YT YouTube2Text

YouTube2Text → Transcripts → AI Does Something Horrifying To Human Thinking

AI Does Something Horrifying To Human Thinking — Transcript

by House of El: AI · 3,709 words · 600 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00College students are rapidly losing the
0:02ability to read. That was one of the big
0:04headlines this week, and if you haven't
0:06seen it, you probably had one of two
0:08reactions. Either, well, obviously AI is
0:11destroying education, or this is
0:13overblown moral panic, or you know,
0:15maybe you had a third reaction, which
0:17was, wait, college students were reading
0:19before this? As someone who has graded
0:21papers at college, the bar was already
0:23kind of on the floor. All of these
0:25reactions, however, skip over the part
0:27that actually matters, which is what the
0:30research says when you sit down and read
0:31it. So, I did. I went through the
0:34studies, the MIT brain imaging data, the
0:36Carnegie Mellon survey of knowledge
0:38workers, the large-scale cognitive
0:40assessments, the counterpoint studies
0:42that say AI can actually improve
0:44learning under the right conditions.
0:47And what I found is more nuanced, more
0:49interesting, and more important than any
0:51headline is going to give you, because
0:53this is ultimately not just a story
0:55about students, but rather a story about
0:58what happens to human cognition, yours,
1:00mine, everyone's, when we start
1:02outsourcing our thinking to machines,
1:05and it has implications that go well
1:07beyond the classroom. My name is L. I
1:09have a PhD in computer science, and I
1:10analyze AI developments to understand
1:12what's actually happening beneath the
1:14hype, which is basically a very polite
1:16way of saying I spend my weekends
1:18reading 80-page PDFs so you don't have
1:20to. You're welcome. Please don't ask
1:22about my social life, though. In this
1:24video, I'm going to walk through what
1:25the science actually says about AI and
1:28how we think, both in children and
1:30adults, the concerning parts and the
1:32hopeful parts, because both of those
1:34exist. Then I want to talk about what
1:36this means for education, for work, and
1:38for the developing minds of the youngest
1:40people in our society. And at the end, I
1:42want to share my vision for how I think
1:44this should be done going forward, and
1:46you can tell me whether you agree with
1:47it or if I got it totally wrong. If you
1:49find this analysis useful, please give
1:51this video a like and subscribe. And the
1:53best way to support this work and keep
1:54it free and without any gatekeeping at
1:56all is through coffee or a channel
1:58membership. The links to everything are
1:59down below. Now, let's start with the
2:01research because the research is where
2:03the nuance lives. It's always academic
2:05papers can be pretty dense, but my job
2:07here is to translate it so we can all
2:09have the same conversation irrespective
2:11of our background knowledge. I won't use
2:14any jargon, I promise. If I accidentally
2:16say back propagation or stochastic
2:18gradient descent, feel free to
2:19unsubscribe, but please don't. Just
2:21judge me in the comments instead.
2:22Anyway, the study that generated the
2:24most attention is from MIT's Media Lab,
2:27a preprint titled Your Brain on ChatGPT
2:30by Cosmina et al. 2025. The researchers
2:33fitted 54 participants aged 18 to 39
2:37with EEG headsets that measured brain
2:39activity across 32 neural regions while
2:42they wrote SAT-style essays. One group
2:45used ChatGPT, one group used Google
2:48Search, one group used nothing, just
2:50their own brains, which in 2026 feels
2:53less like a control group and more like
2:55a dystopian episode of Survivor. The
2:58results were striking. Participants who
3:00used ChatGPT showed the weakest neural
3:02connectivity of all three groups, the
3:05lowest brain engagement across areas
3:07associated with memory, attention,
3:09critical thinking, and creative
3:10reasoning. Search engine users showed
3:13moderate engagement. Brain-only writers,
3:16however, showed the strongest, most
3:18distributed neural networks. 83% of the
3:21AI users could not quote a single line
3:23from the essays they had just written.
