Subprocessors
Where your content goes when a measurement runs
Measuring how AI models answer means sending questions to those models, so the providers below necessarily see some of the content involved. A service that measures what other systems say about you should be plain about which systems it feeds, and the set keeps growing as new model families arrive.
Last updated 29 September 2026
The list
Who processes what
| Provider | What it receives | Where |
|---|---|---|
| Anthropic | Question text, and the answer under assessment when it grades another model | United States |
| OpenAI | Question text, and the answer under assessment when it grades another model | United States |
| Question text | United States | |
| xAI | Question text | United States |
| Brave | Search queries, for measurements that check what is publicly findable | United States |
Anthropic and OpenAI both answer questions and grade answers given by other providers, since a grader has to come from a different vendor than the model it is grading. The remaining providers answer questions without grading anything.
Providers are added steadily as new model families become worth measuring, so treat this as the current set of major providers rather than a fixed one. Which of them run for a given measurement depends on the panel chosen for it, and your report records exactly which models produced which answers.
Run locally
Models that run on our own machines
Some measurements also use open-weight models running on hardware we operate. Content sent to those does not reach the organisations that published the weights, and does not leave the machine, and a report says when local models were used.
This website
The service that runs this website
Cloudflare, a United States company operating a global network, hosts this website and handles its join form. It stores each join request in a Cloudflare database, emails a copy of it to us, and runs Turnstile, the check that a request came from a person. It receives what you enter in the form, and your IP address and browser signals for the Turnstile check. Measurements are not sent to Cloudflare.
What is sent
The content that actually leaves
Two kinds of content travel to providers. The first is the benchmark question, which is a question a person might plausibly ask an assistant and usually contains nothing confidential. The second, for runs that supply evidence to a model, is text taken from publicly available pages on the site being measured.
Credentials never travel, because none are collected in the first place. Nothing behind a login is read, so nothing behind a login is sent.
Their terms
How providers handle it
Each provider processes content under its own terms and privacy policy, and their retention and training practices are theirs to set rather than ours. We use business or API tiers where those offer stronger handling commitments than consumer tiers, though we cannot warrant another company’s practices.
Given that, please do not send us confidential material you would not want processed by the providers listed here. The privacy policy covers the rest of how information is handled.
Changes
Notice of additions
New providers are added regularly, because the set of model families worth measuring keeps growing. This page carries the current major providers and the date it last changed, and material additions are notified to beta participants whose contact details we hold.
Questions can be sent through the Anphase website.