External Model Deprecation

External models used in AI systems can be deprecated, altered, or taken offline by providers and removed, which can result in degraded performance, unexpected behavior, or service outages. When designing AI systems and choosing models to use, it is important to review their policies on deprecation and design backup plans.

This risk is highest when models are accessed through APIs or cloud platforms and the organization does not control the infrastructure or model lifecycle. Some providers give notice before deprecation, but abrupt changes are also possible. AI teams must account for these risks during design by planning contingencies such as self-hosted alternatives, fallback models, or ensemble strategies to reduce dependency on any single external model.

- OpenAI Model Deprecation: text-davinci-003 is not working text rephrase and text spelling correction - As of July 6, 2023, the following models were no longer available for new deployments. However, existing deployments created before this date remained accessible to customers until July 5, 2024. Customers using these models were required to migrate to replacement models before the July 5, 2024, retirement deadline. After this date, any services or applications still relying on the retired models ceased to function.