3:26They had total cognitive amnesia. They
3:28were basically living out the plot of
3:29The Hangover, but instead of waking up
3:31with a tiger in a hotel room, they just
3:34woke up with a 2,000-word essay on The
3:36Great Gatsby that they had absolutely no
3:38memory of creating. Their brains
3:40basically went into screensaver mode the
3:42moment they hit enter. Now, important
3:45caveats here. This is a preprint. It has
3:47not been peer-reviewed yet. The sample
3:49size is 54, which is very small. A 2026
3:53critique by Stankovich et al. flags
3:55methodological concerns. These are
3:58preliminary findings that need
3:59replication at scale, not settled
4:01clinical results. But the directional
4:04signal is consistent with every other
4:06study in this space, and the finding I
4:08want to highlight is this: In a fourth
4:10session, the ChatGPT users were
4:13reassigned to write without AI. Their
4:15brain activity did not return to normal.
4:18The cognitive debt persisted even after
4:20the tool was removed. So, your brain
4:22essentially goes, "Oh, we're not doing
4:25the thinking thing anymore. Cool, I'm
4:26just going to take a nap in the corner."
4:27And then it just refuses to wake up.
4:30The Carnegie Mellon and Microsoft study,
4:322025, is more robust. 390 knowledge
4:35workers, peer-reviewed, published at the
4:38CHI conference. The researchers analyzed
4:41936 real-world examples of AI-assisted
4:45tasks and found a very clear pattern.
4:47The more confident a worker was in the
4:49AI's ability to complete a task, the
4:52less critical thinking they applied.
4:54Workers who trusted their own expertise
4:56evaluated AI output carefully. Workers
4:59who trusted the AI let go of the wheel,
5:02and they didn't just let go of the
5:03wheel, they climbed into the backseat,
5:05put on a sleeping mask, and trusted a
5:07chatbot to drive them through a
5:09gentrified neighborhood at 60 mph. The
5:12study warns of long-term over-reliance
5:15on the tool and diminished skill for
5:16independent problem-solving.
5:19A larger study of 666 participants found
5:22a strong negative correlation between
5:24frequent AI use and critical thinking
5:26scores. In plain English, a negative
5:28correlation just means a seesaw effect.
5:31As one thing goes up, the other one goes
5:33down. So, the more frequently somebody
5:35used AI, the lower their critical
5:37thinking scores dropped. The mediating
5:40factor was cognitive offloading, which
5:42is basically the tendency to let digital
5:44tools do the thinking for you. The
5:46correlation between cognitive offloading
5:49and critical thinking decline was even
5:51stronger. In a 2026 preprint by Shen and
5:54Tamkin, tested software developers
5:56learning a new coding library.
5:58Developers who fully delegated to AI
6:00produced working code, but performed 17%
6:04worse on conceptual quizzes. They had
6:06the output without necessarily the
6:08understanding. And these were adults
6:11with existing programming expertise.
6:13They had a foundation to fall back on,
6:15and they still lost ground. Now,
6:17consider somebody encountering the
6:19subject for the first time ever with no
6:21existing knowledge to compare the AI's
6:23output against, right? It's also worth
6:25noting that one of the few major studies
6:28that claimed ChatGPT improved learning
6:30performance was retracted last month.
6:32The evidence base is not exactly
6:34symmetric. The studies showing cognitive
6:37costs are accumulating. The studies
6:39showing cognitive benefits are sparse,
6:41often methodologically weaker, and in at
6:44least one prominent case, actually
6:46retracted, which in academic terms is
6:48the ultimate chef's kiss of oops, we
6:50made it all up. But, and this is the
6:53part that gets lost in the alarming
6:54headlines, a 2024 study published in
6:57PNAS found that students who critically
7:00engage with AI, so asking questions,
7:02editing ideas, revising thoughtfully
7:05rather than just copying outputs,
7:07actually performed better and reported
7:09less mental fatigue. So again, the tool
7:12is not the problem, the process is. How
7:15you use AI determines whether it builds
7:17your thinking or replaces it. But
7:20ultimately, this is not a story just
7:22about students. I want to be very clear
7:23about that because I think framing it as
7:25a youth problem lets everyone else off
7:27the hook. The MIT study was conducted on
7:30adults age 18 to 39. The Carnegie Mellon
7:33study surveyed adult knowledge workers.
