You own Claude's outputs. You still can't train a competitor on them.

Anthropic has set out the split between owning Claude's Outputs and being allowed to train on them: customers own the Outputs generated from their Inputs, while the terms of service bar using those Outputs to train or develop AI models without written permission from the company. The policy is laid out in a Claude support article dated March 16, 2026.

At a glance

• The restriction runs through the terms of service rather than through ownership: rights to the text stay with the customer, while the permitted uses of that text are set by contract.
• Anthropic gives two safety reasons: models distilled from Claude's Outputs carry none of its pre-release testing or layered safeguards, and the company has no visibility into how such systems are later deployed.
• Non-competing training remains open, including specialized classifiers and tooling, as do product integrations, customer-facing content generation, data structuring and internal workflow improvements built on Claude's Outputs.

Why it matters. The dividing line here appears to be commercial rather than legal-theoretical: ownership of a text and the right to use that text as training data are separated, and the separation is enforced by contract terms. For teams building on the API, that likely turns an architecture question into a compliance one, since whether a downstream model counts as competitive is measured against Anthropic's own model line rather than against the builder's stated intent.

The document states that customers own the Outputs generated from their Inputs, and that Anthropic prohibits use of its services to train or develop AI models without written permission. It describes such restrictions as standard practice across the AI industry. Supporting a third party's attempt to do the same is also a violation of the terms.

Anthropic ties the rule to its own safety process: rigorous pre-release testing, multiple safety layers and continuous monitoring of model behaviour. Outputs used to train new models outside that oversight lose those controls, according to the article, which points to potentially harmful or dangerous systems and to distilled models the company cannot monitor once they are deployed.

The second stated reason is commercial. Where customers generate Outputs that then train competing models, Anthropic says they are using its infrastructure and investment to build direct competitors to its service, and adds that, like other software and service providers, it expects its services not to be used to undermine its product offerings.

The permitted uses are drawn around competition. Training models that do not compete with Anthropic's own remains allowed, with specialized classifiers and tools named as examples. Outputs may also be integrated into applications to power product features, generate content for customers, analyze and structure data, or improve internal workflows and productivity.

What's next. The page does not describe how written permission is requested or on what conditions it is granted, and it carries no effective date beyond the March 16, 2026 stamp. It introduces its lists of permitted tools and prohibited uses without enumerating the individual items in the published text. Adjacent documentation covers Covered Models and the data retention practices that apply to them.