AI Drift

Many computer programs today learn from examples instead of fixed rules. You show them, for instance, a million old bank transfers and mark which of them were fraudulent. The program searches for patterns in this data and applies them to new cases. Such patterns, however, always come from the past. If reality changes, they eventually stop fitting. This exact slow deterioration is called AI Drift.

The silent loss of quality

The tricky thing about drift is that nothing actually breaks. There is no error message and no crash. The system keeps responding, just as fast as before, and sounds just as confident. Only the answers are correct less and less often. Without targeted monitoring, this often only becomes apparent after months.

How costly this becomes depends on the field of application. With a music recommendation, a bad suggestion is annoying, nothing more. With fraud detection in payment transactions, real damage slips through. With a medical evaluation or a credit decision, drift can directly disadvantage people. That is why regulators increasingly require operators to monitor their systems continuously, not just check them once at launch.

A well-known example is the coronavirus period. Models that predicted purchasing behavior in retail were practically worthless in spring 2020. They only knew a world without lockdowns. Such disruptions are rare, but they illustrate the principle particularly clearly.

When the data or the rules change

Experts distinguish between two causes. The first is called Data Drift: the inputs look different than before. A voice assistant is suddenly asked about terms that did not exist during training. A camera in a factory experiences worse lighting because someone replaced the lighting fixtures. The task has remained the same, only the material is new.

The second cause is called Concept Drift. Here, the relationship itself changes. A feature that used to reliably indicate fraud is now completely normal. Fraudsters adapt their methods as soon as a defense system recognizes them. The model has then learned a rule that no longer holds true in reality.

The only remedy for both is measurement. Operators continuously compare how current inputs are distributed compared to the training data. In addition, they check samples to see how often the predictions actually turn out to be correct. If the gap becomes too large, the model is retrained with fresh data. Some systems continue learning automatically, but this brings new risks: if a model learns from its own incorrect answers, it reinforces its errors.

Drift in products and in headlines

You most often encounter drift without anyone using the word. The spam filter lets new tricks through. The search function in an online shop suggests things nobody is looking for anymore. A translation app doesn’t know current expressions. All of these are symptoms of the same problem.

With large chatbots, the term is often used incorrectly. Users report that a model has become dumber over the course of weeks. However, the stored values within the model do not change on their own. Usually, the provider has deployed a leaner version or rewritten the instructions in the background. Genuine drift, by contrast, arises without any action by the provider, purely through the changed environment.

In financial news, drift mainly comes up in the context of model risk. Banks and insurers must document how often they review their systems. The EU’s AI Act also requires post-market monitoring for high-risk applications. A model is thus not considered a finished product, but something that must be continuously maintained.