7:36The cognitive offloading research
7:38measures working professionals. This is
7:40essentially a use it or lose it
7:42phenomenon that applies to everyone
7:44regardless of age. Thinking requires
7:46energy and I want to be honest about the
7:48world we're all operating in. The
7:50economic environment is demanding,
7:52exhausting, and relentless. There are a
7:55bazillion things competing for your
7:56effort and attention every single minute
7:58of every single day. Reaching for a tool
8:01that gives you a cognitive break is not
8:03some sort of a moral failure. It is
8:04profoundly human. Nobody should be
8:07shamed for using AI to lighten a mental
8:09load that was already crushing them
8:11before the technology existed. If you
8:13use ChatGPT to write an email that says,
8:16"Per my last email." in a way that won't
8:18get you sent to HR, you're just
8:20surviving. It's cool. This is a safe
8:22space. And here's a counterpoint that I
8:24think is very important to engage with
8:26honestly. Education Next published a
8:28piece with a very provocative title, "AI
8:31Didn't Destroy Critical Thinking, We
8:32Did." Their argument is that there was
8:35no golden age of critical thinking
8:37before AI came along. NAEP scores have
8:40been stagnant for decades. The
8:42Collegiate Learning Assessment found
8:44college students were academically
8:46adrift and learning almost nothing
8:48across their years of college, and that
8:50research was published well before
8:52ChatGPT existed. AI in this framing
8:55didn't erode critical thinking. It
8:57instead exposed how poorly educational
9:00systems have been teaching it all along.
9:02I think there is a genuine truth in
9:04that, but I also think it doesn't let
9:06anyone off the hook. If anything, it
9:08makes the situation more urgent. If
9:10critical thinking was already fragile,
9:13adding a tool that makes it even easier
9:15to avoid thinking is essentially pouring
9:17water on a chip pan fire. The
9:19pre-existing weakness is not an argument
9:22for complacency. It is an argument for
9:24acting faster. By the way, I did
9:26actually write an ebook on developing
9:28critical thinking, and I'll leave a link
9:29to that below if you want to have a
9:31look. It was written entirely by my own
9:33human brain, which I can prove because
9:35it contains at least three typos and a
9:37deeply questionable metaphor about
9:39cheese. But anyway, the research is also
9:41clear that when you stop exercising
9:43cognitive abilities, they atrophy. The
9:46MIT finding that brain activity didn't
9:48bounce back after AI was removed should
9:51give everyone a pause. Not because it
9:53means you're broken if you use ChatGPT,
9:55but because it suggests there are real
9:57tradeoffs to be aware of. A teacher in
9:59Minneapolis named Maureen Mulvaney
10:01banned all technology from her AP
10:03literature class at Washburn High
10:05School. Phones gone, laptops gone,
10:07everything done with pencil and paper.
10:09In September, 46% of her students felt
10:12confident about their reading ability.
10:14By February, that number was 95%. The
10:17students, after of course initial
10:19resistance, described falling in love
10:22with the analog approach. A Wharton
10:24study on phone bans in college found
10:26that struggling students saw the biggest
10:28grade improvements, roughly double the
10:30average effect. First-year students
10:33benefited most, precisely the group
10:35still forming their study habits.
10:37Sometimes removing the tool is what
10:39rebuilds the muscle, but removing it
10:41permanently is not the answer either.
