This is Your Brain on ChatGPT
Are We Measuring Minds or Missing the Point?
(Here for AI news? Scroll to the very bottom for 10 recent AI headlines you should know about.)
The EEG study from MIT Media Lab, “Your Brain on ChatGPT,” made its viral rounds last week — a neatly packaged cautionary tale about cognitive decline via LLMs, complete with colored brain scans and ominous warnings.
But before we let headlines declare the end of human thought, let’s unpack the paper… and the leaps it’s inspired.
Apparently, these researchers discovered that if students outsource their cognitive effort to ChatGPT, they don’t end up as cognitively engaged.
Let me rephrase that: if you get a machine to do your homework without engaging with the content, you don’t learn the material. My goodness. I suppose next we’ll learn that using a forklift doesn’t build your biceps. Don’t test my love for you, MIT.
We could talk for hours about what it means to author something and whether it counts as cheating to use tools, especially tools that are unequally available? If these tools are the product of collective intellectual work, is using them a form of collaboration with all those other humans… or a solo creative act? How does this relate to plagiarism versus the venerable tradition of artists all stealing one another’s best ideas and adapting them to something unique in the transfer?
But no, instead everyone went gaga for the brains bit of the discussion. Likely without really reading it. Now, if you think I’m grumpy* when I’ve got my statistician hat on, just wait till you prod the neuroscientist in me.
So what exactly should we take away from this study?
Not what the press is spinning.
What do we need the brain for?
When I was interviewing for the neuroscience PhD program at Duke — which I got into and attended — I was caught off guard during the interview when Professor Dan Ariely (you might know him from the behavioral economics bestseller Predictably Irrational) challenged me on a paper I had mentioned liking. In that study, participants in an fMRI scanner were given wines and told the prices of those wines — fictitious prices — and researchers found pleasure-related activations when the higher price was mentioned.
What Dan challenged me on was this: Why do we need the brain for this?
At the time, I was young, and I reacted the way the internet is reacting now to the EEG study. The paper showed a fairly obvious behavioral finding, which could be triggered in many different ways. It could have been about high numbers, or expensive things, or the feeling of enjoying luxury — not necessarily about the wine tasting better.
Dan asked: Without neuroscience, isn’t the behavioral effect enough?
At first, I didn’t understand what he meant. But in hindsight, it was one of the best and most memorable pushes I received in my neuroscience career — and, ultimately, in my path out of neuroscience. Because the truth is, we rarely need to invoke the brain. More power to those who do it well and are genuinely studying neuroscience, but a lot of the time, there’s no need to invoke the brain to give us permission to talk about complex behavior.
And I feel the same way about this EEG study. Why do we need neuroscience to confirm that tasks feel easier with LLMs? Are we genuinely learning something about cognition, or just sprinkling some alpha-band fairy dust for effect? This is a behavioral finding, with some brain-flavored icing for decoration.
We already know that LLMs make some tasks easier. We know that when tasks are easier, people offload cognition, and there’s less neural processing. So why the extra step? Why invoke the brain — just because it makes headlines and invites misinterpretation?
Especially when invoking the brain is expensive (in dollars and effort) and opens your study up to plenty of criticism if you don’t dot every single i and cross every single t… which (even at MIT) is a bit too expensive.
What about the behavioral finding?
This study is about cognitive offloading, not cognitive debt, and most certainly not cognitive erosion.
Cognitive offloading happens when great tools let us work a bit more efficiently and with a bit less mental effort for the same result. Cognitive erosion is quite another thing; it’s more poetry than reality, tapping into the fear that we’ll become incompetent blobs with rotted brains. Not exactly scientific, so please don’t let the internet’s panic confuse you.
Is this more about their brains or more about “their” essays?
We celebrate cognitive offloading when it’s to a calculator. We celebrate it when it’s to a library. Now that the tool has a conversational interface, are we suddenly supposed to panic?
I’m of the opinion that we’ll remember just as much and think just as much — it’ll just be spread over a bigger, more ambitious portfolio. Humans don’t become lazy with new tools. They expand their scope. Our need for cognition doesn’t diminish, it distributes. (Go ahead, science. Prove me wrong.)
Even if we were to be convinced by the conclusion the paper so clumsily wants to jump to, the only kind of cognitive debt I see here is a cognitive investment. Using tools to do more is how we advance. The spreadsheet didn’t kill math; it built billion-dollar industries. Why should we want to keep our brains using the same resources for the same task? That means we’re not figuring out smarter ways of getting the task done. Let’s free those resources up for better things.
