Questions about AI memorization evidence.
The questions litigators, retained experts and rights holders ask before an engagement, answered with the limits stated.
What does IPLYR do?
IPLYR prepares written forensic reports on whether a specific copyrighted work was memorized by a specific AI model.
Each report gives a binary finding, EXTRACTION CONFIRMED or NO EXTRACTION DETECTED, together with the record needed to reproduce it and a plain statement of what the finding does not show.
Who is IPLYR for?
Copyright litigators who need to show what a model retained, the independent experts they retain, and rights holders, publishers and estates working through counsel.
Teams preparing EU AI Act Article 53 training-data documentation use the same work in a regulatory setting.
What is memorization or extraction evidence?
Language models can reproduce passages of their training data, sometimes word for word, when prompted with part of a passage. Published research measures how likely a model is to reproduce a given text and how to tell genuine memorization apart from text a model could produce without having seen it.
When a model reproduces a long, distinctive passage of a work, that reproduction is direct evidence the work shaped the model. It is a different kind of evidence from a statistical score saying a text looks familiar to the model.
Have courts looked at this kind of evidence?
Yes, in a small number of decisions. In Getty Images v. Stability AI ([2025] EWHC 2863 (Ch)), the English High Court weighed expert extraction experiments and treated them as able to show that reproduction can happen, while not establishing how often it happens. In GEMA v. OpenAI (LG München I, 42 O 14139/24, 11 November 2025), the Munich Regional Court held that memorization in a model falls outside Germany's text-and-data-mining exception.
Neither decision ruled on IPLYR's work. IPLYR does not claim that any court has accepted one of its reports.
What do we receive?
A written report with the finding, the work and model tested, the prompts and the model's responses, and a section on limitations. The materials are delivered as a signed, hash-anchored bundle, so any later change to a file can be detected.
The full methodology brief is available to counsel and opposing experts under a mutual NDA.
What does EXTRACTION CONFIRMED prove, and what does it not?
It shows that the model reproduced the work under documented, repeatable conditions. That supports the inference that the work was in the model's training data.
It does not decide infringement, fair use, damages, how many copies were made, or where training took place. Those are legal questions for counsel and the court, and some turn on facts that only discovery can supply.
What does NO EXTRACTION DETECTED mean?
It means the model did not reproduce the work under the test. It is not proof that the work was never used in training.
Heavy paraphrase, refusal training and output filtering can all hide memorization that is present. Image and audio coverage is narrower than text. Reports say which of these apply.
Can a membership-inference score prove a model was trained on our work?
Not on its own. Membership-inference methods estimate whether a text looks like training data to the model, and published work argues that such a score cannot by itself prove training (Zhang et al., Membership Inference Attacks Cannot Prove that a Model Was Trained On Your Data, arXiv:2409.19798).
Reproduction of the work itself is the stronger form of evidence, and it is what the binary finding reports.
What is your accuracy or false-positive rate?
IPLYR does not publish a headline accuracy figure. No courtroom error rate for this kind of evidence has been independently established, and a single number would hide how much results depend on the model and the type of work.
Where a false-positive rate has been measured for the model and type of work in an engagement, the report states it and how it was measured. Where it has not been measured, the report says so in plain words instead of borrowing a figure from the literature.
Does this work on closed commercial models?
Models reachable through an API can be tested for reproduction. What can be concluded differs by model. For open models whose training data is public, whether a passage was in training can be checked directly. For closed models it usually cannot, and conclusions about training generally need material obtained in discovery.
Each report states which situation applies.
Will the report be admitted in court?
Admissibility is decided by the court. IPLYR's work is prepared to be examined under Federal Rule of Evidence 702 and the Daubert factors: whether the method can be tested, its peer-reviewed basis, its known or unknown error rate, and whether it can be reproduced.
IPLYR does not claim prior court acceptance and does not provide legal advice.
Who testifies about the findings?
Counsel's retained expert. Reports are built so that an independent expert can reproduce each measurement from the record and form their own opinion, rather than vouch for a result they cannot check.
What do you need from us, and what happens to it?
The work at issue, or the passages that matter, and which model or product to test. Testing a model hosted by a provider means sending prompts drawn from the work to that provider under its terms of service; tell us at intake if that is a concern.
Confidentiality and retention terms are agreed in writing for each engagement. Questions about them go to legal@iplyr.com.
How is an engagement priced?
Each engagement is scoped to the number of works and models and to what the matter needs. Published starting points are on the pricing page; a scoping call sets the actual figure.
Who is behind IPLYR?
IPLYR is founder-led. Namandeep Singh scopes and reviews every engagement and can be reached at namandeep@iplyr.com.