10:44Where this gets genuinely serious is
10:46with children, small children. There is
10:48a distinction that I think deserves much
10:50more attention than it's getting right
10:51now. A Psychology Today analysis put it
10:54very starkly, adults who offload
10:56thinking to AI lose capacity they built.
10:59Children, however, may never build it at
11:01all. When a developing brain has never
11:03formed independent reasoning
11:06AI doesn't compete with existing
11:08capacity. It replaces the development of
11:10that capacity entirely. At Yale, a
11:13student told CNN that classroom
11:15discussions have become flat and
11:17predictable. She watched a classmate
11:19typing ferociously on a laptop asking
11:21the AI the question the professor had
11:23just asked. Everyone now kind of sounds
11:26the same, she said. When every student
11:28processes information through the same
11:30model, the homogenization isn't just
11:33linguistic, but rather cognitive. The
11:35model's reasoning becomes the student's
11:37reasoning. Its biases become their
11:39defaults. We are accidentally speed
11:42running a dystopian experiment where an
11:44entire generation ends up sounding like
11:46a mildly polite mid-level marketing
11:48executive from San Francisco. And where
11:51is the limit exactly? Where is the line
11:53between a useful educational tool and an
11:56indiscriminate cognitive crutch? Should
11:58a bored toddler spend three hours on a
12:00call with ChatGPT discussing Thomas the
12:02Train? Is that learning? Is that
12:04babysitting? Is that fine? Nobody has a
12:07clear answer, though I suspect Thomas
12:09the Train has some very dark underlying
12:11thoughts about the military-industrial
12:12complex if you ask him correctly. But
12:15just consider this, a 2017 study found
12:17that simply having a smartphone
12:19physically nearby, even face down, even
12:22turned off, reduced available cognitive
12:25capacity. The mere presence of the
12:27device was enough to occupy mental
12:29resources. The phone sits there on the
12:31desk like Lord Voldemort's Horcrux, just
12:34silently draining your soul while you
12:36try to do long division. If a
12:38powered-off phone can measurably affect
12:40your brain, what does an
12:41always-available AI system do to a mind
12:44that is still learning how to think in
12:45the first place?
12:47A few days ago, the UK announced a ban
12:49on social media for under 16s. Snapchat,
12:52TikTok, YouTube, Instagram, X, the
12:54works. This is the furthest any country
12:56has gone, by the way, going even beyond
12:58Australia's landmark legislation from
13:00last year. 20 US states now ban phones
13:04during the school day. France, Spain,
13:06Greece, Slovenia, and South Korea have
13:08all introduced or are working on similar
13:11restrictions. It took a decade of
13:13evidence about what unregulated social
13:15media access does to developing minds
13:17before regulation caught up. The
13:19research on AI and cognitive development
13:21is following the same trajectory, only
13:24faster, because the cognitive impact of
13:26AI is more direct than social media.
13:29Social media captured attention. AI
13:31captures thinking itself.
13:33A RAND Corporation survey from March
13:362026 found that 70% of students
13:39themselves are worried that AI is
13:41eroding their critical thinking skills.
13:43That number grew from 48% to 65% among
13:47high schoolers in just 1 year. When the
13:49people being affected are telling you
13:51they are concerned and the concern is
13:53growing, that is data. And the UK
13:56technology minister Liz Kendall
13:58specifically flagged AI chatbots as a
14:00separate concern, noting that children
14:03were forming one-to-one relationships
14:05with AI systems that were not designed
14:07with child safety in mind. The social
14:09media ban is one regulatory response.
14:12The AI dimension hasn't been addressed
14:14yet, but the pressure is building.
14:17So, what should be done? Because I don't
14:18think the answer is banning AI from
14:20education, and I don't think the answer
14:22is unrestricted access, either.
14:24I've discussed on this channel before
14:26how the body of human knowledge is
14:28expanding constantly. Every field has
14:31more to know than it did a generation
14:32ago. The time it takes to become
14:34competent keeps growing, but the human
14:37lifespan doesn't extend at the same
14:39pace. So, to any anti-aging scientist
14:42watching this video, chop-chop, please.
