RAI (Responsible AI)

RAI is the University of Stuttgart's AI tool that complies with data protection regulations and can be safely used in both everyday work and teaching. This information page brings together all key details regarding features, access, data protection, and potential uses, and helps employees use RAI safely and effectively at the university.

RAI is particularly helpful with text-based tasks, such as writing, reorganizing, summarizing, translating, or developing ideas. RAI is accessible only on the university network or via VPN. 

Depending on the model, RAI is operated by GWDG in Göttingen, via Microsoft Azure in the EU, or via Microsoft Azure outside the EU. User input is not used to train the AI. RAI may also be required in courses, where appropriate. There is also a group chat feature available.

Video tutorial: what is RAI?

05:09

Note: The video shows the interface of the previous version. Some controls look different in RAI 2.0.

Sample prompts to get you started

RAI can assist you in your daily work – whether as a voice assistant, a translation tool, or for quickly transcribing long documents. The following sample prompts illustrate specific use cases for different target audiences. Important: Always critically review AI-generated results to ensure their accuracy. AI models can generate content that sounds plausible but is actually incorrect (so-called “hallucinations”). Before entering any data, be sure to check its confidentiality rating (C1–C4).

“Please edit the following text for style: Improve the grammar, sentence structure, and clarity without altering the content, my argument, or the tone. Mark the parts you have changed. [Insert your own text]"

Key benefit: The content remains yours – RAI just helps with the finishing touches. This is comparable to proofreading and is one of the least critical yet most useful applications of AI.
Model recommendation: For public or internal texts (C1–C2), GPT-6 Luna via Azure (EU) is sufficient. If the text contains confidential content (C3, e.g. committee decisions, personnel data), use Gemma 4 or DeepSeek V4 Flash 0731 via GWDG. RAI is not suitable for highly confidential information (C4).

“Translate the following German text into academic English. Keep the technical terms as they are and mark any places where you are unsure whether the technical term has been translated correctly. [Insert your own text]"

Key benefit: Scientific translations are time-consuming, and DeepL has its limitations when it comes to technical terminology. RAI can provide a solid initial draft, which you can then review for technical accuracy.
Model recommendation: For public or internal texts (C1–C2): GPT-6.1 Sol via Azure (EU). For confidential texts (C3, e.g. unpublished research texts): Gemma 4 via GWDG (supports over 140 languages). RAI is not suitable for highly confidential information (C4).

“Summarize the following document in no more than 10 sentences. Focus on: key points, agreed-upon actions, and outstanding issues. [Insert document]"

Key benefit: In everyday work, minutes, guidelines, or reports need to be recorded quickly. RAI saves time during the initial review – but the summary should always be checked against the original.
Model recommendation: For public or internal documents (C1–C2): GPT-6 Luna via Azure (EU) (fast and resource-efficient). For confidential documents (C3, e.g. committee meetings, personnel matters): DeepSeek V4 Flash 0731 via GWDG – with around one million tokens, it can process the largest volumes of text. RAI is not suitable for highly confidential information (C4).

“Rewrite the following draft email to [target audience, e.g., students / international partners] so that the tone is friendly and clear, while retaining all the relevant information. [Insert your draft text]"

Why this is a good idea: The message comes from you – RAI simply helps you craft it in a way that is tailored to your target audience and easy to understand. This is especially helpful in multilingual communication or when it's difficult to strike the right tone.
Model recommendation: If the email contains only public information (C1): any model is suitable. If it contains internal content without sensitive personal data (C2): GPT-6 Luna via Azure (EU). If it contains names with personal feedback or sensitive personnel matters (C3): Gemma 4 or DeepSeek V4 Flash 0731 via GWDG. RAI is not suitable for highly confidential information (C4).

“I get the following error when running my Python script: [insert error message]. Here is the relevant code snippet: [Insert code]. Please explain what is causing the error and how to fix it.”

Key benefit: Interpreting error messages is a clearly defined task where AI has already demonstrated its effectiveness.” The learning effect is preserved because the explanation promotes understanding rather than simply spoon-feeding the solution.
Model recommendation: Source code is usually public or internal (C1–C2): GPT-6.1 Sol via Azure (EU) is particularly strong here. If the code contains confidential research logic (C3): GLM 5.3 Flash or Qwen 3.8 via GWDG, both with strong programming capabilities. RAI is not suitable for highly confidential information (C4).

