offline ai systems for businesses

Blog

  • In 10 minutes, you’ll be using offline, local, private A.I.

  • Clickbait ‘Research’ claims GPT causes brain rot.

    Welcome to 2025 where the sources are made up, the science is weak, and peer review doesn’t exist if the peeked result is ‘scary’.

    MIT Media Lab, arXiv:2506.08872

    This study has been making the rounds on LinkedIn, X, and Facebook. It’s being parroted by the kind of people who see a heatmap and think they’ve discovered God. I read the study. I read the methodology. I read the EEG data. And I’m here to explain, in clear language, why this is not science, it’s narrative engineering.

    If you’re sharing this paper as evidence of “ChatGPT causing brain rot,” you’re falling for scientific theater dressed up in academic branding.

    1. Sample Size is Statistically Meaningless

    The entire study hinges on 54 participants across three groups. Only 18 completed the full cycle including Session 4. These are MIT and Boston-area students, which means it’s already a culturally and academically skewed population. This is not generalizable.

    “We recruited a total of 54 participants… 18 participants among them completed session 4.”
    Section: Abstract

    In neuroscience, and especially in EEG work, small sample sizes are a red flag. Brain signals vary massively between individuals. That’s why robust cognitive EEG studies need large participant pools and replication across multiple environments to mean anything. None of that is present here.

    2. EEG is Misused and Misinterpreted

    They used the Enobio 32-channel EEG headset. While decent for entry-level neuroscience, it is not remotely close to clinical-grade equipment for what they’re claiming to measure.

    They analyzed differences in brainwave patterns—alpha, beta, theta, delta bands—and concluded that LLM users had “less neural engagement.”

    “LLM assistance elicited the weakest overall coupling.”
    Section: Abstract

    That’s not a conclusion, that’s an assumption. Lower EEG activity doesn’t automatically mean “bad.” It could mean improved efficiency, familiarity, or reduced stress. The authors lean into a bias: less effort = less intelligence. That is neither proven nor even well-argued in the data.

    3. Overreach on Cognitive Load and “Debt”

    They invent the term “cognitive debt” and attempt to position it as a real measurable outcome. It’s not. It’s marketing language that sounds technical. Nowhere in the study is “cognitive debt” defined in a mathematically or scientifically rigorous way.

    They rely on EEG band correlations and vague metrics like “ownership of essay” and “ability to quote” to imply that using ChatGPT damages your ability to retain knowledge. But there is no proven causal mechanism, and certainly no long-term follow-up to back that up.

    “LLM group… fell behind in their ability to quote from the essays they wrote just minutes prior.”
    Section: Abstract

    What they’re describing is short-term recall of an essay you barely wrote with your own hands. That’s not news. If you ask someone to read a legal contract generated by someone else, then recall it word for word five minutes later, the same issue arises. This isn’t about AI. It’s about task ownership.

    4. Emotional Language is Not Science

    The paper contains emotionally charged phrases like:

    • “We stopped early because the findings were too striking.”
    • “Cognitive offloading leads to deficiency.”
    • “It was too scary not to publish.”

    Fear is not a scientific outcome. This is not a measured response based on falsifiable results. It is narrative crafting designed to trigger sharing and debate on social platforms. And it worked.

    Scientific conclusions must be replicable, not reactionary. This is not the language of research. It is the language of clickbait.

    5. The Experimental Conditions Were Flawed

    Participants were assigned to one of three groups:

    • Group 1: LLM-only (ChatGPT)
    • Group 2: Search Engine (Google)
    • Group 3: Brain-only (no tools)

    Each participant wrote essays using prompts pulled from SAT questions. They then swapped conditions in Session 4, which the authors use to compare cognitive state shifts.

    But here’s the problem: The control variables weren’t held constant. Participants were not equally experienced with the tools. Some were first-time ChatGPT users. Others were seasoned. Some used LLMs to translate foreign languages. Others used it to bypass writing entirely.

    “Some participants used LLM for translation.”
    Summary Table, Page 3

    This is not a controlled study. It’s a user-behavior diary pretending to be neuroscience.

    6. Scoring Was Subjective and Internally Reinforced

    They used an “AI judge” to grade the essays. Let’s be clear: they used another LLM to score LLM-generated work, then compared it to human scoring. This is not a valid cross-check! It’s a recursive loop.

    “We performed scoring with the help from the human teachers and an AI judge (a specially built AI agent).”
    Abstract

    There is no external validation set. There is no independent scoring system. The outcomes were shaped by the very biases embedded in the original tools they were supposed to test.

    7. The “Ownership” and “Quoting” Metrics are Built to Fail

    The study found that LLM users were less able to quote their essays or felt lower “ownership.” That’s expected. If the AI generates the structure, content, and phrasing, you’re not going to remember it as well. But this isn’t cognitive damage, it’s basic memory encoding.

    They position this as a failure state, but it’s actually a feature. Cognitive offloading is the entire point of using tools. We don’t remember phone numbers because we use contact lists. That doesn’t mean our brains are rotting!!! It means we are prioritizing higher-order tasks and offloading the minutiae.

    8. No Real Peer Review

    This is a preprint on arXiv. It has not passed formal peer review. Yet it’s being shared on social media as if it’s a settled fact. People are reposting it with captions like “Proof that AI is making us stupid,” despite the obvious limitations and speculative leaps.

    Summary: What This Actually Is

    This study is not worthless. It’s just not what it claims to be. At best, it’s exploratory research on how different toolsets affect short-term engagement with a writing task. At worst, it’s a content-marketing campaign in the skin of science.

    The emotional charge, early stoppage, invented terminology, and tiny sample size all point to one thing: this was designed to be viral, not rigorous.

    Final Thought

    If you’re worried about ChatGPT “damaging brains,” this study is not your proof. It’s a hypothesis, dressed up like a conclusion, passed around by people who haven’t read it.

    If you want to talk about real risks of LLMs, we can talk about bias replication, dependency, epistemic narrowing, or training data ethics. But this paper isn’t it. This is theater.

    If anyone needs help understanding what real science looks like, I’ll be here.