In this edition of The EDiT Talks, Laurie Forcier, VP of Strategy at EDT&Partners, speaks with Professor Eric Tschirhart, Professor of Physiology at the University of Luxembourg. Eric has spent decades in Higher Education, watching knowledge transfer evolve from laptops and Wikipedia to large language models (LLMs). He has become one of the most vocal advocates inside his institution for a different approach to AI adoption: one led not by committees, but by what he calls the "cowboys" or outliers.

‍

Every university has them. The academics who do not wait for institutional consensus. The ones who test, break things, and prove what works before anyone has written a policy paper about it. Eric is one of those people, and his argument is that universities need to find more of them, fast. Because while institutions deliberate, students are already using AI on their own terms, and the gap between what generic tools deliver and what learners actually need is growing.

‍

[insert YT link]

‍

The EDiT talks is a series of conversations with leaders shaping the future of education and technology.

‍

Laurie asks Eric how the University of Luxembourg approached AI adoption. His answer starts with a problem he could see in front of him: students were using general-purpose LLMs and getting 25-page responses to simple questions. The outputs were technically correct but educationally useless. Find out how he approached this challenge through Lecture, the open-source genAI framework developed by EDT&Partners in collaboration with AWS.

‍

Key points from the conversation:

• The problem with generic AI in education is not accuracy, it is relevance. Students using general-purpose LLMs were getting unfocused outputs that bore no relationship to their actual learning objectives.
• Eric's response was to rebuild his courses around tightly defined learning objectives, then work with EDT&Partners to develop a purpose-built LLM using Lecture, an open-source generative AI framework designed for teaching and learning. A smaller, constrained model that gives students exactly what they need and nothing more.
• The biggest barrier to AI adoption in Higher Education is emotional. Experienced professors can see AI as a threat to academic authority, and roughly half of Eric's colleagues remain hesitant. Universities, he notes, change like elephants.
• The way past that inertia is not top-down mandate. It is finding the "outliers" and "cowboys", the willing early adopters who can test, prove value, and build momentum from the inside. Waiting for institution-wide policy, Eric argues, means falling behind while students adapt without guidance.
• The impact is already measurable. Before exams, Eric's students generated 600 to 800 queries to the course-specific LLM in just a couple of days, questions that would previously have flooded his inbox or gone unasked entirely.
• Students are still assessed on their own thinking. They may use any AI tool for assignments, but must defend their reasoning in a 15-minute oral examination, explaining how they used AI and how they interpreted the output. The result is stronger critical thinking.

‍

"In medical school, ranking is vital and one point can change a life. Fairness is a superpower." - Professor Eric Tschirhart

‍

Eric is also making the case that human evaluation depends on mood, and that AI can level that out by assessing against the same knowledge base consistently. "Fairness is a superpower" is reframes AI's value away from efficiency and toward something more human.

‍

What makes the University of Luxembourg example worth paying attention to is that the experimentation is disciplined. The LLM is constrained to course material. The learning objectives act as guardrails. Students get focused, relevant responses instead of noise. And the results, hundreds of queries replacing unanswered emails, oral examinations replacing blind submission, are already changing how both teaching and assessment work in practice.

‍

Eric's message to universities still in early-stage deliberation is direct: do not wait for consensus. Find your cowboys, give them room to prove the model, and build outward from there. Good AI implementation starts with the right design centered around learning and the right people.

‍