Spot-Checking

A computer program can churn out millions of results in a single day. No human can check that entire volume. So one takes a small, randomly drawn selection and looks at only those very closely. This is exactly what is called spot-checking, or in German roughly 'Stichprobenprüfung'. If hardly any errors are found in the sample, one assumes that the rest is largely fine as well. The principle is familiar from everyday life: customs doesn’t open every suitcase either, but pulls out individual travelers.

Why nobody reads all AI responses

Systems with artificial intelligence produce text, images, or decisions in volumes that make any complete check unaffordable. A chatbot from a major provider answers hundreds of millions of requests daily. Even if a reviewer needed only ten seconds per response, that would require tens of thousands of full-time positions. Spot-checking is therefore often the only practical form of quality control.

A second problem adds to this: many errors made by AI systems don’t stand out automatically at all. If a model invents a source that doesn’t exist, the response still looks clean and fluent. A computer detects nothing suspicious about it. Only when a human looks up the source does the error come to light. It is precisely such cases that are found almost exclusively through manual spot checks.

For companies, this also has a legal dimension. Anyone deploying an AI system must be able to show that they monitor its quality. Documented spot checks are a common form of evidence for this. They do not replace a thorough review, but they demonstrate that someone is paying attention.

What makes a good sample

What matters is how the examples are selected. If one simply takes the first hundred cases of the day, the selection is biased. Perhaps entirely different users write in the morning than in the evening. That’s why cases are drawn as randomly as possible, so that every result has the same chance of being checked. Only then is it valid to draw conclusions about the whole from the excerpt.

The second question is size. If one checks a hundred cases and finds three errors, the actual error rate probably lies somewhere between one and eight percent. This range is called the uncertainty interval and it shrinks the more one checks. With a thousand cases, the estimate becomes considerably sharper. Anyone drawing far-reaching conclusions from ten checked examples is therefore vastly overestimating their numbers.

Random sampling is often combined with targeted selection. Alongside the random sample, one additionally checks particularly sensitive cases, such as medical questions. However, these two figures must not then be mixed together. The targeted selection shows how bad things can get in the worst case. The random selection shows what things look like in normal operation.

From the factory floor to model comparison

The method originally comes from industry. In factories, individual parts have been pulled from production and measured for decades, instead of inspecting every single screw. Food authorities and financial auditors work this way too. The AI industry simply adopted the idea.

In practice, one encounters spot-checking today wherever models review content. Social networks have software filter posts and pass on only a fraction of them to human review teams. Companies that have customer inquiries answered automatically read through some of the dialogues themselves every day. And when a company introduces a new language model, employees compare the responses of the old and the new system on a handful of typical tasks.

A common misconception is that spot-checking is proof of being error-free. It is not. Rare but severe errors almost never show up in small samples. An error that occurs in one out of ten thousand cases will, with high probability, remain undetected among a hundred checked examples. Samples reveal tendencies, not guarantees.