Are you wondering if using artificial intelligence daily to summarize your case files is crossing a legal line? It is, and this is a legal reality that is accelerating across Europe.
Beyond mere ethical recommendations, a landmark ruling in Munich in November 2025 established that processing protected texts with AI can constitute proven copyright infringement. In Switzerland and elsewhere, uploading documents from licensed databases to open models violates both copyright law and professional ethical obligations.
So, is daily legal AI breaking the law? On the agenda:
- What is the Munich II ruling that changed the game for European lawyers?
- Why is copy-pasting into a public chatbot no longer considered "personal use"?
- How do the license agreements for your legal databases implicitly prohibit generative AI?
- How is the commitment to professional confidentiality compromised by these unsecured tools?
Let's get started.
Does the Munich II ruling change the game for AI training?
The legal landscape shifted on November 11, 2025, with the Munich Regional Court's decision in the case of GEMA v. OpenAI. For the first time, a European court has established that training an artificial intelligence model on copyrighted works constitutes a criminal and civil offense. The Court rejected the argument that neural networks merely memorize "statistical patterns." Instead, it ruled that when content is permanently integrated into the model's parameters to be retrieved later, it constitutes reproduction within the meaning of German copyright law.
This decision is fundamental because it does not target technology companies alone. It applies whenever an end-user transfers protected content into a system designed to learn from it. The reasoning is relentless: if an AI can reproduce near-verbatim passages of a work, it means that work has been "copied" and stored in the model's memory. The Court characterized this process as permanent reproduction, dismissing any defense based on an analogy to the MP3 format, which simplifies data without retaining it in an identifiable form.
The direct implication for legal professionals is clear: uploading a legal commentary or case law annotation into a public chatbot is, under this judicial logic, equivalent to authorizing the reproduction and distribution of a protected work without a license. The excuse of technical "transformation" no longer holds up against the reality of how these models memorize data.
Is copy-pasting into a public chatbot copyright infringement?
The scene plays out daily without anyone worrying: a lawyer downloads a doctrinal commentary from their database, pastes it into ChatGPT with the prompt "summarize the key arguments," and proceeds to draft their brief. This practice, once considered pragmatic and harmless, is now identified as a potential copyright violation. The Munich Court specifically rejected the exception for temporary reproduction for analysis (Section 44b of the German Copyright Act) because the content is not merely analyzed and then deleted, but durably integrated into the model's weights.
If an AI program memorizes the exact terms of a text to reproduce them later, the user's action of feeding that system constitutes participation in illicit reproduction. The technical distinction between "compression" and "memorization" is no longer a valid defense: the model's ability to generate sequences of tokens recognizably similar to the original proves the copy. The lawyer who provides this raw material then becomes co-responsible for the illicit act.
This extends beyond the initial training of the AI by its developer. Every human interaction where protected content is ingested to be processed, summarized, or generated transforms the machine into a vector for unauthorized distribution. Copy-pasting is no longer a neutral act; it is an operation of integrating protected data into a proprietary or open system not licensed for that purpose.
Do database licenses prohibit generative AI?
Legal risk stems not only from general copyright law but also from the specific contracts governing your access to legal content. An increasing number of content platforms—whether for music, stock photography, or legal databases—are updating their terms and conditions to explicitly prohibit the use of their works in machine learning models. Publishers like Spotify have, as of September 2025, included clauses prohibiting the ingestion of their content by any AI.
This restriction logically and legally extends to providers of legal literature. Your subscriptions to databases or analysis services almost certainly include limiting clauses that cover only online consultation by a qualified human. These texts often specify that the content may not be used for training or "ingested" into an artificial intelligence model. Using an automatic summarization feature via an external tool is therefore a direct violation of the license agreement.
This "minor inaccuracy" in contractual terms is no longer anecdotal. It transforms a common practice into a contractual breach with serious consequences. By uploading an excerpt of legal doctrine to a third-party server, you are not only committing an ethical breach; you are violating the intellectual property and exclusive rights of the publisher whose content you are leasing.
Did the German Bar Association provide a warning?
Even before the Munich court ruling, the profession's regulatory authority had sounded the alarm. In July 2025, the German Federal Bar published the text 32/2025, an explicit recommendation addressed to all practitioners regarding the use of artificial intelligence. This document warned against uploading protected literature to third-party servers, emphasizing that the reproduction resulting from these transfers requires express authorization or a specific legal basis.
The recommendation is unequivocal: copyright requirements are essential when using AI systems. It specifies that any form of reproduction, particularly through uploading to AI provider servers, must be covered by a license or legal permission. This means that intuitive use without prior verification of rights is now non-compliant.
The fact that this warning received little initial attention does not diminish its current legal value. It demonstrates that the legal community itself recognizes the risk. Lawyers can no longer complain about the opacity of the law when a professional authority has already defined the boundaries of illegality in relation to artificial intelligence models.
Is professional privilege compromised by public AI?
Although this analysis focuses on copyright, the violation of professional privilege constitutes an immediate and distinct danger. Public artificial intelligence models are not designed to guarantee the confidentiality of sensitive data. When a lawyer submits case facts, party names, or defense strategies to a public AI, this information may become visible to the model's publisher or used to refine its algorithms.
This practice jeopardizes the very essence of the lawyer-client relationship. Professional privilege requires that information processing remain confined to the circle strictly necessary for defending the client's interests. Using an unsecured third-party tool, where data is potentially stored or processed outside of any Swiss legal control, constitutes a breach of this protection.
