In an era where artificial intelligence is becoming an indispensable tool for every student and scientist, institutions are hitting an unexpected wall. It is not built from technical limitations, but from the need to protect data. According to a report by The Daily Pennsylvanian, the University of Penn decided to tackle this problem by creating a controlled environment that functions as a "safe harbor" for interacting with LLMs (large language models).
Why don't users want to use standard versions of ChatGPT and Claude?
The main reason is what is known as data leakage. If a regular user enters sensitive research, application code, or personal data into the free version of ChatGPT or Claude, this data may (unless explicitly turned off) be used for further model training. For a university or corporation, this is an unacceptable scenario that threatens both intellectual property and privacy.
This phenomenon is called Shadow AI — a situation where employees or students use unofficial tools for work because the official channels are too rigid. Penn's model seeks to "pull these users back" into a secure ecosystem by offering them the same quality of outputs, but with a guarantee that their data remains within a closed circle.
Battle of the giants: Claude vs. ChatGPT in the academic environment
The pilot program at Penn is not limited to a single model. The strategy lies in providing a choice between the two most significant players on the market, which is enormously beneficial for users.
Anthropic Claude: King of safety and context
Models from the Claude family (currently dominant versions in 2026 include Claude Sonnet 5 or Opus 4.8) are known for their emphasis on so-called Constitutional AI. This means the models are trained with a clear set of ethical principles, minimizing the risk of generating harmful content. For academics, Claude is often the preferred choice due to its ability to work with an enormous context window — it can "read" and analyze entire books or extensive scientific studies at once without losing the thread.
OpenAI ChatGPT: Universal multimodal power
On the other side stands ChatGPT, which excels in its versatility and integration into a broad ecosystem of tools. While Claude excels in nuanced writing and text analysis, ChatGPT often wins in logical tasks and programming thanks to its ability to use various tools (so-called agentic functions) in real time. In benchmark tests, these two models constantly trade first place, with the choice often depending on the specific task.
| Feature | Claude (Anthropic) | ChatGPT (OpenAI) |
|---|---|---|
| Main strength | Safety, long context | Multimodality, ecosystem |
| Response style | Humanistic, nuanced | Direct, analytical |
| Suitable for | Document analysis, writing | Programming, agents |
Impact on the Czech market and EU regulation
Although Penn is an American institution, its approach is extremely relevant for the Czech academic sphere and businesses alike. In the Czech Republic and across the EU, when implementing similar solutions we must take into account the strict regulations of GDPR and the newly implemented EU AI Act.
For Czech companies, this means that simply using public chatbots can be risky from a compliance standpoint. Implementing one's own "AI portal" (similar to Penn's) via an API interface is the only way for a Czech mid-sized company or university (such as Charles University or CTU) to legally and securely harness the power of these models. The good news is that both Claude and ChatGPT models already have excellent Czech language support, enabling seamless work with local context.
Pricing policy: How much does it cost?
