noun
Model drift is the gradual degradation of an AI model’s accuracy as the data it was trained on no longer matches the real-world conditions it encounters. This happens because the world evolves—user behavior shifts, language changes, or new trends emerge—and the model fails to adapt.
It typically arises in large language models (LLMs) or image generators when the training data becomes outdated. For example, a chatbot trained in 2020 might struggle with slang from 2024, or an image generator might produce outdated fashion styles if its training data is old.
For developers, model drift means constant monitoring and retraining. It forces teams to allocate resources for updates, which can be costly. A real-world case: a 2023 study found that a popular language model’s accuracy dropped by 15% over two years without retraining.
A customer service AI trained in 2021 starts giving incorrect answers in 2024 because new product features were never added to its training data.