What the MIT cognitive debt study says, and what it does not say.

The MIT Media Lab project page describes the study as exploring the neural and behavioural consequences of LLM-assisted essay writing (project page; preprint at arXiv 2506.08872). Participants were assigned to one of three conditions: LLM-assisted, search-engine-assisted, or unaided ("brain-only"). They wrote essays on assigned prompts while researchers collected EEG data alongside written outputs. The researchers then applied recall tests and linguistic analysis to compare across groups.

The reported findings, summarised by TIME in 2025 (TIME coverage), included measurable differences in neural engagement during composition, with the LLM-assisted group showing reduced activation patterns associated with effortful processing relative to the unaided condition. Recall tests administered afterwards showed the LLM-assisted group reconstructed their own essay content less reliably. Linguistic analysis indicated more homogeneous language patterns in the LLM-assisted group, consistent with the model's stylistic influence.

The work introduces the phrase "cognitive debt" — a way of naming the gap between using a tool to produce a result and being able to do that same thinking yourself afterward. It is a framing borrowed loosely from technical-debt language in software, not a clinical diagnosis.

The study is a preprint, not a peer-reviewed publication. It is task-specific (essay writing on assigned prompts), it studies a specific population (the MIT participant pool), and it does not claim that all AI use is harmful. It does not claim that LLMs cause general cognitive decline, dementia, or long-term loss of intelligence. It does not produce a clinical recommendation.

A formal 2026 commentary by Stanković et al. (Stanković critique) raised methodological concerns: short task duration, single-session design, possible confounds in how the unaided group's effortful processing was interpreted, and the gap between EEG patterns and any claim about durable cognitive change. The critique does not say the MIT findings are wrong — it says the inference jump from "differences during one writing session" to "cognitive debt as a general phenomenon" is larger than the study's design supports.

For the same reason, Senwitt does not call the study "peer-reviewed evidence" or use it to support broad transfer claims. It is one carefully designed signal that asks a good question, and the careful posture is to read it alongside the critique rather than treating either as the final word. The broader cognitive-offloading literature (Risko & Gilbert, 2016) provides more durable context than any single 2025 result.

The honest read of the study is the one Senwitt builds around: when a tool can do more of the cognitive work, deliberate practice becomes more valuable, not less. That is not a panic about AI. It is a design principle for habits.

Senwitt's response to that signal is not a clinical claim. It is a daily place to practise the thinking you still want to own — writing, math, code, memory, reading, and reasoning — in one short Set. The product's position does not rest on the MIT study being right in its strongest form. It rests on the older, broader cognitive-offloading literature, of which the MIT preprint is one recent contribution.

For the longer argument on why the broader literature supports daily practice without requiring strong claims about any single study, see cognitive offloading and scope of evidence.