Grader

A grader is a program that evaluates the output of a computer program and assigns it a grade or score. The English word “to grade” means exactly that. In AI development, a grader checks the responses of a chatbot or another learning system. It answers a simple question: was this response good or bad? Instead of a human reading through thousands of answers, software does it in seconds. You can think of a grader as the automatic correction of a class test, except here it’s not students being examined, but machines.

No progress without evaluation

An AI system only learns if someone tells it what is better and what is worse. This is exactly the feedback the grader provides. It is thus the yardstick against which the entire training is aligned. Without it, the system wouldn’t know in which direction it should change.

The second reason is simply the sheer volume. Large providers test their models on hundreds of thousands of tasks before a new version is released. Human reviewers would need months for that and cost a lot of money. A grader accomplishes the same work overnight on a server.

But this gives rise to a risk that experts take seriously. A system always optimizes for the yardstick it is given, not for what one actually wants. If the grader rewards long answers, the answers become long instead of good. This exploitation of weaknesses in the yardstick is called reward hacking. A poorly built grader therefore doesn’t just fail to improve a model — it measurably makes it worse.

From fixed rules to evaluating models

The simplest design is a grader with a fixed rule. For a math problem, it compares the AI’s result with the stored solution. If it matches, it gets a point, otherwise zero. For coding tasks, it runs the written code and checks whether it passes the test cases. Such graders are fast, cheap, and completely unambiguous.

It gets harder with tasks that don’t have a single correct solution. For summarizing a text or giving a polite reply to a customer, there is no fixed answer key. Here, a second AI model is often used as an evaluator. It is given the task, the answer, and a list of criteria, and assigns points based on that. This procedure is called LLM-as-a-Judge, meaning a language model acting as a judge.

Both designs have typical weaknesses. The fixed rule is rigid: it counts a factually correct answer as wrong if only the wording differs. The evaluating model, on the other hand, is lenient and often favors confident-sounding answers, even if they are factually wrong. That’s why developers first check their graders against human evaluations. Only once both judgments largely agree is the grader considered usable.

Graders in benchmarks and developer tools

Graders are most commonly encountered in reports about benchmarks, i.e. standardized test suites for AI models. When a company reports that its model solves 71 percent of a task collection, a grader calculated that percentage. How strict it is determines the result. That’s why comparisons between two providers are only fair if both used the same grader.

The term also appears in developer products. The major providers offer tools that let companies write their own graders. An insurance company, for example, uses one to check whether its chatbot names the correct tariff number. Such checks run automatically with every new model version, similar to a security check before an update.

A common misconception is that a grader measures an objective truth. It only measures what its developers built into it. Anyone reading a number from a benchmark should therefore always ask who built the grader and by what criteria it judges.