You have probably received a message like this in a discussion:
C
A coworker you probably value a lot 2:47 PM
ChatGPT said: <A lot of raw LLM output> Sometimes even without the label. Someone asked an LLM about the topic you were discussing and pasted the answer straight into the conversation.
Chances are you have been on the sending side too - I know I have. The problem: Pasting an answer you don't understand gives you words, not understanding - people confuse reading an answer with understanding it1 - and it shifts the effort to the other person. You spend seconds prompting, while they must read the answer, understand it, fact-check it2,3 and decide what is relevant. When they reply, all you can do is paste their answer back into the LLM, because you still don't understand what the conversation is about.
I have a suggestion for a better way to handle this: Use the LLM as a learning tool. Ask it to explain unclear points, give examples and challenge your understanding. Then put the answer away and explain the topic in your own words. Research shows that this improves understanding and memory.4,5 Once you understand the topic, return to the other person with actual knowledge instead of copied words and continue the discussion at eye level. That knowledge goes into your backpack: The next time the topic comes up, you can join the discussion without pulling out your phone - and one day it may even help you solve a problem that has nothing to do with it.
Thinking is also a use-it-or-lose-it skill. When you stop doing the thinking yourself, you don't remember what you hand off,6 you perform worse without the tool,7 and people who lean on AI a lot tend to score worse at critical thinking.8 Check important claims and come back when you can explain what you think and why.
We're all still figuring this out. We need to find the balance: LLMs make us faster by default, but the ones who come out ahead use them to get smarter. We need to make sure that, when there's no LLM at hand, there's still a capable person - someone who can discuss, think critically and work things out on their own.
Sources #
- Rozenblit & Keil (2002), The misunderstood limits of folk science: an illusion of explanatory depth. https://cogdevlab.yale.edu/sites/default/files/files/rozenblit%20%26%20keil%20%202002.pdf
- Lin, Hilton & Evans (2022), TruthfulQA: Measuring how models mimic human falsehoods. https://aclanthology.org/2022.acl-long.229/
- Ji et al. (2023), Survey of hallucination in natural language generation.
[https://doi.org/10.1145/3571730](https://doi.org/10.1145/3571730)
- Chi et al. (1994), Eliciting self-explanations improves understanding.
https://www.sciencedirect.com/science/article/pii/0364021394900167
- Karpicke & Blunt (2011), Retrieval practice produces more learning than elaborative studying with concept mapping. https://learninglab.psych.purdue.edu/downloads/2011/2011_Karpicke_Blunt_Science.pdf
- Grinschgl, Papenmeier & Meyerhoff (2021), Consequences of cognitive off: Boosting performance but diminishing memory. https://doi.org/10.1177/17470218211008060
- Bastani et al. (2025), Generative AI without guardrails can harm learning: Evidence from high school mathematics. https://www.pnas.org/doi/10.1073/pnas.2422633122
- Gerlich (2025), AI tools in society: Impacts on cognitive off and the future of critical thinking.
[https://www.mdpi.com/2075-4698/15/1/6](https://www.mdpi.com/2075-4698/15/1/6) (correlational, self-report)
[Back](../index.html)