Photo by Bhautik Patel on Unsplash
What We Found
55%. That is the measured gap in cognitive engagement between participants who wrote essays entirely unaided and those who used ChatGPT — a figure reported by MIT Technology Review on June 5, 2026, drawing on a study conducted at MIT Media Lab and published in June 2025. As of June 15, 2026, this research is forcing a harder look at the hidden costs of AI-assisted workflows across education and the workplace.
According to Google News aggregation of the coverage, the Washington Post reported the study tracked 54 participants over four months using EEG (electroencephalography — electrodes that measure electrical brain activity in real time) while they wrote essays under three conditions: pen-and-paper, Google Search, or ChatGPT. Brain-only participants showed the strongest neural connectivity. ChatGPT users displayed the weakest alpha and theta brainwave activity — the specific patterns tied to deep memory formation and reflective processing. Dr. Nataliya Kosmyna, the MIT Media Lab neuroscientist who led the research, stated plainly: "The convenience of having this tool today will have a cost at a later date."
The recall data made the gap harder to dismiss: as of the study's June 2025 publication, 83% of ChatGPT users could not accurately recall key arguments from essays they had just written, nor accurately quote from papers they themselves produced. Google Search users fell somewhere in between.
The Evidence
That MIT study does not arrive in a vacuum. Frontiers in Psychology published a meta-analysis drawing on 70 research papers and data from roughly 17,000 college students collected between 2001 and 2019. Three out of four measured emotional intelligence components — specifically well-being, self-control, and emotionality — showed significant decline over that period, with increased technology usage identified as a contributing explanatory factor.
Gloria Mark, a psychologist at UC Irvine who has spent decades measuring digital distraction, brings another dimension. Her longitudinal research found that average attention spans on screens fell from approximately 2.5 minutes in 2004 to 75 seconds by 2012, and then to just 47 seconds across the 2014-to-2020 window. Her research also documents that after a single interruption, workers take an average of 23 minutes and 15 seconds to fully recover their focus on the original task. In her interview with MIT Technology Review, Mark pointed to emotional intelligence as the specific capacity at greatest risk: "The muscle we risk atrophying is emotional intelligence, which surveys suggest is already on the decline. The key factor is effort, and the more effort we put into something, the deeper the satisfaction we stand to gain."
A Harvard University survey of 1,400 American workers adds a workplace layer: as of that survey's publication, approximately 14% of respondents described experiencing what they called "mental fog" after intensive AI chatbot sessions. Separate lab experiments found that participants who used AI assistance for as little as 10 minutes on math or reading problems showed diminished unaided performance immediately afterward, and abandoned difficult problems more quickly than baseline groups.
What It Means for Your AI Tool Stack
Chart: Measured attention spans on screens across three time periods, per UC Irvine psychologist Gloria Mark's longitudinal research. The decline predates generative AI — suggesting the trend has only accelerated since.
The Conversation pushed back on the bluntest version of the "AI damages your brain" framing, arguing that timing and manner of use matter more than use itself. Their analysis of the MIT data noted that AI can support deeper thinking when introduced after an initial independent attempt — functioning as a verification tool rather than a draft generator.
That distinction matters enormously for how productivity professionals structure their workflows. ChatGPT, Claude, and Gemini are all engineered to minimize friction to a first output. That is their core value proposition. But the MIT evidence suggests the specific habit pattern — prompt first, think second — is precisely what drove the weak neural connectivity the EEG captured. The tools themselves are not inherently harmful; reaching for them before any independent cognitive engagement may be.
For enterprise contexts, this lands differently. Companies deploying AI chatbots for financial analysis, document review, or customer decision support are, in effect, building workflows where the effortful-thinking phase is delegated by default. Researchers coined the term "cognitive debt" in 2025 — explicitly analogous to technical debt in software — to describe how that pattern compounds quietly until a moment requiring unassisted judgment arrives. As the AI agent ecosystem continues expanding, as Smart AI Trends has been tracking, the question of which human cognitive capacities get maintained versus outsourced is becoming harder for organizations to ignore.
The Limit Nobody Is Marketing
No AI vendor's pricing page mentions cognitive cost curves. The pitch is always productivity, democratized intelligence, and speed. What the research literature is now documenting — across MIT Media Lab, UC Irvine, Harvard, and Frontiers in Psychology — is that for specific task types (analytical writing, math problem-solving, reflective reading), repeated AI delegation may quietly erode the very skills it appears to be replacing.