14:44We're on a clock here. AI-powered tools
14:46for navigating, synthesizing, and making
14:48sense of that expanding knowledge are
14:50becoming genuinely necessary in
14:52research, in science, in medicine, in
14:54education itself. I really believe that.
14:56I've experienced it personally and
14:58actually made an expanded thesis on this
15:00that I'm going to link below if you want
15:01to dig further into that. But, what we
15:03currently have is general-purpose large
15:06language models, ChatGPT, Claude, Grok,
15:09being used as the default tool for
15:11everything. For professional research,
15:13for casual queries, for children's
15:15homework, for clinical questions, for
15:17emotional support, for education at
15:19every level. One tool, every context. No
15:23differentiation whatsoever between a
15:25post-doctoral researcher exploring
15:27quantum chromodynamics and a 12-year-old
15:30working through their first algebra
15:31problem. We are using the exact same
15:34tool to unlock the secrets of the
15:36universe as we are to write a Tinder bio
15:38for a guy named Chad who lists crypto as
15:41a personality trait. We are misusing the
15:43infrastructure.
15:45I think this is where the field needs to
15:46go and I want to say this carefully
15:48because it's a vision rather than a
15:50prescription. Instead of putting
15:52guardrails on general chatbots,
15:54filtering out smart, restricting mental
15:56health conversations, adding safety
15:58warnings exposed, what if the investment
16:00went into developing highly specialized
16:03purpose-built AI applications? An
16:05educational AI designed around the
16:07Socratic method, for example. I'm
16:09personally a huge fan of the Socratic
16:10method as a framework for learning that
16:12develops cognitive abilities by
16:14questioning, challenging, and genuinely
16:16refusing to just hand over the answer.
16:19Just imagine a tool that when a student
16:21asks, "What caused the French
16:22Revolution?" doesn't give them a summary
16:24and instead it asks, "Well, what do you
16:26know about the conditions in France
16:27before 1789? What was happening
16:30economically? What was happening
16:31politically?" Basically, the AI acts
16:33like a tough love personal trainer or
16:36Gordon Ramsay in a kitchen refusing to
16:38serve the dish until you actually chop
16:40the onions yourself and then builds on
16:42whatever the student provides,
16:44correcting misconceptions, filling gaps,
16:46but always making the student do the
16:48thinking. That is a fundamentally
16:50different interaction from here's your
16:52answer, would you like it in bullet
16:54points? Then there's healthcare AI built
16:56specifically for clinical context with
16:58appropriate safeguards. Research AI
17:00designed for synthesis and verification
17:03rather than generation. These things do
17:05exist in early forms. Khan Academy's
17:07Khanmigo is a notable example of this,
17:10but they are nowhere near as prevalent
17:12as the general purpose tools. The
17:14investment, the research energy, the
17:16venture capital, the deployment urgency,
17:18it overwhelmingly flows towards general
17:21LLMs, and that balance, I think, needs
17:23to shift. And critically, different
17:25people learn differently. Visual
17:27learners, kinesthetic learners, verbal
17:29processors, social learners. An
17:31educational AI built with this diversity
17:33in mind would be a fundamentally
17:35different product from a chatbot that
17:37gives the same kind of response to
17:39everyone. The research says that
17:41critical engagement with AI, questioning
17:43it, editing its output, arguing with it,
17:46actually improves learning outcomes. So,
17:48build a tool that forces that
17:50engagement. Let's build a tool that
17:52makes us think harder, not less. The
17:54technology to do this exists, the
17:56research base to inform it exists.
17:58What's missing is the priority, because
18:00we have reached the ultimate holy grail
18:02of this channel, my friends, incentives.
