AI skills in computer science degrees

What matters now

30.07.2026

Artificial intelligence, Career

More than 90 percent of students in Germany use AI tools for their studies. The nationwide longitudinal study by Hochschule Darmstadt documents a jump from 63.2 percent in 2023 to 91.6 percent in winter 2024/25.

For engineering team leads, that’s highly relevant. The graduates you’ll be hiring in the coming months completed their computer science or STEM degree with a toolkit that didn’t exist in this form three years ago. The interesting question is: What did students actually learn while using it?

Good results are not yet proof of competence

The OECD Digital Education Outlook 2026 describes a finding worth paying attention to. Students with access to general-purpose AI tools produce better work than peers without that access. But the advantage disappears as soon as exams are taken without AI, and in some cases even reverses. The OECD calls this the “illusion of learning.”

For hiring, that has a very practical consequence: a working project and a cleanly documented repository say less about a candidate’s actual abilities today than they did just a few years ago. What’s more telling is how a candidate got there. What decisions did they make? How did they recognize that an AI suggestion wasn’t sound? Universities are currently going through exactly this same shift from evaluating the product to evaluating the process.

The pressure to act is clearly measurable there. According to the 2025 AI Monitor, 97 percent of the surveyed universities are examining how AI affects exams. Yet only 43 percent have actually adapted their examination regulations. Between recognizing the problem and doing something about it, there’s still quite a bit of ground to cover.

Why this matters especially in safety-critical projects

In regulated industries, code that nobody fully understands is a documentation problem. Anyone developing software under IEC 62304 or ISO 26262 needs to be able to justify why an implementation looks the way it does. AI tools for code generation, testing, and documentation are treated as computerized systems in such environments and require their own validation, for instance under GAMP 5. The EU AI Act adds the obligation to ensure sufficient AI literacy within the organization.

This shifts the bottleneck from writing to reviewing. When a new hire generates more code in an afternoon than an experienced colleague can review in a week, review capacity becomes the actual constraint. And review requires understanding.

This shift is already showing up in the labor market. Stanford’s Digital Economy Lab finds a relative employment decline of 13 percent among 22- to 25-year-olds in highly AI-exposed occupations in payroll data, around 20 percent among young software developers compared to the peak at the end of 2022. Employment among experienced professionals stayed stable over the same period. Whether the cause actually lies in AI or in interest rate conditions is scientifically disputed. Still, the direction of the signal fits what the World Economic Forum’s Future of Jobs Report 2025 projects: 39 percent of today’s core skills will change by 2030.

Orchestration becomes a learning objective

Some universities are already actively shaping this shift. Prof. Dr. Christoph Neumann at OTH Amberg-Weiden, with whom sepp.med works closely, has for several years used prompt-based exercises in foundational courses, coursework on prompt engineering and vibe coding, and AI-supported project work in web development, cloud computing, and big data. His assessment model turns productive AI use into a learning objective rather than a rule violation. His thesis: what matters is no longer simply executing tasks, but the ability to deliberately orchestrate AI systems.

This direction aligns with the institutional line. Since July 2026, the German Science and Humanities Council (Wissenschaftsrat) has recommended, under the guiding principle of intellectual sovereignty, a dual path: systematically building AI skills while also preserving AI-free spaces in the curriculum where independent thinking is practiced. The developer community is arriving at a similar approach. Andrej Karpathy, who coined the term “vibe coding” in 2025, now describes “agentic engineering” as its successor: working with AI agents, but with considerably more control and review.

What this means for your team

Evaluate process, not just product: In interviews, ask about verification strategies, not only about results.
Plan for review capacity: More generated code means more review effort, not less.
Validate AI tools: In regulated environments, risk assessment and usage policies belong in every tool rollout.
Build targeted skills: Prompt and context engineering can be taught, and so can evaluating AI-generated test cases.
Stay in touch with universities: Internships, theses, and joint projects show early on which skills are coming up through the pipeline.

A single prompt workshop doesn’t create review competence on its own. It’s the starting point that structured practice can build on. That’s exactly the principle behind our AI-supported software testing, which follows a human-in-the-loop approach: AI handles the routine work, experienced testers validate the critical cases, and the review stays documented and audit-ready. For AI tools and AI components themselves, quality assurance for AI systems applies. If you want to get your team up to speed, the sepp.med Academy offers suitable seminars.

AI skills in computer science degrees

Good results are not yet proof of competence

Good results are not yet proof of competence

Why this matters especially in safety-critical projects

Orchestration becomes a learning objective

What this means for your team

FAQ

Sources

Get in touch with us!