Who comes to take, and how you prove it
Neighbouring rights exist. A creator can object to their work being taken. But to object, you first have to establish the fact. And today, almost no creator can say who came to their site, when, or what they carried away.
That is the gap. Not the law: the proof. This page reports what can be measured when you decide to look, on a single server — the one serving this site.
What was measured today
Over the past twenty-four hours the doorkeeper logged 47,098 visits. 44,498 came from the house itself — the server talking to itself, our own robots, the office connection. The actual outside world accounts for 2,600 visits, five in a hundred. Any traffic figure that does not subtract your own is wrong by a factor of twenty. That is the first lesson, and it stings.
Among those 2,600, every forced door aimed at the same thing:
/.env, the file where a site keeps its passwords, and
/.git/config, the address of its source code. Fourteen attempts on
the first, eleven variants of the same file elsewhere, eight on the second.
None obtained anything.
Naming, not guessing
A visitor announces itself in a short line called a user agent. That line is free text: anyone can write anything in it. A robot claiming to be Googlebot proves nothing.
The reader installed here names twenty-six actors — search engines, AI crawlers, measurement tools, social networks. Among them: OpenAI, Anthropic, Perplexity, Meta, ByteDance, Common Crawl, Amazon, Apple, Google, Microsoft, Yandex, Baidu.
googlebot.com. A machine rented in the same
building ends in googleusercontent.com. Confusing the two means
letting anyone in — or accusing Google wrongly. Both mistakes were made here
before being corrected.
Some publishers list the addresses of their crawlers, which makes verification possible. Others do not. When a name has no official registry, the only honest answer is: "we can neither believe it nor accuse it". That is written as such in the table, rather than an invented verdict.
Recognising a machine that changes its name
Disguise is the real problem. A robot that gets refused comes back under another name, from another address, with another user agent. The name stops being useful.
But something remains that a program cannot easily change: the way it opens the encrypted conversation. The first words exchanged to establish a secure connection form a pattern — a handshake fingerprint. It follows the machine, not the disguise.
What this changes for a creator
Today, when an author suspects an AI model has fed on their work, they have nothing to show. No date, no trace, no name. They can neither negotiate, nor refuse, nor complain usefully. They do not know.
What is described here does not solve that. It is a server, not an institution. But it shows the observation is technically feasible, with free tools, on an ordinary machine, by one person. And that what is missing is not the technique: it is that nobody has yet put it in creators' hands.
What is not proven, and will not be claimed
Seeing a crawler pass does not prove a work was used to train a model. We measure a visit, not a use. The path between the two — which page, which model, which version — remains entirely open.
Likewise, a closed address may be shared by thousands of innocent people. That is why every closure carries its reason: without a reason, a blocklist is blind punishment.
One lesson from the same day, to illustrate: this server's firewall permanently closed access to a phone, and to a visitor who was simply reading a CV. Neither had done anything. The cause was a contradiction in our own configuration. A verification tool that does not watch itself produces false accusations — and that is the main danger in everything described here.
What comes next
This is an open project, in France, documented as it goes — including the mistakes, which are the useful part. The code, the method and the measurements are available.
If you work on protecting creators against AI models, or want the same measurement on your own sites, the door is open: get in touch.