Verification

Verification means: one checks whether a result is really correct. It is not about how good or plausible it looks. It is about genuine proof. An example: a program claims that 4879 is divisible by 7. One can recalculate this claim in a second and settle the matter. Exactly this separation is the core of the concept: producing something is one task, checking it is another. Often the checking is much easier than the producing.

Why convincing answers can still be wrong

Modern language programs like ChatGPT predict, word by word, what most likely fits next. This makes them produce texts that sound fluent and confident. Whether the content is true, they do not check. Fabricated statements of this kind are called hallucinations. For the reader they are hard to recognize, because they appear in the same assured tone as correct statements.

That is why verification is the decisive safeguard. It determines whether an AI system may be used in a serious domain. In a poem, an error is harmless. In a medication dosage, a tax calculation, or the control code of a vehicle, it is not. Companies therefore almost always build a checking stage behind the model before an output is passed on anywhere.

Verification has also become economically important. Where an answer can be checked automatically, expensive human oversight is saved. Where that is not possible, experts are needed who read every output. This is the reason why AI has gained a foothold faster in programming than in legal advice: a program can be run and tested, a legal assessment cannot.

From spot-checking to mathematical proof

The simplest form is the test. One gives the result concrete examples and checks whether it behaves correctly. For software this means: one runs the code and compares the output with the expected value. But tests only reveal errors that were actually tested for. Passing all tests is a good sign, not a proof.

More rigorous is formal verification. Here one proves mathematically that a program satisfies a certain property for all possible inputs. Special tools, so-called proof assistants, mechanically check every step of this proof. This is costly and is used mainly where errors would be catastrophic, for instance in chips, aircraft control systems, or encryption.

In between lie practical hybrid forms. One has an AI back up its answer with sources and checks the sources. One has a second model criticize the output of the first. Or a human reviews every proposal before it takes effect. It is important to distinguish this from validation: verification asks whether something was built correctly, validation asks whether it solves the right problem at all.

Verification in chatbots, code, and AI training

In everyday life one usually encounters it invisibly. When a chatbot displays links to websites, that is an invitation to verification: one can look up where the statement comes from. The notice that answers may contain errors also belongs here. It explicitly shifts the duty of checking to the user.

In software development, checking is firmly built in. An AI assistant proposes code, automated tests run over it, and only then does a developer accept the proposal. In the news, verification problems often surface as scandals, for instance when a lawyer submits fabricated court rulings because he did not check the model’s statements.

Verification also now plays the leading role in training AI. Models train preferentially on math and programming tasks, because there every solution can be judged mechanically as right or wrong. The model tries many paths, a checking program gives feedback, good paths are reinforced. This is called learning with verifiable reward. It is the reason why the computational abilities of new models have grown markedly faster than their reliability in domains without clear checkability.