“Say as I Don’t Say”

In a new publication from our chair on the use of Large Language Models (LLMs), published in the Journal of Management Information Systems (VHB A; FT50), Prof. Adam and other researchers shows that LLMs are most useful when they do more than simply agree with us. The best outcomes arise when users critically reflect on their own assumptions while the AI challenges their thinking.


The takeaway: treat LLMs as sparring partners – question, test, and refine ideas instead of passively accepting their answers.

Martin Adam: “I am particularly proud of this project because it originated through back-and-forth reflections with others while teaching the use of LLMs.”

The paper can be accessed here:
https://www.tandfonline.com/doi/full/10.1080/07421222.2026.2692283

The following interview highlights the main idea of the publication:

AI as a Sparring Partner: How Critical Thinking Improves Collaboration with LLMs

Large Language Models like ChatGPT can make knowledge and ideas quickly available – but achieving high-quality results requires more than simply using them. A recent study in the Journal of Management Information Systems examines how people can improve the quality of collaboration with LLMs through reflection. We spoke with the author, Prof. Dr. Martin Adam, about the key insights.

Professor Adam, what was the core question of your study?

We are interested in how humans can collaborate with LLMs without either blindly trusting them or prematurely rejecting their support. Both extremes are problematic. Therefore, we view reflection as a central mechanism: users should consciously compare what they themselves know about a topic with the LLM's responses and actively question contradictions or new perspectives.

What form of collaboration leads to the best results?

Particularly effective is the combination of dialogic thinking and an adversarial interaction with the LLM. Dialogic thinking means not just slightly adjusting one's own assumptions, but being able to fundamentally question them. At the same time, the LLM should not simply agree, but provide counterarguments, point out weaknesses, and introduce alternative perspectives. In our findings, precisely this combination leads to a higher factual quality of the work outcomes.

What common mistakes do users make when working with generative AI?

A common problem is that convincingly formulated answers are accepted too quickly. As a result, people give up a part of their own cognitive responsibility to the system. We speak here of the importance of cognitive primacy: the human should retain mental control. An LLM can provide new information and perspectives, but the evaluation and contextualization should remain with the human.

What does this mean in practical terms for companies and universities?

We should view LLMs less as all-knowing assistants and more as sparring partners. This can also be put into practice very easily: instead of just asking for a solution, one can prompt the model to criticize one's own arguments, take up counter-positions, or identify potential errors.

For organizations, this means not only providing access to generative AI, but also developing the competencies needed to use it reflectively. The goal should not be to replace human thinking with AI, but to strengthen it through productive friction.

What is your main recommendation for working with LLMs?

Leverage the strengths of AI—but remain in control of your own thinking. Effective human–AI collaboration does not come from the system agreeing with us as often as possible. It emerges when humans and AI challenge each other.