18:05And seriously, I talk about incentives
18:06so much on this channel, I should
18:08probably just tattoo the word directly
18:09onto my forehead. That way, if I ever
18:11pass out from reading too many academic
18:13papers, the paramedics will at least
18:15know exactly what my core personality
18:16trait was. But this requires studies,
18:20real studies with adequate sample sizes
18:22across diverse age groups and
18:24populations with longitudinal
18:26follow-ups, not 54 participants over 4
18:29months, important as that MIT study is
18:31as a starting point. We need hundreds,
18:33ideally thousands, followed over years.
18:36We need studies that track cognitive
18:38development in children who use AI from
18:40an early age versus those who don't. We
18:43need studies that differentiate between
18:45types of AI use, passive consumption
18:47versus active engagement. We need
18:49studies that account for the diversity
18:51of learning styles and neurological
18:53profiles. If we frame this as a further
18:56research is urgently needed, rather than
18:58ban this whole thing, we arrive at a
19:00much better destination for everyone,
19:03irrespective of who makes money or who
19:05pays how much tax. I started this video
19:07with a headline about college students
19:09losing the ability to read, and the
19:11research tells a story that is more
19:13complicated and more interesting than
19:15that headline suggest. AI is measurably
19:17changing how human brains engage with
19:19information. The cognitive cost of
19:22indiscriminate use are real, documented,
19:24and very concerning, especially for
19:27developing minds that may never build
19:28the capacities they're outsourcing. But
19:31AI used critically, actively, as a tool
19:34for challenge rather than a shortcut to
19:36answers, can actually make people better
19:38thinkers. The difference between those
19:41two outcomes comes down to design,
19:43deployment, and intention.
19:45The field of computer science, my field,
19:47needs to take this seriously, arguably
19:49in an interdisciplinary effort with
19:51education, not as a PR problem to manage
19:54with safety filters, but as a genuine
19:57research and development challenge. How
19:59do you build AI that makes humans
20:01smarter rather than lazier? How do you
20:03deploy it in education in a way that
20:05respects the diversity of how people
20:07learn and the vulnerability of
20:08developing minds? How do you do that
20:11while still allowing the technology to
20:12do what it genuinely does well? These
20:15are not rhetorical questions. They are
20:16research questions, and they deserve the
20:19same intensity of funding and attention
20:21that currently goes towards making
20:23models faster, cheaper, and more
20:25capable.
20:26I think about this a lot, and I keep
20:28coming back to the same thought. The
20:29technology is extraordinary. It is
20:31awe-inspiring. The potential is real.
20:34But potential without intentionality is
20:36just capability waiting to be misused.
20:39We saw it with social media, a
20:41technology with genuine capacity to
20:42connect people deployed without
20:44guardrails, optimized for engagement
20:47rather than well-being, and it took a
20:49decade of documented harm before the UK
20:51banned it for children. We are at an
20:54earlier point on the same curve with AI,
20:56and we have the advantage of seeing the
20:58pattern this time. The question now is
21:00whether we act on it, or we wait for the
21:03damage to accumulate and the headlines
21:05to get even worse.
21:06I don't have all of the answers, but I
21:08think if we don't start building AI that
21:10makes us smarter, we're going to wind up
21:12in a world where the machines are doing
21:14the art and the philosophy, while humans
21:16are relegated to doing the dishes and
21:18filling out captchas to prove we still
21:20exist. And frankly, I really don't want
21:22to live in a world where a robot laughs
21:24at my bad jokes or do the dishes for
21:26that matter.
21:27If you want the full context on how AI
21:29is changing education, assessment, and
21:32what a degree even means in 2026, I made
21:34a dedicated video about Princeton ending
21:36its 133-year honor code and it connects
21:39directly to everything we discussed
21:41today. That's the video that I would
21:42watch next. Thank you so much for
21:44watching this one. Subscribe and I'll
21:45see you all on the next one.

About this transcript

This page contains the full transcript of AI Does Something Horrifying To Human Thinking by House of El: AI, generated from the public captions YouTube serves with the video. The transcript has 3,709 words across 600 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.