KIT made its AI strategy a campus operating system, not a memo

THE SIGNAL IN ONE SENTENCE

Most institutional AI strategies are written as if the technology lives in a conference room. It does not. It lives in the researcher's code, the student's draft, the instructor's assessment, the administrator's spreadsheet, the library search, the meeting record and the computing system humming behind the campus wall. Karlsruhe Institute of Technology is trying to govern all of that as one institution. KIT published an English announcement on October 5 describing a new institution-wide AI strategy. Its Digital Office had listed the strategy on September 29 as a binding framework for generative AI, supplementing existing guidelines and defining the institutional role of AI in research, teaching, technology transfer, administration and infrastructure. That breadth is the signal. The strategy is not presented only as a list of forbidden uses or a cheerful note telling everyone to experiment responsibly. KIT connects principles to shared tools, training and its own computing capacity. It says data sovereignty, transparency, sustainability and regulatory requirements should be considered from the beginning. The campus already has some of the machinery. KIT's AI Toolbox gives employees and students access to several local and external language models through institutional accounts. The university says users must complete an AI competency module before they can use the toolbox. Its guidance tells users to consider data categories, protection classes and model choice because some requests stay in KIT's environment while others may be processed externally. That is not glamorous, which is exactly why it matters. The important institutional AI product may not be one brilliant model. It may be the boring layer that helps thousands of people answer five questions before a prompt leaves the building. What data is this? Which model is appropriate? Where will the request be processed? What evidence must the user keep? Who remains responsible for the result? When those questions are answered separately by every student, researcher and office, governance becomes a scavenger hunt. When they are built into the campus service, governance begins to behave like infrastructure. That is what makes the operating-system metaphor useful. An operating system does not perform every job itself. It provides shared rules, permissions, interfaces and records so many applications can work without each rebuilding the same foundation. A university AI strategy should do the same. The policy layer states the principles and boundaries. The literacy layer helps people understand what a model can and cannot do. The access layer provides approved tools and routes work to the right environment. The infrastructure layer supplies computing capacity and keeps important workloads under institutional control. The evidence layer shows whether the whole arrangement improves work without quietly damaging research integrity, teaching, privacy, accessibility or labor. KIT has visible pieces in the first four layers. The institution says its strategy grew from a participatory process. In October 2025 it invited employees and students to contribute to the strategy's development. It has generative AI guidelines, training modules, workshops, a community of practice and a toolbox available across the university. It is also investing in computing infrastructure. The October 5 announcement points to the new HoreKa 2 supercomputer, a 17 million euro system intended to support research and AI applications with powerful, data-sovereign capacity. A separate KIT announcement from September describes a planned AI and scientific computing data center backed by 24 million euros from Germany's federal government and Baden-Württemberg. That facility is expected to use waste heat in the campus heating network, with first computers planned for 2030. Those investments put substance behind the word sovereignty. Data sovereignty does not mean every task must run on a machine physically owned by the university. It means the institution can make meaningful choices about data location, access, models, contracts and switching instead of discovering that a convenient external service has become the only workable option. That requires model routing, not one blanket rule. A researcher analyzing public literature may reasonably use a different service from a team working with unpublished patent material. A student asking for help understanding an equation carries a different risk from an administrator preparing a document containing personnel information. A local model may offer stronger control but weaker performance for a particular language or discipline. An external model may be useful while still being unsuitable for protected data. The institutional service should make those differences legible before the user starts typing. Give every tool a plain-language data label. Show whether prompts leave the institution. Show the retention rule. Name the permitted information classes. Warn when file uploads carry a different risk from ordinary text. Record the selected model and policy version alongside important outputs. Then make the safe path easier than the improvised path. If approved tools are slow, confusing or missing the features people need, students and staff will route around them. A beautiful policy document will not win a fight against a free account opened in thirty seconds. KIT's toolbox is promising because it creates one institutional entry point while still exposing multiple models. Its requirement that users complete competency training adds a deliberate pause before access. The next question is what the training changes. Completion is an administrative metric. Competence is a behavioral one. A useful program should test whether people can identify sensitive data, verify a citation, detect a fabricated claim, disclose meaningful AI assistance, choose an appropriate model, document an important workflow and recognize when the tool should not be used. Those skills differ by role. A first-year student needs help understanding academic integrity and protecting personal information. An instructor needs assessment designs that still reveal what a student understands. A researcher needs reproducibility, provenance and rules for confidential data. An administrator needs procurement, records management and a clear boundary around automated decisions. A software team needs evaluation, security testing and incident response. One universal slide deck will not carry that load. The same is true for transparency. Telling people that AI exists is not enough. The institution should make consequential use inspectable. If an AI tool helps screen applications, summarize a performance review, recommend research funding, flag academic misconduct or convert material for a student with a disability, the affected person needs to know what role the system played. They need the evidence behind the outcome, a human contact and a route to correction. The threshold should rise with the consequence. Drafting a meeting agenda may need a simple disclosure and a human edit. Producing an accessible version of course material needs quality checks by people who can judge accessibility. Assisting with a disciplinary decision needs a far stricter record and should never turn a probabilistic suggestion into anonymous authority. Sustainability also needs a ledger. KIT names sustainable AI as a concern and pairs its strategy with local computing investments. That creates an opportunity to publish more than a broad promise. Track energy per workload class, utilization of local systems, external model spending, waste-heat recovery, model size choices and whether a smaller system completed the task well enough. A university does not need to turn every prompt into a moral crisis. It does need evidence that infrastructure decisions match its sustainability claims. Then there is the missing fifth layer: outcomes. KIT's public materials establish direction, services and infrastructure. They do not yet provide a shared implementation calendar, adoption targets, incident statistics, assessment results or a public scorecard connecting the strategy to observed campus outcomes. That is the next useful publication. An institutional AI scorecard could be compact. Report how many people completed role-specific training and passed practical checks. Publish the share of toolbox use routed to local and external models by data class, without exposing private prompts. Count documented incidents, corrections and appeals. Measure whether students and staff can access approved tools regardless of income or disability. Audit citations, reproducibility and disclosure in selected workflows. Compare energy and cost across appropriate model choices. Show which proposed uses were rejected and why. The rejected uses matter. A strategy earns trust when it can say no to a tempting deployment, not only when it launches another tool. Universities have an unusual position in the AI transition. They build models, buy services, teach future workers, employ thousands of people, manage sensitive data and produce evidence society relies on. They are laboratories and institutions at the same time. That makes a campus a good place to demonstrate a more mature bargain. Provide capable tools. Teach people how to use them. Keep important infrastructure under meaningful control. Make the rules visible inside the workflow. Preserve human authority over consequential decisions. Measure whether the system works in practice. KIT has assembled many of the right components. The plain signal is that an AI strategy becomes real only when the safest useful action is also the easiest one to take. The policy should appear when a user chooses a model. The training should change how that person handles data. The infrastructure should create genuine alternatives. The audit record should show what happened. The outcome report should reveal where the institution still falls short. At that point, the strategy is no longer a memo. It is how the campus runs.