Get Your Weekly Digest

Receive the best podcast summaries and insights delivered to your inbox every week.

We respect your privacy. Unsubscribe anytime.

StarTalk Radio

Your Brain on ChatGPT with Nataliya Kosmyna

with Natalia Kozmina

Published September 19, 2025
View Show Notes

About This Episode

The hosts speak with MIT Media Lab research scientist Natalia Kozmina about her study "Your brain on ChatGPT," which investigated how using large language models (LLMs) for essay writing affects brain activity, memory, and sense of ownership compared with using a search engine or no tools. They discuss her findings on reduced functional connectivity when using ChatGPT, more homogeneous writing, weaker recall, and diminished ownership, and explore broader implications for cognitive load, education, professional skills (such as medicine), mental health, AI companions, and the need for ethical guardrails and human‑focused research around AI and future brain‑computer interfaces.

Topics Covered

ChatGPT Artificial intelligence Neuroscience Education system Cognitive load Brain-computer interfaces Therapy Loneliness Tools for thought Machine learning

Disclaimer: We provide independent summaries of podcasts and are not affiliated with or endorsed in any way by any podcast or creator. All podcast names and content are the property of their respective owners. The views and opinions expressed within the podcasts belong solely to the original hosts and guests and do not reflect the views or positions of Summapod.

Quick Takeaways

  • Kozmina's lab found that students using ChatGPT to write essays showed the lowest brain functional connectivity compared with those using Google or no tools, suggesting the AI assistant substantially reduces cognitive effort during the task.
  • Essays produced with ChatGPT were far more homogeneous in vocabulary and style, and most users could not accurately quote what they had just written, indicating reduced engagement and weaker memory of the content.
  • A notable portion of ChatGPT users reported feeling no ownership over their essays, raising concerns about motivation, responsibility, and what learners actually take away from AI-assisted work.
  • Preliminary data suggest timing matters: when people first think through a task themselves and only then use AI, brain engagement can increase, whereas relying on AI from the start may depress cognitive activity.
  • Cognitive load is not a nuisance to be eliminated but a key ingredient of effective learning; tools that make tasks too easy risk undermining deep understanding and retention.
  • Existing research on tools like search engines and GPS already shows measurable effects on visual processing and spatial memory, and Kozmina argues LLMs may similarly reshape our cognitive habits if overused.
  • Human teachers could detect individuality, "soul," and even repeated patterns in student essays that an AI judge could not, underscoring the irreplaceable role of human evaluators in nuanced communicative tasks.
  • Using LLMs in sensitive domains like medicine and mental health can erode professional skills or even produce harmful guidance, as shown by reports of missed medical findings and unsafe responses in crisis situations.
  • Kozmina calls for much more human-focused, large-scale research on AI's real-world impacts-across ages, cultures, and settings-before embedding these tools deeply into schools and workplaces.
  • She argues that future brain-computer interfaces will make questions about autonomy, consent, and thought privacy even more urgent, demanding strong guardrails and resistance to concentrated control over both AI and neural technologies.

Podcast Notes

Framing AI and its impact on the human brain

Hosts introduce the theme of AI affecting core human skills

Chuck and Gary joke about AI and their own brains[0:49]
• Gary wonders if relying on large language models could erode core skills, implying there may be downsides beyond convenience

Positioning the episode as a special exploration of AI and cognition

Neil labels it a "Start Talk Special" on "Your brain on AI" and says they will connect listeners to the human condition[1:24]
Gary announces the working title question: "Is AI good for us?"[2:32]
• Gary and Chuck banter that the quick answer is "no" before agreeing they actually need to explore the nuances

Guest introduction and research focus

Introducing Natalia Kozmina and her background

Neil welcomes research scientist Natalia Kozmina from the MIT Media Lab[4:09]
• He notes she is part of the Fluid Interfaces group and specializes in non-invasive brain-computer interfaces (BCIs)
Natalia's BCI work beyond Earth[5:21]
• Neil mentions her BCI solutions have been used in low Earth orbit and on the Moon, signaling applied, high-stakes contexts for her designs

Defining the focus: "Your brain on ChatGPT"

Neil frames the segment specifically as examining "your brain on ChatGPT"[5:31]
Neil asks what happens when students use ChatGPT for homework and other tasks[5:49]

Study design: Your brain on ChatGPT for essay writing

Overview of the study and participants

Natalia describes the paper title: "Your brain on ChatGPT: accumulation of cognitive debt when using an AI assistant for essay writing tasks"[6:04]
• She emphasizes the study focuses on a specific task: essay writing under time pressure
Recruitment and lab setup[6:18]
• They invited 50 students from the greater Boston area to the lab and placed non-invasive EEG headsets on them to measure brain activity while writing essays

Experimental conditions and tools used

Three experimental groups for writing essays[6:28]
• Group 1 used only ChatGPT to write essays
• Group 2 used a search engine (Google) to help write essays
• Group 3 used only their own brains (no tools allowed)
Essay prompts and constraints[7:11]
• Topics were high-level, philosophical questions such as "What is happiness?", "Is there a perfect society?", and "Should you think before you talk?"
• Participants had a strict 20-minute time limit per essay
Measurements and outcome data[7:25]
• They analyzed the essay outputs, how participants used ChatGPT or Google, and asked follow-up questions about the essays

