Does ChatGPT Affect Learning? What One Essay Study Can and Cannot Show

Matthew Berman discusses a 2025 MIT-affiliated study of AI-assisted essay writing and asks whether using ChatGPT can impair learning. His opening claim of lasting “damage” is stronger than the study supports: it examined a particular writing task, not permanent brain injury or general intelligence.

He outlines questions about mental effort, memory and ownership when people write with a language model, a search engine or no external tool. These are distinct outcomes and should not be collapsed into a single measure of intelligence.

Berman says the paper makes him reconsider how he uses language models for learning. He remains interested in their value but wants to be more deliberate about when they generate answers versus when he works through a problem himself.

He notes that a generated citation can be fabricated or fail to support the associated claim. Checking the actual source remains necessary even when an answer contains links.

The discussion considers social learning and the opportunity to reason with teachers and peers. Berman worries that delegating too much work to a model could reduce those opportunities, though the essay study does not directly measure every form of social learning.

He compares web search with model-generated synthesis. Search can require evaluating multiple sources and navigating clutter; a concise answer can reduce some effort while also changing what the learner chooses to inspect.

Berman argues that lower task effort can be a tradeoff: it may speed up an output while providing less practice in assembling and testing one’s own ideas. The balance depends on how the tool is used and the task being performed.

He describes a shift from actively reasoning to supervising an AI-generated response. This is a hypothesis about a possible workflow, not a result that applies automatically to every user or interaction.

The experiment compares essay writing with an LLM, search and no tool. According to the [researchers’ preprint](https://arxiv.org/abs/2506.08872), 54 people participated in the first three sessions and 18 completed a fourth crossover session; Berman’s cited counts differ.

The video also discusses other research about short-term task performance and knowledge transfer. Findings from different studies should not be combined as though they came from the same experiment.

Berman acknowledges the small sample and describes the groups’ tools and later switching of conditions. The [MIT Media Lab project page](https://www.media.mit.edu/projects/your-brain-on-chatgpt/overview/) identifies the June 2025 report as a preliminary preprint at that time.

He asks whether readers can identify AI-written text and offers his own impressions about style and originality. Feeling that prose is generic is subjective and does not provide a reliable authorship test.

The study reports differences in participants’ perceived ownership of their essays. Berman notes that the responses within the AI group were not uniform, so a single average cannot describe every participant.

He discusses the authors’ interpretation of differences in EEG connectivity patterns. EEG connectivity is a specific measurement and analysis, not a direct percentage of the brain being used, a measure of intelligence or proof of injury.

The video highlights differences in participants’ ability to quote their essays and in the small group that switched tools. Those findings are limited to the tested tasks and sample; they do not demonstrate lasting cognitive damage from ordinary ChatGPT use.

Berman interprets the crossover results as a reason to think independently before seeking AI assistance. That can be a useful learning strategy to test personally, but the small crossover sample cannot establish a universal sequence or long-term effect.