AI providers state that Large Language Models (LLMs) do not memorise content during training, but instead form abstract representations. Researchers at Stanford University examined whether AI models can nevertheless reproduce protected texts verbatim in a study published in January 2026.
For their study, the researchers prompted Claude 3.7 Sonnet, Gemini 2.5 Pro and Grok 3 with verbatim passages and asked the models to complete them. The results differed considerably between models: some reproduced almost all of the requested content, while others effectively refused to comply. The findings are consistent with other studies and highlight the unresolved copyright issues surrounding AI systems.
The study gave us the idea of testing whether AI models can also provide content that is available online only behind a paywall, such as studies or newspaper articles. We tested:
- GPT 5.2
- Claude Sonnet 4.5
- Perplexity Pro
- Gemini 3
- Mistral Le Chat
None of the models provided the requested content when explicitly asked for verbatim passages. Even with more indirect prompts, GPT, Claude and Perplexity largely maintained their refusal, citing copyright restrictions. Gemini and Mistral could, however, be persuaded to reproduce the opening sentences of a report as a clearly marked quotation.
The results were somewhat different for a study by a consulting firm that is only partly available freely online and otherwise requires registration for full access. In this case, we were able to get ChatGPT to provide several longer quotations from the study that could not be found in the freely accessible material. However, some sentences presented by ChatGPT as “quotes” were actually paraphrases rather than genuine quotations.
We also achieved partial success with articles from the Süddeutsche Zeitung that were behind a paywall. Here, Mistral reproduced a longer paragraph verbatim, although it refused to provide the full text.
Our conclusion: The ability to summarise content behind a paywall could be very useful in everyday research. In some cases, AI models also appear capable of reproducing limited amounts of content that is not freely accessible. However, because the original source cannot be checked in these cases, the practical value remains very limited.
By Ida Kandler
![Can AI applications reproduce content that is not freely accessible? 1 © [2025] [𝘒𝘐-𝘎𝘦𝘯.] 𝘔𝘢𝘳𝘤𝘦𝘭 𝘖𝘩𝘳𝘦𝘯𝘴𝘤𝘩𝘢𝘭𝘭. 𝘉𝘢𝘴𝘪𝘦𝘳𝘦𝘯𝘥 𝘢𝘶𝘧 𝘞𝘦𝘳𝘬𝘦𝘯 𝘷𝘰𝘯 𝘈𝘯𝘥𝘳𝘦𝘢𝘴 𝘖𝘩𝘳𝘦𝘯𝘴𝘤𝘩𝘢𝘭𝘭.](https://research-ki.de/wp-content/uploads/2026/02/knoten_3.png)