Exposure bias, in the context of artificial intelligence, denotes two distinct but related phenomena. The first relates to the psychological mere-exposure effect, whereby people show an increased preference for AI systems simply because they are repeatedly exposed to them. This familiarity can lead to uncritical acceptance of the technology and underestimation of its potential risks, such as algorithmic errors or ethical shortcomings. Trust and positive perception thus arise not from objective assessment but from mere habit.
The second, more technical sense is known as exposure bias proper and concerns a specific problem in training autoregressive models, for example language models for text generation. During training the model is exposed only to flawless, real data (so-called teacher forcing), where it learns to predict the next element from perfect context. When generating in real deployment, however, the model builds on its own previously generated outputs, which may contain errors. This discrepancy between the training and inference processes can lead to a gradual accumulation of errors and drift away from the desired output.
Put another way, “exposure bias” in artificial intelligence can be understood in two ways. First, it works much as it does when you come to like a song you often hear on the radio. The more you encounter an AI application, a voice assistant say, the more you may trust it and feel good about it, even though you know nothing about how it works. You get used to it and it strikes you as reliable simply because it is familiar.
Second, in the technical sense, picture a robot learning to write sentences. In “driving school” it is shown only correctly written texts so that it learns to add the next word. But when it is sent off to write on its own, it has to continue from words it wrote itself. If it makes a small mistake, the next word builds on it and the error can grow until the whole sentence loses its sense. So it is the difference between learning from a perfect model text and real composition, where the AI has to cope with its own imperfections.