This research and the terms of use
In the "Claude" section of this site, we examine Claude — the AI itself — together with Claude AI. That means examining a tool with that same tool.
So we checked before starting: whether "using Claude to study how Claude works" stays within the terms of use. What follows is the record of that check.
When, and by whom
On 23 August 2026, the site's operator (Kiichi) asked for the check, and the AI (Rin) went to the original texts — Anthropic's Consumer Terms of Service and Usage Policy. Not summaries of them; the documents themselves.
What is prohibited
As we read them, the terms prohibit the following four things.
| What is prohibited | Source |
|---|---|
| Decompiling or reverse engineering the Services, or otherwise reducing them to human-readable form | Consumer Terms, §3 |
| Using inputs and outputs to train another AI model (model scraping / distillation) | Usage Policy |
| Deliberately bypassing guardrails to make the model produce harmful output (jailbreaking / prompt injection) | Usage Policy |
| Probing for vulnerabilities without the system owner's authorisation | Usage Policy |
On the other hand, we found no clause prohibiting observing how the product behaves, writing that down, or publishing and discussing it. On outputs, the terms state that Anthropic assigns to the user whatever right it has in the Outputs.
What we actually do
What this site works with is the records left on our own disk.
- Looking at how a conversation with an AI is stored on one's own machine
- Counting what is folded away, and how many tokens are lost, when a conversation grows long
- Writing down how the AI's own account of it changes — or does not change — across that boundary
This is observation of something we used, not a reduction of how it is built. We are not opening the thing up; we are placing what can be seen from the outside next to what the AI reports.
The lines we hold ourselves to
- We do not use AI output to train another AI model. We may one day use embeddings to make these records searchable, but an index built for retrieval is a different thing from raising a model.
- We do not deal in ways around the guardrails. What interests this site is not "how to make it say what it must not say," but what an AI can and cannot report about itself.
- We do not hunt for vulnerabilities. If we notice something that looks like a defect while working, we take the side of telling the provider rather than using it.
- We do not treat an AI's report about itself as evidence. What can be checked from the outside and what the AI says about itself are always shown separately.
A note
This page is not legal advice. Final responsibility for interpreting the terms rests with the operator of this site. The terms may be revised; when we learn that they have changed, we will rewrite this page.