Evaluation

When a company releases a new AI program, it usually claims it is better than the previous one. How would it know that? It has tested it. The program is presented with many tasks whose correct solutions are already known, and it is counted how often it gets things wrong. This systematic checking is exactly what is called evaluation. The comparison to a school exam works well here: there is a fixed set of questions, a model answer, and a grade at the end.

Why claims without numbers are worthless

AI systems often appear more convincing than they actually are. A language model formulates fluently and confidently, even when the content is wrong. A human who asks a few questions can easily be fooled by this. Only a structured measurement across thousands of tasks reveals where the actual weaknesses lie.

For companies, a lot of money hinges on this. Anyone deciding whether to switch their customer service system to a new model needs hard comparative figures. Investors, too, look at test results when assessing the value of an AI company. That is why evaluation results regularly appear in stock market news and not just in trade journals.

There is also a legal dimension. The European AI Act requires proof for high-risk applications that a system operates reliably. Proof without measurement is not possible. Evaluation thus turns from a voluntary quality control into an obligation.

Test sets, metrics, and the process of a measurement

The basis is a test set: a collection of tasks along with the stored correct answer. It is important that the model has not seen these tasks during training. Otherwise one is only testing memorization, not ability. For this reason, the data is separated from the start into a training portion and a test portion.

From the results, key figures are calculated, so-called metrics. The simplest is the accuracy rate: the proportion of correctly solved tasks. For tasks without a clear-cut solution, such as summarizing a text, it becomes more difficult. Then either humans evaluate the outputs, or a second AI model takes on the role of judge. Both approaches have well-known weaknesses: humans are expensive and disagree with one another, while AI judges often favor long, wordy answers.

A common problem is called contamination. Here, the test questions have accidentally ended up in the training data, because both originate from the internet. The model then already knows the answers and achieves dream results that collapse in real-world use. For this reason, research groups continuously develop new, secretly kept test sets.

From leaderboards to practical testing within companies

The most visible are public rankings, so-called leaderboards. There, models from different providers stand side by side with their scores. Well-known test collections examine school and university knowledge, programming tasks, or mathematical reasoning. When the media report that a model has passed an exam, such a test is almost always behind it.

In addition, there are comparisons in which real users see two anonymous answers and choose the better one. From many such duels, a ranking emerges similar to chess ratings. Such methods measure less the subject-matter knowledge than the subjective impression. Both perspectives complement each other, and neither replaces the other.

In everyday business, companies additionally build their own evaluations. An insurance company tests its system with a hundred typical claims reports from its own archive. These internal tests reveal more about the concrete benefit than any public leaderboard. A common misconception is that a top spot on the leaderboard automatically means the best model for one's own task. Often a smaller, cheaper system wins because it fits the specific use case better.