Anthropic is managing its AI training capabilities for Claude through specific data handling protocols. These policies define how different types of user interactions are utilized for model development.
Text-based conversation training for consumer accounts transitioned to an opt-out model in September 2025. For users who participate in training programs, Anthropic maintains specific protocols for data handling and retention that differ from standard conversation logs.
According to Anthropic’s training policies, de-identified audio and its corresponding transcripts may be retained for up to five years when a user opts in for training purposes. This is a significant extension compared to the standard 30-day retention period applied to non-training data.
Training programs are currently limited to consumer-facing accounts, including Free, Pro, and Max plans. Conversations from Claude for Work, Enterprise, Education, and Government plans remain excluded from model training by default under current commercial terms. This exclusion for commercial and public sector clients applies regardless of whether the users are interacting via text or voice.
Data already processed under the five-year retention policy may remain in de-identified training sets. This distinction highlights the handling of data “donation,” as information requires de-identification processes before being reviewed by human contractors or used in model fine-tuning.
As Anthropic continues to develop its “Covered Models”—including the Mythos and Fable classes—specific classes often have stricter safety-related retention requirements that operate independently of general training settings.
