Cruz: LLM Outsourcing Breeds Illusions of Understanding

A new paper in Computational Brain & Behavior by Universität Potsdam researcher Nicole Cruz makes an argument that runs against the mood of a year when almost every research group is figuring out where to slot LLMs into its workflow. Her claim, stated bluntly in the paper, is that LLMs can be useful but they cannot think, and delegating research activities to them raises the risk of avoidable illusions of understanding.

The framing is not a general anti-AI complaint. Cruz separates two kinds of illusion. Some are baked into the research process itself and are inevitable. Others are avoidable, she argues, if researchers put in the work: precision and transparency in language and observation, domain expertise, and careful, critical and creative thinking. What worries her is that outsourcing to LLMs sits squarely in the avoidable category and is expanding fast.

Why this matters if you run a lab or edit a journal: the surface quality of AI-assisted writing has raced ahead of anything reviewers can easily interrogate. A literature review that reads fluently is not the same as one whose author has understood the sources, and Cruz's argument is that the gap between reading like understanding and actually being understanding is exactly the space where illusions grow. As a small tell, the paper carries an author note that no LLMs were used to write it.

The honest caveat is that this is an argument, not an experiment. Cruz does not quantify how often LLM-assisted research produces illusions, propose a test editors could apply, or draw fine distinctions between tasks where the risk is high versus low. Readers looking for a benchmark, an incident count or a domain-by-domain risk map will not find one here.

What is worth taking away is the direction of the recommendation. If Cruz is right that there are no shortcuts to understanding, then the researchers and institutions that keep investing in precise language, domain training and slow critical thinking are the ones who will end up with a widening quality gap over those treating LLMs as a thinking substitute rather than a tool.