Outsourcing thinking to AI? Think about it
What AI does to our judgment: cognitive debt through AI
The other day I flinched for a moment: before I think for myself, I first type a question into the window. Further thoughts of my own come afterward—mostly as a reaction to what the machine has already said. If I’m honest, that is also a little embarrassing for someone who also makes a living from people thinking for themselves.
So I am not an exception here, but evidence. What happens to me on a small scale shows up in research on a large scale—and quite consistently. A study with more than 600 people found that those who use AI tools frequently show weaker critical thinking skills, mediated by cognitive offloading. That means outsourcing the thinking instead of finishing it yourself. Hardly surprising if you take an honest look at your own chat history. PsyPost
MIT did not settle for assertions; it measured. People who wrote essays with ChatGPT ended up with the weakest neural connectivity of all three groups, and of all places in the bands associated with creative ideation. If you like, picture it as a muscle that is told it no longer needs to exert itself, and that believes this without objection. Bonus punch line: the texts of the AI group were strikingly homogeneous—so we are producing uniform mush. Even when the AI was taken away from the group again, their thinking remained weaker than that of those who had worked alone from the start. This creates a training deficit that we apparently also bring upon ourselves. TimeWBUR News
Not all of it is something to smile about. Three months after AI support was introduced, clinical staff lost six percent of their ability to detect tumors without AI assistance. That is where the sarcasm ends for me. arxiv
I consider the following important, because when it occurs, we should take notice:
- Asking the machine first, then thinking for yourself—not the other way around
- Adopting a solution without understanding how it came about
- Believing something is right because it reads fluently—not because it has been checked
- Uncertainty with tasks that used to be a matter of course
- The feeling of increasingly hearing the same three thoughts in conversations
- It becomes harder to come up with several options of your own on the spot
Fortunately, none of this is irreversible. It is enough to reverse the order. First wrestle with it myself, then bring in the machine, then hold the result up against my own idea instead of swallowing it. And also to allow yourself phases in which you deliberately work without any tool, so that the muscle does not completely unlearn what it is there for. AI can do a great many things much better than we humans can. It would make sense, however, to design these use cases deliberately, and not automatically toss much (or everything) over the fence to the AI.
But there is one thing the machine cannot do (or only very rarely): hand me back a question that makes me falter myself. Most of the time I can answer perfectly well. Questions like these, which are no longer based on pattern recognition but on the present moment, on emergence, on insight, remain human work—which is probably also the reason why I will not become replaceable in the medium term either.
PsyPost (study on cognitive offloading, 600+ participants): https://www.psypost.org/ai-tools-may-weaken-critical-thinking-skills-by-encouraging-cognitive-offloading-study-suggests/
Time (MIT study, neural connectivity): https://time.com/7295195/ai-chatgpt-google-learning-school/
WBUR (also on the MIT study): https://www.wbur.org/news/2025/09/16/ai-study-essays-brain-cognition
arXiv (clinical staff, 6% loss in tumor detection): https://arxiv.org/pdf/2602.21012