Definition

A definition is a determination of exactly what a word means. It draws a boundary: what still belongs to it, and what no longer does? An example: A square is a quadrilateral with four sides of equal length and four right angles. This description is built so that anyone can check it. One can measure any given shape and afterward say clearly yes or no. This is exactly what distinguishes a definition from a mere description like “a square looks neat.”

When a single word decides money and law

In news texts, definitions seem inconspicuous, yet they are often the actual point of contention. The EU’s regulation on artificial intelligence, the AI Act for short, divides systems into risk classes. Whether a piece of software counts as a “high-risk system” hinges on a definition fixed in the text of the law. If a product falls under it, the manufacturer must organize testing, documentation, and oversight. That costs millions. If it narrowly misses the mark, almost all of that falls away.

With technical figures too, the definition decides whether a comparison is valid. Companies advertise that their model achieved a certain percentage on a test. Does the test allow multiple attempts per task? Does only the first attempt at an answer count? Depending on how this is fixed, the result changes considerably. Two numbers placed side by side are only meaningful if the same thing was measured both times.

A common misconception is that a definition is simply true or false. It is, at first, an agreement. But one can afterward judge whether it is useful: whether it is clearly decidable and whether it separates the cases one actually wants to separate.

The blueprint of a clean conceptual definition

The classic structure consists of two parts. First, one names the broader category to which the thing belongs. Then one names the feature that distinguishes it from all other members of that category. For a square, the broader category is the quadrilateral. The distinguishing feature is the equal sides and the right angles.

Three mistakes keep recurring. The first is circular reasoning: explaining a word with itself, as in “intelligence is the ability to act intelligently.” The second is a definition that is too broad, one that also covers things that were never meant to be included. The third is a definition that is too narrow, one that excludes clear-cut cases.

For this reason, standards and laws often work with so-called operational definitions. Here, what is stated is not the essence of a thing, but a measurement procedure. Instead of explaining what “fast” is, one specifies: the time is measured from sending the request to the first character output. This sounds bureaucratic, but it makes statements verifiable.

Definitions in dealing with AI and in everyday school life

Anyone using a chatbot is constantly working with definitions without calling them that. The instruction “summarize the text briefly” is unclear because “briefly” is not defined. The instruction “summarize the text in no more than five sentences” delivers more reliable results. One has drawn the boundary oneself, instead of leaving it to the program.

The same scrutiny pays off with product announcements. When a provider talks about an “agent,” this sometimes means a program that independently plans multiple steps, and sometimes just a chatbot with access to a search function. Similarly unclear is the term “open source” for AI models. Sometimes the training data and code are open, sometimes only the finished model weights are, under a restricted license.

In practice, this means: with every bold claim, it is worth asking how the terms used are defined. Who is defining this, and what is deliberately left out? This question costs little time and protects against numbers that look impressive but measure nothing comparable.