Education institutions are beginning to respond. New AI literacy programs increasingly emphasize "effortful learning" — designing sequences where students attempt tasks independently before accessing AI assistance. Mark's practical prescriptions, detailed in her MIT Technology Review interview, are deliberately low-tech: read complete books rather than AI-generated summaries; navigate familiar routes without GPS to exercise spatial memory; engage with challenging material before requesting an AI synthesis. These are not anti-technology positions. They are maintenance routines for the cognitive substrate that makes any tool use meaningful.
My read: the industry will not solve this. The incentive structure runs entirely the other way — engagement, retention, and session depth are all maximized when AI handles more, not less. The behavioral design choices have to come from users and organizations, not product teams.
How to Act on This
Identify tasks where you currently open ChatGPT, Claude, or a similar tool as the opening move. Flag any that involve synthesis, judgment, or analysis — the categories where the MIT data shows cognitive cost accumulates most visibly. The goal is not eliminating AI assistance; it is noticing where habit has replaced intention. Works fine for a team of three; becomes a cultural problem at thirty if nobody names the pattern.
For documents, analyses, or decisions that carry real weight, commit to an initial independent draft before opening an AI assistant. Even a five-minute independent attempt changes the neural engagement profile, according to the MIT findings. This mirrors the "attempt first, assist second" sequencing that AI literacy researchers are now building into curricula. For long desk sessions, pairing this focused-thinking practice with a quality ergonomic keyboard reduces physical friction and helps sustain the habit.
Gloria Mark's attention-span research shows that consistent exposure to interrupting technology degrades task-return speed — 23 minutes and 15 seconds per interruption in her measurements. Blocking screen-free windows for demanding analytical work is not nostalgia; it is maintenance for the capacity that makes AI-augmented work effective in the first place. Noise canceling headphones can help carve out those intervals in open-plan environments without requiring a separate room.
Frequently Asked Questions
Does using ChatGPT actually weaken your thinking skills over time?
The current research documents specific short-to-medium-term effects rather than confirmed permanent decline. As of June 15, 2026, the MIT Media Lab's June 2025 study found that habitual ChatGPT use during essay writing correlated with significantly weaker neural connectivity (measured via EEG) and substantially lower recall of one's own written arguments compared to pen-and-paper or Google Search users. The Conversation's analysis of the same data argues the effect depends heavily on when AI enters the workflow — use after independent effort may carry a different cognitive cost than use as the first step.
Is AI bad for your brain and memory specifically?
The evidence points to specific mechanisms rather than a blanket conclusion. The MIT EEG data found weaker alpha and theta brainwave activity in ChatGPT users — patterns associated with deep memory encoding and reflective processing. The Frontiers in Psychology meta-analysis covering 70 papers and 17,000 students across 2001 to 2019 found declines in three emotional intelligence components, with technology usage as a partial explanatory factor. Separate lab experiments showed that just 10 minutes of AI-assisted problem-solving reduced unaided performance immediately afterward. The consistent thread: offloading the effortful phase of a cognitive task reduces the neural engagement that consolidates learning and memory.
How does AI affect cognitive function and learning differently for students versus professionals?
The MIT study used essay writing — a task common to both students and knowledge workers. The Harvard survey of 1,400 workers found 14% reporting mental fog after intensive AI chatbot use, suggesting the workplace effect is real, not limited to academic settings. For students, the risk is that skills under active development get bypassed before they consolidate. For professionals, the risk is subtler: capabilities that exist but go unpracticed may atrophy quietly, surfacing as a gap precisely when unassisted judgment is most needed. The "cognitive debt" framing coined in 2025 applies to both groups — the compounding just looks different depending on career stage.
Bottom line: The research as of June 15, 2026 is specific enough to change behavior without warranting a wholesale rejection of AI tools. The consistent finding across MIT Media Lab, UC Irvine, Harvard, and Frontiers in Psychology is that sequencing is the critical variable — who engages first, the human or the model. Workflows where AI eliminates the effortful thinking phase carry documented cognitive cost. Workflows where AI extends or validates independent effort may not. That distinction will not appear on any product roadmap. It has to come from the user.
Disclaimer: This article presents editorial commentary based on publicly reported research and does not constitute medical or professional advice. Research based on publicly available sources current as of June 15, 2026.