Is AI Giving You Donkey Brains?

A four-month MIT study using EEG brain scans found that people who wrote with ChatGPT assistance showed weaker memory and quotation accuracy than people who wrote without AI—alongside measurable differences in brain connectivity. The most striking behavioral result: 83.3% of ChatGPT-assisted participants failed to correctly quote from essays they had written minutes earlier, compared with 11.1% in both the search-engine and “brain only” groups. The implication is less about “AI makes you faster” and more about what gets offloaded: when the tool handles parts of drafting and synthesis, the brain may engage differently, leaving fewer durable traces for later recall.

Neural data in the study pointed in the same direction. Across frequency bands, the brain-only group showed stronger neural connectivity, especially in theta and high alpha bands—patterns the researchers linked to heavier internal cognitive work, including working memory and planning demands of composing from scratch. In contrast, the LLM-assisted group showed reduced connectivity overall, with the transcript describing a drop in measured connections (79 vs. 42) and differences in connectivity pathways such as frontal-to-posterior communication. The discussion also highlighted how alpha connectivity is often associated with internal semantic processing and creative ideation; higher alpha connectivity in the brain-only group was interpreted as greater internally driven idea generation, while the LLM group appeared to rely more on external suggestions.

The study also surfaced a “productivity paradox.” ChatGPT can reduce cognitive load by streamlining information presentation and synthesis—making tasks feel easier and faster—but that reduction may come at the cost of deeper learning. In the transcript’s summary of the cognitive-load findings, LLM users experienced about a 32% lower “germane cognitive load,” described as mental effort tied to learning, schema construction, and building lasting mental models. That trade-off matters because learning isn’t only about producing an answer; it also depends on effortful integration and the mental “work” that helps knowledge stick.

Beyond memory, the study examined essay ownership and cognitive agency. Brain-only participants largely reported full authorship, while the LLM-assisted group showed fragmented or conflicted authorship—described as a diminished sense of agency and weaker self-monitoring/error-monitoring signals. Teachers in the study reportedly could detect AI-written essays even when language sounded polished: they described “soulless” writing—close to perfect structure and accuracy but lacking personal insights and distinctive content.

The transcript’s broader takeaway is not a call to ban AI, but a warning about habit formation. If people repeatedly offload drafting, synthesis, and revision to LLMs, they may build “cognitive debt”—a learned helplessness-like pattern where independent thinking and recall degrade when the tool is removed. The practical recommendation offered is to use AI strategically while still scheduling time to write, plan, and think without assistance, so the brain keeps practicing the skills that make learning and ownership durable.