What AI vendors commit to, in their own words.

Usage Policy (effective September 15, 2025)

Anthropic may not train models on Customer Content (Inputs + Outputs) under Commercial Terms

Customer retains all rights to Inputs; owns Outputs

Pending verification (Claude Enterprise terms)

API: 30 days for safety. Enterprise: 0 days available.

At least 60 days notice before retirement for publicly released models

Responsible Scaling Policy v3.3 (effective May 26, 2026)

AWS Acceptable Use Policy

Customer data in Bedrock not used to train base foundation models

Customer owns Outputs

Uncapped IP indemnity for covered Bedrock services

Not retained beyond inference unless customer enables logging

Ad-hoc per underlying model provider. Bedrock relays provider deprecation.

Relies on underlying model providers (Anthropic RSP, Meta, etc.)

Llama Acceptable Use Policy

Open-weights distribution. Llama license does not require Meta to train on customer data; in-product Meta AI has separate terms.

Customer / licensee owns Outputs subject to Llama Community License terms

No standalone indemnity program for Llama itself. Some distributors (AWS Bedrock, Databricks) provide their own.

Not applicable for open-weights local deployment. Hosted Meta AI separate terms.

Ad-hoc per Llama version release. No formal cross-version commitment.

Frontier AI Framework v1.1