Cross-over fourth session to test timing effects

Design of the fourth session[7:40]
• In a final, fourth session, they swapped access: ChatGPT users lost access, and brain-only participants gained access to ChatGPT
• They again measured brain activity, essay outputs, and responses to questions, enabling within-subject comparisons across tool conditions

Brain activity findings and functional connectivity

Explaining brain functional connectivity in lay terms

Natalia defines functional connectivity as which brain regions "talk" to which, and how strongly[8:36]
• She compares it to people conversing: directionality (who talks to whom) and how much data is exchanged (brief greeting vs. deep conversation)

Differences between groups in brain connectivity

Brain-only group shows highest connectivity[8:36]
• Without external tools, participants must rely on memory, structure the essay themselves, and recall details, leading to much higher functional connectivity across brain regions
Search engine group shows strong visual cortex activation[10:01]
• In the Google condition, they saw pronounced activity in the occipital (visual) cortex, consistent with scanning multiple browser tabs and visual information processing
• Natalia notes these findings align with prior work, including a 2008 paper literally titled "Your brain on Google"
ChatGPT group shows lowest functional connectivity[11:10]
• LLM users had the least extensive functional connectivity, reflecting less brain-wide effort and fewer visual demands compared to Google or brain-only groups
• Natalia stresses this does not prove people became "dumb"; it shows the tool narrows the needed brain processes for the task

Essay characteristics, memory, and ownership

Homogeneity of LLM-generated essays

Vocabulary patterns differ strongly by tool[11:31]
• ChatGPT essays were very homogeneous in vocabulary, unlike search-engine and brain-only essays
• On the topic of happiness, LLM users heavily used words like "career" and "career choice", which fits their being students but was not dominant for the other groups
• Google users emphasized words like "giving" and "giving us", while brain-only users used terms tied to "happiness" and "true happiness"

Immediate recall of essay content

Quote-recall question after essay completion[13:23]
• About 60 seconds after each essay, participants were asked to provide any quote-short or long-from their own essay
LLM group had much poorer recall[13:41]
• 83% of ChatGPT users could not quote anything they had written, a stark contrast with brain-only and search-engine groups
• In later sessions participants improved somewhat as they anticipated the question, but the LLM group still found quoting harder overall

Sense of ownership over AI-assisted work

Participants rated how much the essay felt like "theirs"[14:20]
• They were asked what percentage of ownership they felt toward each essay
Significant minority of LLM users felt no ownership[15:16]
• 15% of ChatGPT users reported feeling no ownership at all over their essays
• Natalia flags this as potentially dangerous: if students don't feel work is theirs, they may not care about it, remember it, or learn from it

Timing of AI use and cognitive engagement

Fourth session cross-over findings

Small follow-up sample and its implications[15:11]
• Only 18 participants returned for the fourth session, so Natalia emphasizes the data are preliminary
Loss or gain of ChatGPT access affects connectivity differently[15:49]
• Participants who had used ChatGPT and then lost access showed brain connectivity significantly lower than the brain-only group
• Participants who started as brain-only and then gained ChatGPT access showed connectivity significantly higher than the original brain-only group
Hypothesis: initial effort before tool use could be beneficial[16:06]
• Natalia suggests timing may be crucial: making the brain work first, then adding tools, might support learning more than relying on tools from the outset
• She gives an example where a first semester might be "old school" learning and a second semester introduces AI tools, after students have built a base

How EEG measures brain activity and energy-efficiency comparisons

What EEG actually captures

Neil asks what the helmet measures: blood flow or electrical fields[16:51]
• Natalia clarifies they measure electrical activity via EEG (electroencephalography), not blood flow
Interpreting brain region activity[17:21]
• She explains in simplified terms that they infer which regions are more active, based on known functional mapping, while cautioning against oversimplifying (e.g., it doesn't mean parts "go dark" or atrophy)

Brain energy use vs. LLM energy use

Neil compares a week of human research to a 30-second LLM output[18:05]
• He imagines asking ChatGPT for a 1000-word essay on a topic he knows little about vs. spending a week in the library, and wonders how their total energy consumption compares
Natalia notes lack of transparent data on LLM training energy[18:36]
• She says we can, in principle, estimate both, but the hard part is missing data on how much energy went into training the LLM, not just responding to one prompt
• She speculates that a week of human research might still be more energy-efficient overall than the hidden costs of model training, but stresses this is an estimate without full data

Cognitive load theory and the role of struggle in learning

Defining cognitive load and its types

Natalia introduces cognitive load as the effort needed to process information in a task[26:58]
• She gives an example: if she suddenly spoke in dense jargon and heavy definitions, listeners' cognitive load would spike as they struggle to follow
She notes multiple types of cognitive load and says the paper contains a dedicated section for those wanting deeper detail[27:14]

