Collected
A crawler copies the public page your work sits on. At this moment it still carries your name and terms.
Give it a handle, a paragraph you wrote, or a melody you made up. The simulator walks it through the stages a large AI platform can put creative work through, and ends with the one thing you never get: a receipt.
An example is already loaded so you can see a finished run. Replace it with your own, then press Run.
Six stages. The first three happen when data is gathered and prepared. The fourth is training. The last two happen every time someone asks the finished model for something.
A crawler copies the public page your work sits on. At this moment it still carries your name and terms.
Cleaning steps commonly keep the content and drop the surrounding page, where your name and terms live.
The content is cut into small units the model can count. Your work is now a list of fragments.
Training does not file your work in a drawer. It nudges millions of numbers so the model gets better at producing things like it. Gold dots are your fragments. Grey dots stand for everyone else's work. Watch the gold fade into the clusters.
A user types a prompt. The model produces new output shaped by what it absorbed. Below, a tiny model trained only on what you entered makes its own version, right here in your browser.
The output ships under the platform's terms, to a paying customer.
This is the Latent Space Loophole. Your work was used at every stage, and no stage produced a record that it happened.
No notice that your work was collected. No record of what it was used for. No line in the output that points back to you. If you later ask "was my work in there?", there may be no document anywhere that can answer.
The fingerprint is a real SHA-256 hash of what you entered, computed in this page. Change one letter above and it changes completely. That is what makes a record tamper-evident. Recording authorship at creation, then sealing it so it can be checked later, is the idea behind the first two floors of the Mirror Protocol.
A teaching tool that overstates its case teaches the wrong lesson. Here is where this one compresses or simplifies.
Collecting and training happen over weeks or months, across enormous datasets. The seconds part is real only at the end: once a model is trained, producing something in a style it absorbed takes seconds, every time anyone asks.
Data cleaning steps commonly extract the main text or audio and discard page furniture such as bylines, captions and license notices. Practices vary by company and dataset, and some datasets do keep source links.
Training generally adjusts numbers rather than storing a copy of your file. That is exactly why there is no record. Models can still sometimes reproduce pieces of training data closely, which researchers call memorization.
Your piece is one among billions, so its influence on any single output is small and hard to trace. That difficulty is the loophole: it makes "prove it was used" close to impossible for an individual.
Some companies license data, honor opt-out signals, or exclude certain sources. Whether training on unlicensed work is lawful is being argued in courts right now. This page does not decide that question; it shows the mechanism.
Stage 05 runs a small pattern model trained only on what you typed. It is far cruder than a commercial system, and it uses nothing but your words or notes, which is the point: everything recognizable in it came from you.
"The AI didn't just find my work. It dissolved my identity into its system without my receipt."
The fix is not to stop sharing. It is to make sure that when work is created, a record of who made it is created too, and sealed so it can be checked later.