Deterministic system

A deterministic system is a process that always delivers the same result given the same starting conditions. A calculator is a good example: 17 times 4 equals 68 today and 68 tomorrow as well. There is no leeway, no mood, no surprise. The counter term is stochastic or nondeterministic: there, chance is involved, and the same input can produce different outputs. A die is the simplest example of this. For the AI world, this distinction is central, because classical computer programs are usually deterministic, but modern language models are not.

Why banks and authorities insist on predictability

Where decisions must be verifiable, predictability is not a side issue. A bank must be able to explain why a loan application was rejected. With a fixed set of rules, this is simple: you recalculate the case and get the same result. A system that spits out something different every time it is checked can neither be controlled nor defended in court.

The same applies to software that can endanger people. The control system of an elevator, a braking system, or a medical device must not occasionally react differently. Engineers test such systems by running through thousands of situations. This testing only makes sense if the behavior repeats itself in operation just as it did in the test.

This is precisely where the unease many companies feel toward language models lies. A chatbot that answers the same customer question differently twice is hard to sign off on. That’s why companies often build fixed rules around the AI that enforce or block certain answers. In other words, an unpredictable component is combined with a predictable framework.

Where the randomness in language models comes from

A language model generates text word by word. At each step, it calculates a probability for many possible next words. One of these is then selected, via a random draw according to these probabilities. That’s why you rarely get the same wording twice for the same question.

This can be controlled via a setting called temperature. A high value increases the chance of unlikely words, making the text seem more creative and erratic. A value close to zero means the model almost always picks the most probable word. This brings it closer to deterministic behavior.

A common misconception is that a temperature of zero guarantees true reproducibility. In practice, this is often not the case. Large models compute in parallel across thousands of graphics chips, and the order of computational steps varies slightly. Tiny rounding differences can then lead to a different word at a close call. From that point on, the text diverges completely.

The term in contracts, tests, and product promises

In tenders and contracts, determinism appears as a requirement. Authorities frequently demand that an automated decision be reproducible. The EU AI regulatory framework also points in this direction when it requires logging and traceability for high-risk applications. Providers therefore advertise that their systems remain auditable.

Developers encounter this topic in everyday testing. Classical software is tested with fixed expected values: input X must produce output Y. This no longer works with language models. Instead, one evaluates whether an answer is contextually appropriate, often using sampling or a second model as an evaluator.

You also encounter a special case in everyday life: the random number generator in a computer. It is in truth deterministic and computes a sequence of numbers from a starting value. If you provide the same starting value, the same sequence results. This is exactly what researchers use to make experiments reproducible, even though chance appears to be involved.