systems clearer.
Access Official Technical PDF