“Please rewrite the following text in simple language (level B1) that is easy to understand for international students and non-native speakers. The content must remain in its entirety. [Insert text]"

Key benefit: Accessibility and clarity are essential at an international university. RAI can help make existing texts more accessible. This is a clear benefit that doesn't harm anyone.
Model recommendation: For public or internal texts (C1–C2, e.g. website texts, internal circulars): GPT-6 Luna via Azure (EU). For confidential texts (C3, e.g. examination regulations): Gemma 4 or DeepSeek V4 Flash 0731 via GWDG. RAI is not suitable for highly confidential information (C4).

Background: HAWKI and GWDG

HAWKI is an open-source platform for generative AI that is used at many universities. HAWKI was developed with the aim of providing university staff with low-threshold, pedagogically supported access to generative AI and fostering discussion about meaningful ways to use it. RAI uses a HAWKI 2 user interface.

RAI offers three model backends: OpenAI models via Microsoft Azure in the EU, Anthropic's Claude models via Microsoft Azure outside the EU, and open-weight models hosted on servers operated by the Gesellschaft für wissenschaftliche Datenverarbeitung mbH Göttingen (GWDG).

Models in RAI and selection guide

RAI offers several AI models. The plans differ in terms of performance, data limits, and hosting.

  • Azure (EU): OpenAI models running on Microsoft Azure servers in the EU (permitted for data classes C1–C2). Input data is stored by Microsoft for 30 days.
  • Azure (outside the EU): Anthropic's Claude models running on Microsoft Azure servers outside the EU (permitted for data class C1 only). Input data is stored by Microsoft for 30 days.
  • GWDG (Göttingen): Open-weight models running on GWDG servers (permitted for data classes C1–C3; documented approval is required for C3).

Which information may be entered into which system depends on the confidentiality level and the type of system. If in doubt, err on the side of caution: until the matter is clarified, use only public content (C1). The AI systems available in the portfolio are generally not suitable for highly confidential information (C4).

Note on models of Chinese origin (Qwen, DeepSeek, GLM): These models also run exclusively on GWDG servers in Göttingen. Your input is not passed on to the developers. However, their training may give the models a political slant. On politically or socially sensitive topics, answers may therefore be one-sided or incomplete. Please review such answers particularly critically.

The following overview (as of November 4, 2026) will help you choose the model that best suits your needs.

The top model of the GPT-6 family for particularly demanding, multi-step tasks. It is considerably more resource-intensive than GPT-6.1 Sol and GPT-6 Luna, so please use it selectively.

Details:

  • Developer / Country: OpenAI / USA
  • Hosting: Microsoft Azure (EU)
  • Permitted data classes: C1-C2
  • Image input: yes
  • Context: 272,000 tokens (approx. 550 pages of text)
  • Strengths: very demanding analyses, programming, mathematics and logical reasoning

Balanced general-purpose model of the GPT-6 family and successor to GPT-6 Sol. A good choice for demanding tasks; according to OpenAI, it makes significantly fewer factual errors than previous versions.

Details:

  • Developer / Country: OpenAI / USA
  • Hosting: Microsoft Azure (EU)
  • Permitted data classes: C1-C2
  • Image input: yes
  • Context: 272,000 tokens (approx. 550 pages of text)
  • Strengths: demanding text tasks, programming, working thoroughly through complex questions

The fast and resource-efficient tier of the GPT-6 family and successor to GPT-5.6 Luna. The right choice for most everyday tasks.

Details:

  • Developer / Country: OpenAI / USA
  • Hosting: Microsoft Azure (EU)
  • Permitted data classes: C1-C2
  • Image input: yes
  • Context: 272,000 tokens (approx. 550 pages of text)
  • Strengths: fast responses, everyday text and communication tasks, summarising and extracting information, resource-efficient

The most powerful Claude model in RAI. For complex, open-ended tasks that require careful judgement. It is more resource-intensive than Claude Sonnet 5.5 and Claude Haiku 5.5. Permitted for public data (C1) only.

Details:

  • Developer / Country: Anthropic / USA
  • Hosting: Microsoft Azure (outside the EU)
  • Permitted data classes: C1 only
  • Image input: yes
  • Context: 1,000,000 tokens (approx. 2,000 pages of text)
  • Strengths: complex and open-ended tasks, demanding texts, programming, thorough analyses

Faster, more resource-efficient general-purpose model of the Claude 5.5 family for everyday work. Permitted for public data (C1) only.

