What model providers' usage rules say about training on…
3 model providers' own usage documents restrict using what their models produce to train another AI model, each rule quoted from the provider's page, read 28 October 2026.
Using its outputs to train another AI model
The rule, in its own words
Each model provider, answer by answer
A model provider per row. Every answer opens onto the line it comes from, quoted exactly, with the usage document it sits in and the day of the read. "Not stated" marks a question the usage document leaves unanswered.
| Model provider | Using its outputs to train another AI model | The rule, in its own words |
|---|---|---|
| Anthropic | barred unless the provider authorises it first | Use outputs to train an AI model (e.g., “model scraping” or “model distillation”) without… |
| OpenAI | varies by document | These changes are effective November 20, 2026. |
| xAI | barred where the model or product competes with the provider | Scraping, harvesting or reselling any Input or Output, or distilling model data or Outputs |
The 8 sentences the answers come from
- Anthropic, usage policy, read 2026-10-28, Using its outputs to train another AI model:
Use outputs to train an AI model (e.g., “model scraping” or “model distillation”) without prior authorization from Anthropic
- Anthropic, usage policy, read 2026-10-28, The rule, in its own words:
Use outputs to train an AI model (e.g., “model scraping” or “model distillation”) without prior authorization from Anthropic
- OpenAI, terms of use for the UK and Europe, read 2026-10-28, Using its outputs to train another AI model:
Using Output to develop models that compete with OpenAI.
- OpenAI, terms of use for the UK and Europe, read 2026-10-28, The rule, in its own words:
Using Output to develop models that compete with OpenAI.
- OpenAI, usage policies, read 2026-10-28, Using its outputs to train another AI model:
You must not use OpenAI models or their outputs to generate, prepare, select, or supply training material or signals for models that are intended as alternatives to or substitutes for OpenAI models or services.
- OpenAI, usage policies, read 2026-10-28, The rule, in its own words:
These changes are effective November 20, 2026.
- xAI, acceptable use policy, read 2026-10-28, Using its outputs to train another AI model:
Using the Service or any Output to develop (or assist anyone in developing) machine learning models or any products or services that compete with SpaceXSI, whether directly or indirectly
- xAI, acceptable use policy, read 2026-10-28, The rule, in its own words:
Scraping, harvesting or reselling any Input or Output, or distilling model data or Outputs
Method
Each model provider's own model providers' own usage policies, acceptable use policies and regional terms of use were read through estate-fetch and stored whole (sha256 on every row).
For every usage document, two questions were answered from its own words: using its outputs to train another AI model, the rule, in its own words.
Every stated answer carries the sentence it rests on, checked against the stored text; 'not stated' means the usage document read does not address the question.
A model provider with several usage documents reads 'varies' where they disagree.
3 model providers, 4 usage documents.
The 4 pages we read
- OpenAI, usage policies (from the publisher's own model providers' own usage policies, acceptable use policies and regional terms of use; reproduced for comment)
- OpenAI, terms of use for the UK and Europe (from the publisher's own model providers' own usage policies, acceptable use policies and regional terms of use; reproduced for comment)
- Anthropic, usage policy (from the publisher's own model providers' own usage policies, acceptable use policies and regional terms of use; reproduced for comment)
- xAI, acceptable use policy (from the publisher's own model providers' own usage policies, acceptable use policies and regional terms of use; reproduced for comment)
Quote this figure or put it on your site
3 model providers' own usage documents restrict using what their models produce to train another AI model, each rule quoted from the provider's page, read 28 October 2026.
Using its outputs to train another AI model, by model provider: barred unless the provider authorises it first 1 of 3, barred where the model or product competes with the provider 1 of 3, varies 1 of 3 (3 model providers, read 28 October 2026).
The rule, in its own words, by model provider: stated 3 of 3 (3 model providers, read 28 October 2026).
Citation: "What model providers' usage rules say about training on their output", updated 2026-10-28, https://sellagentlogstoai.com/datasets/agent-model-output-rules/.
Download as CSV (8 answers as of 2026-10-28)