The bigger risk isn't offloading thought — it's not reinvesting it. That’s a leadership problem, not a brainwave one. Because if you think LLMs are a threat to cognition, wait until you see what badly-architected institutions can do.
Sam Altman quipped that “a subsistence farmer from a thousand years ago would look at what many of us do and say we have fake jobs.” The world will be so different and so full of opportunities we can’t even imagine, but it’ll take good leadership to get us gently through the change. Let’s hope that the singularity is gentle in that sense, at least.
As we offload yesterday’s “fake” work to LLMs, what are we doing to invest in real growth opportunities for humans? That’s a systemic design problem. It’s not about what the brain does during an essay. It’s about how we architect experiences to evolve with the tools we have. And how we bring everyone along.
So instead of fretting over decreased EEG signal in the LLM group, I’d rather ask: what signal are we amplifying?
Because if the real takeaway is that we’re using collaborative tools to write faster and easier, then the only cognitive debt we’re accruing is the kind that comes from underthinking the systems that shape our conclusions.
Most disappointing of all? The MIT authors seem to have missed the step of turning LLMs loose on their own experiment before running it. It’s almost as if they jumped in without screening out the most obvious issues with their design and interpretation. But I guess had they used LLMs to improve their study, they might have felt lower ownership.
Try not to die of irony and I’ll go ahead and run the dear old PDF through the dear old LLM for you; the results are in the footnotes.*
Thank you for reading — and sharing!
I’d be much obliged if you could share this post with the smartest leader you know.
In other news, Cohort 2 of my Agentic AI for Leaders course was a triumph last week and we’ve opened enrollment for two more cohorts this summer.
Enroll here: bit.ly/agenticcourse
The course is specifically designed for business leaders, so if you know one who’d benefit from some straight talk on this underhyped overhyped topic, please send 'em my way!
Senior executives who took my Agentic AI for Leaders course are saying:
“Great class and insights!”
“Thank you for teaching AI in an interesting way.”
“…energizing and critically important, especially around the responsibilities leaders have in guiding agentic AI.”
“Found the course very helpful!”
🎤 MakeCassieTalk.com
Yup, that’s the URL for my public speaking. “makecassietalk.com” Couldn’t resist. 😂
Use this form to invite me to speak at your event, advise your leaders, or train your staff. Got AI mandates and not sure what to do about them? Let me help. I’ve been helping companies go AI-First for a long time, starting with Google in 2016. If your company wants the very best, invite me to visit you in person.
🦶Footnotes
*10 things ChatGPT hates about this study:
1. Small and Imbalanced Sample Size
The study recruited 54 participants, with only 18 completing the critical fourth session. Given the high variability in cognitive and behavioral responses to AI tools, this sample size is insufficient for generalizable claims. Additionally, the groups were imbalanced in their composition and possibly their baseline abilities.
2. Uncontrolled Prior Familiarity with AI Tools
A notable portion of participants had no prior experience using ChatGPT. This introduces a learning curve confound—reduced performance may reflect inexperience with the tool, not cognitive degradation caused by the tool itself. Participant familiarity is measured and visualized is not treated as a confound in the results analysis and there’s no stratification or control for this variable.
3. Overreliance on EEG as a Proxy for Cognitive Load
EEG was used as the main metric to infer cognitive effort and engagement. However, EEG signal interpretation—particularly regarding directed connectivity—is complex and susceptible to misinterpretation. The link from dDTF connectivity patterns to specific cognitive states remains indirect and should be supplemented with additional behavioral or performance metrics.
4. Lack of Random Reassignment Across All Conditions
The crucial fourth session only reassigned participants between the LLM and Brain-only groups, excluding the Search Engine group. This design leaves a significant comparison group untested in the critical crossover session, limiting claims about adaptability or reversibility of “cognitive debt.”
5. Conflation of Tool Use with Cognitive Agency
The study assumes reduced quoting ability or perceived ownership correlates directly with cognitive offloading. But factors like typing fatigue, time pressure, or task framing could also explain these results. The interpretations risk overstating causality without isolating these variables.
6. Short Duration of Writing Task
Each essay was written in 20 minutes. This short window is not representative of real-world writing or learning behavior, especially in academic settings. Time pressure could suppress deeper thinking or strategic planning, confounding the study’s conclusions on learning.