Details:

  • Developer / Country: Anthropic / USA
  • Hosting: Microsoft Azure (outside the EU)
  • Permitted data classes: C1 only
  • Image input: yes
  • Context: 1,000,000 tokens (approx. 2,000 pages of text)
  • Strengths: everyday text and analysis tasks, programming, working with documents and spreadsheets

The fastest and most resource-efficient model of the Claude 5.5 family. Permitted for public data (C1) only.

Details:

  • Developer / Country: Anthropic / USA
  • Hosting: Microsoft Azure (outside the EU)
  • Permitted data classes: C1 only
  • Image input: yes
  • Context: 1,000,000 tokens (approx. 2,000 pages of text)
  • Strengths: fast responses, summarising, classifying and extracting information

Open model from Google based on Gemini technology. Understands text and images and covers over 140 languages.

Details:

  • Developer / Country: Google / USA
  • Hosting: GWDG (Göttingen)
  • Permitted data classes: C1-C2 as well as C3 with documented approval
  • Image input: yes
  • Context: 256,000 tokens (approx. 500 pages of text)
  • Strengths: multilingual tasks (over 140 languages), combining long documents and figures, solid all-rounder
  • As of: 04/2026

Compact open-weight model from Alibaba's Qwen 3.8 series with adjustable reasoning effort. In addition to text, it also understands images.

Details:

  • Developer / Country: Alibaba Cloud / China
  • Hosting: GWDG (Göttingen)
  • Data classes: C1-C2 as well as C3 with documented approval
  • Image input: yes
  • Context: 256,000 tokens (approx. 500 pages of text)
  • Strengths: research, programming, multi-step tasks, image analysis
  • As of: 08/2026

Low-cost variant of the GLM 5.3 series and the first model in the series to process images natively.

Details:

  • Developer / Country: Z.ai / Singapore (parent company Zhipu AI, China)
  • Hosting: GWDG (Göttingen)
  • Data classes: C1-C2 as well as C3 with documented approval
  • Image input: yes
  • Context: 1,000,000 tokens (approx. 2,000 pages of text)
  • Strengths: programming, multi-step tasks, very large amounts of text, image analysis
  • As of: 09/2026

High-throughput model with a very large context window. Text-only; image input is not supported.

Details:

  • Developer / Country: DeepSeek / China
  • Hosting: GWDG (Göttingen)
  • Data classes: C1-C2 as well as C3 with documented approval
  • Image input: no
  • Context: 1,000,000 tokens (approx. 2,000 pages of text)
  • Strengths: very long documents, fast processing of large volumes of text, general text and programming tasks
  • As of: 07/2026

Efficient open-weight model from Alibaba's Qwen 3.6 series: of 35 billion parameters in total, only 3 billion are active per request. In addition to text, it also understands images.

Details:

  • Developer / Country: Alibaba Cloud / China
  • Hosting: GWDG (Göttingen)
  • Data classes: C1-C2 as well as C3 with documented approval
  • Image input: yes
  • Context: 256,000 tokens (approx. 500 pages of text)
  • Strengths: fast responses, programming, multi-step tasks, image analysis
  • As of: 04/2026

Omni-modal model that understands text, images, speech and video. Its strength lies in the breadth of modalities, not in raw performance.

Details:

  • Developer / Country: Alibaba Cloud / China
  • Hosting: GWDG (Göttingen)
  • Data classes: C1-C2 as well as C3 with documented approval
  • Image input: yes
  • Context: 256,000 tokens (approx. 500 pages of text)
  • Strengths: understanding images and videos, broad coverage of modalities
  • As of: 09/2025

Guidelines for responsible use

Confidentiality

Check the confidentiality class and select the appropriate hosting option (Azure EU: C1–C2; Azure outside the EU: C1 only; GWDG: C1–C3, C3 only with documented approval).

Disclosure requirements

AI-assisted content must be disclosed if an AI system significantly contributes to the creation of content or if content is published without further human editing or quality assurance (e.g., website content, official correspondence, emails to external recipients). Additional information on transparency obligations

Training courses

Complete the mandatory training before using the system for work purposes. More information on the mandatory training

Report errors

In the event of a mistake (wrong system or wrong data entered), report the incident to Data Protection and Information Security as soon as possible. More information on what to do in the event of a mistake

Training, tutorials, videos

The AI learning portal in ILIAS offers video tutorials on RAI and the training courses. Training is mandatory before using AI systems for work purposes. You can find out which courses are eligible and how to submit the certificate in the Framework for the use of AI.