Case law and professional ethics agree on the duty of care regarding security. Ensuring that the tool used does not store or leak confidential information is a legal obligation. Using a public chatbot to process client data is equivalent to handing the keys to a case to unauthorized parties, nullifying the fundamental protection of professional privilege.
Can AI memory be erased by law?
An argument often raised to defend AI tools is the possibility of "asking to erase" or "forgetting" content. However, recent computer science research shows that this "erasure by forgetting" mechanism is illusory given the complexity of language models. Studies conducted by Stanford and Yale in January 2026 proved that it is possible to extract up to 95.8% of an entire book even after access has been blocked or the model has been updated.
This means that the model's memory is not simply "forgotten" but remains intrinsically present in the global parameters. Erasure requested by law or by a user does not remove the AI's ability to generate responses that reproduce the original content. Deletion is therefore neither effective nor legally guaranteed.
Consequently, once a document has been ingested, the copyright violation is perpetual. The risk does not disappear with time or a deletion request. It persists as long as the model remains capable of reproducing the protected works it has integrated, rendering any retroactive action ineffective in alleviating the legitimate concerns of rights holders.
Does the European approach favor innovation or protection?
The European legislative landscape is evolving rapidly to address these challenges, marking a shift away from the old "move fast and break things" approach. Now, the focus is on the necessity of licenses and strict compliance with intellectual property rights. Lawsuits against technology companies are multiplying, showing that the sector will no longer benefit from a legal vacuum.
The Munich Court has set a precedent that applies beyond German-speaking law. The logic adopted—that permanent memorization constitutes reproduction—can be easily transposed to Swiss and European jurisdictions, where copyright law is similar in its protection of intellectual works. The European Union is beginning to demand total transparency regarding training data.
There is no longer any room for ambiguity. Innovation should not come at the expense of legality. Practitioners must now adopt a proactive approach where every interaction with AI is validated by a robust legal framework, ensuring that the data processed is either free of rights or authorized in advance.
How does Silex ensure the legality of your research?
Faced with these major risks, the solution is not to give up on AI, but to use a tool designed for compliance. Silex, the Swiss legal AI, differs fundamentally from consumer-grade tools. Its architecture is based on proprietary technology that does not learn from your data and does not store any sensitive information to train future models.
The advantage of Silex lies in its ability to analyze your files and internal documents via Silo, without ever exposing them to external servers or storing them illegally. Unlike chatbots that copy and collect, Silex processes information in an isolated and secure environment, ensuring that confidentiality and copyright are respected.
This positioning allows lawyers to benefit from the power of artificial intelligence without fearing lawsuits for copyright infringement or breach of professional secrecy. Silex is the partner that provides human validation of results, ensuring full compliance with Swiss and international legal requirements.
Why is Swiss hosting crucial for data?
Server location is not a technical detail; it is a legal guarantee. International tools often store data in undefined clouds or under foreign jurisdictions with weak protections. Silex is hosted in Switzerland by Infomaniak, ensuring that all your data remains subject to Swiss law, one of the most rigorous in the world regarding professional secrecy and data protection.
This choice of local infrastructure ensures that your case information does not transit to jurisdictions where it could be subpoenaed by foreign authorities or used to train global models. The advantage is twofold: total legal security and strict compliance with the Federal Act on Data Protection (FADP).
This provides reassurance to both the firm and its clients, whose files are processed without the risk of leaks to uncontrolled third-party entities. Data sovereignty is at the heart of trust in a world where AI is becoming omnipresent.
Are Swiss firms adopting this new approach?
The Swiss market is already following this trend toward responsible and secure AI. Bordier & Cie, in particular, has chosen to integrate Silex into its workflows to enhance the quality and security of its legal research.
This adoption reflects a growing awareness: efficiency is no longer enough if it comes with legal risk. Leading firms now prioritize tools that guarantee data sovereignty and compliance with Swiss ethical standards, moving away from unverified, generic solutions.
This evolution shows that the use of AI is not a matter of technical choice, but of strategic positioning. Firms that secure their research strengthen their credibility and protect their long-term reputation by avoiding legal pitfalls that could slow down or compromise their practice.
Does Silex help avoid substantive errors in analysis?
The risk of infringement is not the only danger. Substantive errors, hallucinations, or misinterpretations of the law remain major pitfalls. Silex incorporates human validation that is rigorous for every generated response. The algorithm does not simply guess; it analyzes the file and cross-references information with a verified legal database, drastically reducing the risk of error.
This transforms AI into a reliable assistant rather than a potentially misleading autonomous source. The user retains final control over the produced content, validating every statement before using it in a brief or conclusion.
By combining algorithmic power with human expertise, Silex offers a secure alternative to raw models that produce plausible but incorrect results. This hybridization is the key to secure daily use.
What checklist should you follow before uploading a document to an AI?
Before interacting with any artificial intelligence tool, ask yourself these essential questions:
- Licensing: Does my database explicitly authorize the use of its content in external AI models?
- Security: Does the tool guarantee that my data will not be stored or used for future training?
- Sovereignty: Are the servers located in Switzerland and subject to Swiss law?
- Validation: Can I manually verify and validate every result before using it?
In short : If the answer to any of these points is no, refrain from uploading the document. Security and legality must take precedence over speed.
To learn more: Legal AI for individuals, Silex in WinLex and SkyLex, Silex and EXPERTsuisse.