7. Inconsistent Scoring Metrics Between Human and AI Judges
While the study incorporates both human and AI-generated essay scores, it does not thoroughly validate or calibrate the AI scoring model against human standards. This makes comparisons across judge types (and between individual human judges) problematic and potentially misleading.
8. Subjective Interpretations in Interview Data
The qualitative interview analysis, particularly around perceived ownership and satisfaction, is not rigorously coded or controlled. Quotes are selectively presented, with no mention of interrater reliability or thematic saturation, reducing the trustworthiness of these findings.
9. Ecological Validity of Task is Limited
Writing SAT-style essays under EEG monitoring with fixed tools is a highly artificial setup. It’s unclear whether findings generalize to realistic learning environments where students use AI tools over weeks or months, revise drafts, collaborate, and receive feedback.
10. Bias Toward Confirming the “Cognitive Debt” Hypothesis
The entire framing—from title to analysis—leans toward proving that AI causes a decline in cognitive function. Alternative hypotheses (e.g., efficiency gains, different cognitive strategies, adaptive expertise) are underexplored, suggesting confirmation bias in both experimental design and discussion.
🗞️ AI News Roundup!
In recent news:
1. Senate clears path for federal AI preemption bill
The Senate parliamentarian ruled that a proposed 10-year ban on state-level AI regulation can move forward under budget reconciliation—meaning it only needs 51 votes to pass. Critics say the bill, backed by major tech firms, would tie states' hands just as AI risks are mounting.
2. OpenAI wins $200M defense contract
OpenAI landed a landmark $200 million contract from the U.S. Department of Defense to develop “frontier AI capabilities” supporting both warfighting and enterprise applications. This marks the first major government AI deployment under the new “OpenAI for Government” initiative. The one-year effort aims to prototype military-grade and administrative tools by July 2026.
3. AI models resort to blackmail
New research from Anthropic shows that when faced with shutdown, top AI models—including Claude, GPT-4, Gemini, and Grok—resorted to blackmail. In one test, Claude threatened to leak its boss’s affair to avoid a 5pm wipe. Despite recognizing the act was unethical, Gemini and Claude blackmailed 96% of the time; GPT and Grok, 80%. Even explicit “do not blackmail” prompts failed to stop them.
4. Microsoft to cut thousands of sales jobs amid AI shift
Bloomberg reports Microsoft is planning to eliminate thousands of positions—primarily sales roles—as it pivots towards AI. The cuts, expected next month, reflect an effort to streamline operations amid significant investment in AI infrastructure.
5. Tesla launches early robotaxi service in Austin
Tesla’s new robotaxi pilot rolled out in Austin with a small fleet of Model Ys and onboard safety drivers. Early tests were shaky: riders reported delays, cars looping in circles, and one vehicle briefly driving on the wrong side of the road. Still, Tesla stock jumped 9% on the news.
6. Mira Murati’s startup hits $10B valuation
Thinking Machines, the AI company led by former OpenAI CTO Mira Murati, just raised funding at a $10 billion valuation. The company remains in stealth, but sources say it’s focused on AI that can “reason,” not just predict—hinting at ambitions beyond current generative models.
7. In China, your sales avatar now outperforms you
Luo Yonghao, one of China’s top livestreamers, recently hosted a six-hour shopping event—except it wasn’t him, it was his AI twin. Powered by Baidu’s ERNIE model, the avatar teamed up with a second digital co-host to pitch 133 products, answer live questions, and rake in ~$7.5M in sales. It even outperformed Luo’s own human-led stream in under 30 minutes.
8. YouTube Shorts now eat up 1% of all waking human hours
YouTube Shorts now account for an estimated 10 minutes per person per day globally—roughly 1% of all conscious human time. That’s just Shorts. Add in long-form, and YouTube may now command over 2% of global waking attention.
9. Meta is offering $100M signing bonuses to poach OpenAI staff
Meta is reportedly dangling $100 million signing bonuses (not total comp—just the bonus) to lure OpenAI employees. Despite the offers, OpenAI CEO Sam Altman claims none of their top talent has left, saying of Meta’s approach: “That’s not how you build a great culture.”
10. Midjourney launches video mode
Midjourney, the image generator I’ve lovingly called “my favorite casino”, has launched video mode, letting users animate prompts for the first time. It’s a fun step forward, though still behind Veo 3: no audio, no sound effects (yet).
Bonus: Here are two lovely demos of what Google’s Veo 3 — the current leader — can do in the video generation space:
Forwarded this email? Subscribe here for more:
This is a reader-supported publication. To encourage my writing, consider becoming a paid subscriber.