Frequently asked questions about RAI

Yes. RAI is accessible only on the university network or via VPN. If you are off campus, you must first establish a VPN connection. You can download the VPN client here. Instructions can be found on the TIK website.

After logging in, the model currently in use is shown in the input area. Clicking on it opens the list of models. The models are labelled by hosting: “Azure-EU” (Azure in the EU), “Azure-Other” (Azure outside the EU) and “GWDG”. The square brackets show the highest permitted data class for each model, e.g. “[≤ C2–Intern]”. You can find a visual guide in the RAI tutorials.

That depends on your use case and how confidential your data is. For confidential data (up to C3), use the GWDG models. For general tasks with internal data (up to C2), the OpenAI models via Azure (EU) often offer higher quality and speed. For exclusively public data (C1), the Claude models via Azure outside the EU are also available. You can find a detailed overview in the section “Models in RAI and selection guide” on this page.

That depends on the confidentiality class. Azure outside the EU is intended for C1 only, Azure (EU) for C1–C2. The GWDG option is intended for C1–C3; documented approval is required for C3. The AI systems available in the portfolio are generally not suitable for C4. Overview of confidentiality levels

The use of this service is free for all university members. However, depending on the model, the university incurs costs per unit of use. Therefore, please use RAI responsibly and in a manner that conserves resources, in line with the principle of fair use.

That depends on the confidentiality level. Azure (EU) is intended for C1–C2. The GWDG variant is intended for C1–C3; documented approval is required for C3. The AI systems available in the portfolio are generally not suitable for C4. Overview of confidentiality levels

No. Input data is not used for training in RAI – neither for the GWDG models nor for the Azure/OpenAI models.

In the Azure (EU) version, input data is stored by Microsoft for 30 days. For GWDG-supported models, operations take place on GWDG servers. Prompts and chat content are not stored there.

First, check what type of data is affected. If personal or confidential data (C2 or higher) has been entered into an unsuitable system, promptly notify your supervisor and report the incident to the Data Protection Officer and the Information Security Office. Reporting an incident promptly helps limit risks and ensure it is handled properly. For purely public content (C1), notification is generally not required.

Your chats in RAI are stored with client-side encryption. When you log in for the first time, you will be prompted to create a personal key. Choose a password that is easy to remember. You will be required to enter it manually on other devices or in private browsing mode. You can reset the key, but doing so will result in the loss of all saved chats.

Yes. Mandatory training is required for the professional use of AI systems – including RAI. It covers the basics of confidentiality classes, system types, and the university’s AI policy, and raises awareness of issues such as bias, discrimination, and hallucinations, among others. A certificate will be issued upon successful completion of the course, which must be presented to your supervisor before AI can be used for work purposes. You can find the training courses on the AI Learning Portal in ILIAS. One of the reasons for this is the EU AI Regulation (Art. 4), which establishes minimum standards for AI literacy for all employees who work with AI systems.

No. RAI is a secure all-round AI system for working with text, for example for ideas, rephrasing, translations and analyses within the scope of the guideline. It does not replace specialist research, special analyses or databases. For web-based research, you can use other approved tools from the whitelist.

RAI supports the understanding and analysis of images and PDFs (Vision) in select models. Image generation is currently not available in RAI. You can use additional approved tools from the whitelist to generate images of for other specialized tasks.

Yes, RAI offers a chat export feature. Occasional formatting issues (e.g., truncated lines) may currently occur during PDF export. A solution is being worked on.

Personal API access is currently unavailable. The platform's capacity is limited, and custom integration (e.g., with VS Code or your own applications) is not currently possible. We are aware that there is high demand for API access – this is a top priority for us, and we are actively working on it. As soon as API access becomes available, an announcement will be made.

RAI works best in Chrome and Firefox. Display issues may occasionally occur in Safari. Refreshing the page (page reload) often helps as a workaround. If the problem persists, try using a different browser.

Models with a pronounced reasoning component in particular (e.g. GLM 5.3 Flash, Qwen 3.8) require more computing time and can cause timeouts with complex requests. In this case, try to make your prompt shorter or less complex, or switch to an all-round model such as GPT-5.6 Luna. If the problem occurs repeatedly, contact RAI support.

That is a well-known problem. If a request is still in progress and a new chat is opened at the same time, replies may appear in the wrong chat, or display errors may occur. If this is the case, close the chat and start a new one.

 

RAI Support

 

Datenschutz

 

Stabsstelle Informationssicherheit

To the top of the page