Summary

• KIT’s AI strategy covers research, teaching, technology transfer, administration and digital infrastructure.
• Data sovereignty, transparency, sustainability and regulatory compliance are built into the institution-wide framework.
• The university is investing €17 million in its HoreKa 2 supercomputer to provide local computing resources for research and AI applications.

Karlsruhe Institute of Technology has adopted an institution-wide artificial intelligence strategy that places computing infrastructure, staff guidance and governance inside the same programme rather than treating AI as a collection of separate research and software projects.

KIT says the strategy will govern the use and development of AI across research, teaching, technology transfer, administration and infrastructure. Its principles include data sovereignty, transparency, sustainability and regulatory compliance, while the university is simultaneously expanding access to AI tools and investing in computing capacity under its own control.

Universities face an unusually broad implementation problem because researchers are building advanced AI systems, students are using generative tools, administrators are looking for automation and departments are handling sensitive information inside the same institution.

A single policy cannot resolve all of those demands, but an institutional strategy can establish common boundaries for systems that might otherwise spread through departments independently.

AI adoption arrives through several doors

Research organisations encountered artificial intelligence long before the current wave of generative tools because machine learning is already used to analyse scientific datasets, model physical systems and automate experimental work. Large language models have added another route through writing, coding, research assistance and administration.

That combination makes central governance difficult. Rules written only for academic research miss software used by administrative employees, while restrictions designed around student coursework may be inappropriate for scientists developing new models themselves.

KIT’s framework therefore establishes broad principles while recognising different operating environments. The university says it has already produced guidance for generative AI, introduced training and expanded access to modern tools through an AI toolbox for employees and students.

Training alongside access matters because many generative AI risks are behavioural as well as technical. Staff and students need to understand when information can be entered into a service, how generated output should be checked and which responsibilities remain with the person using the system.

Those questions become particularly important when research is unpublished, commercially sensitive or subject to legal and ethical controls. A convenient external AI service can create unwanted data exposure if users do not understand what happens to prompts or uploaded material.

Data sovereignty requires computing capacity

KIT is supporting the governance framework with its own infrastructure, including a €17 million investment in the HoreKa 2 supercomputer intended to provide powerful and data sovereign computing resources for research and AI applications.

That investment illustrates the difference between writing a sovereignty requirement and being able to implement it. An organisation that wants sensitive workloads to remain within infrastructure it controls needs enough local computing capacity to make that option practical.

The constraint is particularly acute in AI because modern models can require expensive accelerator hardware. Smaller organisations may have little alternative to public cloud services, while a large technical university can justify shared infrastructure across many research teams.

Local compute does not eliminate external dependency because accelerator chips, software frameworks and other components still come from global supply chains, while researchers will continue using external models and cloud services where those are the most appropriate tools.

It does give KIT another deployment option. Projects involving sensitive datasets or particular sovereignty requirements can operate within institutional infrastructure rather than forcing policy teams to choose between breaking their own rules and preventing researchers from carrying out computationally intensive work.

A university has to govern innovation without freezing it

Experimentation forms part of a university’s purpose, making its AI governance problem different from that of an organisation with a tightly defined operating environment.

An overly prescriptive framework could become obsolete quickly or obstruct legitimate research, while a framework that remains too general provides little practical help when employees have to decide whether a particular dataset or model can be used.

KIT says its strategy was developed through an institution-wide participatory process. That can improve legitimacy and expose requirements from different disciplines, although the more demanding test comes when broad principles are translated into procurement rules, system access and daily decisions.

Sustainability creates another tension because AI can accelerate research and automate work while the computing systems behind it consume substantial electricity and hardware resources. Universities studying environmental impact increasingly have to apply similar scrutiny to their own digital infrastructure.

The strategy explicitly includes sustainable AI alongside transparency and sovereignty, bringing those trade-offs into institutional planning rather than treating computing capacity as an unlimited input.

Governance increasingly depends on architecture

AI governance cannot remain confined to an ethics statement when decisions about infrastructure determine where data can be processed, procurement decides which suppliers receive access and technical controls establish what users can actually do.

Those layers become more important as AI shifts from optional experimentation into routine institutional work. Once assistants, analysis tools and automated workflows spread across departments, inconsistent local decisions become harder to manage and audit.

KIT is therefore treating artificial intelligence as part of institutional architecture rather than another application category. Its strategy still has to prove itself in implementation, particularly where research freedom and central controls collide.

The combination of common rules, user training and controlled computing capacity nevertheless moves the programme beyond policy documents alone. Responsible AI becomes easier to require when the institution has built an infrastructure capable of offering a practical alternative.