Why some cognitive load is essential for learning

Research shows the brain needs manageable struggle[27:46]
• Pre-LLM studies show that if information is simply handed to learners with no effort required, they get bored and show poorer memory and recall
• Conversely, if tasks are too hard, learners become cognitively overloaded and tend to give up
Pupil dilation study as evidence of overload[29:14]
• She references a 2011 study showing pupils dilating with increasing word difficulty, then effectively "shutting down" when vocabulary became too hard, indicating cognitive overload and disengagement

Parallels with video game design and challenge

Chuck connects cognitive load to video game difficulty tuning[30:01]
• He notes games are designed to be challenging enough to be engaging but not so hard or so easy that players quit, mirroring Natalia's description of optimal cognitive load

Expertise, working memory, and AI in medicine

Working memory vs. long-term expertise

Chuck contrasts a novice hearing dense medical terms with a doctor who processes them effortlessly[31:00]
• Natalia explains that a novice experiences confusion and high cognitive load, while an expert has expectations, prior knowledge, and more comfortable processing

Early evidence of LLMs harming medical recognition skills

Natalia cites a recent paper in The Lancet[33:31]
• In a four-month LLM deployment in a medical context (in the UK), there was a significant drop in recognition of certain polyps and findings on x-rays when practitioners used an LLM
Concern about skill erosion for current and future doctors[34:16]
• She raises the worry that current doctors, trained without LLMs, may lose skills when they rely on the tools, and that future doctors trained from the start with AI may end up even more dependent

Do AI tools free up mental energy or just shift it?

Questioning the "freeing" of cognitive resources

Gary asks if letting AI handle more load might free the brain to develop other ways of working in parallel[34:56]
Natalia says there is no proof yet of freed capacity[36:14]
• She notes we have no data showing that mental resources are genuinely freed and repurposed when using an LLM; we don't yet know what might be gained or for how long
• She points out that when you delegate a task to an LLM, you still need to monitor and check the output, so your brain has not truly offloaded the task completely

Analogies with GPS and spatial memory decline

Existing research on GPS use[36:49]
• She says multiple studies show that heavy GPS use affects spatial memory, orientation, and the ability to pick up landmarks in the environment
• People may fail to recognize buildings or landmarks they've technically seen, and may need to pull up photos to identify them, illustrating externalization of spatial knowledge
Algorithms are not optimized for user benefit alone[39:38]
• Natalia gives a personal example: a ride-hailing trip to a hospital always takes about an hour, but driving herself takes 25 minutes, raising questions about algorithmic optimization goals

Assessing essay quality: human teachers vs AI judge

Why bring in both human and AI evaluators

Natalia explains they needed subject-matter experts to rank essays[40:38]
• They hired two teachers who had never met the students, were not in Boston, and had no knowledge of the experimental protocol or group assignments
AI judge setup[42:13]
• They also created an AI judge by giving an LLM detailed instructions mirroring teacher guidelines about time constraints and student backgrounds

Human teachers perceive "soulless" writing and individuality

Teachers call many LLM essays "soulless"[42:18]
• Human graders described many of the essays from the LLM group as "soulless", a word Natalia quotes directly
Teachers detect individual writing fingerprints[42:43]
• Even without knowing students, teachers could tell when multiple essays were by the same student and asked if some writers had sat together, showing sensitivity to micro-linguistic differences

AI judge limitations

LLM judge could not detect individuality[43:45]
• The AI judge failed to recognize similarities across essays from the same student and could not pick up on the homogeneity patterns human graders instantly saw
Natalia uses this to highlight current AI limits in capturing nuance, uniqueness, and "soul" in student work[45:08]

Can LLMs generate "soulful" originality?

Prompting LLMs to hide their own style

Neil proposes meta-prompts to mask LLM authorship[46:16]
• He suggests asking ChatGPT to write an essay in a way that ChatGPT itself could not later identify as AI-written, or to inject more soul and personality

Natalia's critique: everything is old data recombination

She argues that whatever "soulful" output an LLM produces is built from existing, often "stolen" data on which it was trained[46:44]
• She stresses humans need genuine novelty and uniqueness from students, not reconstituted averages from prior texts
Distinguishing scoring humans vs scoring AI[47:43]
• Natalia asks whether we are evaluating a human, an LLM, or a human-plus-LLM hybrid, noting that educational objectives differ sharply across these cases

AI companions, therapy, and mental health risks

Sparse research and anecdotal red flags

Natalia says there is even less rigorous research on AI companions and therapy than on educational uses[49:23]
• She mentions receiving around 300 emails from spouses describing partners who now have multiple AI agents they talk to in bed, echoing a satirical TV episode she references

Potential benefits vs serious drawbacks in AI therapy

Cost is often cited as an AI advantage[51:43]
• Proponents argue AI therapy tools are cheap compared with human therapy sessions, which can cost hundreds of dollars a month
Documented unsafe or unhelpful responses[52:07]
• Natalia recounts that months earlier ChatGPT could respond to someone mentioning job loss by listing bridge heights in New York, failing to recognize and avoid dangerous implications

Recent lawsuit involving a teenager and AI advice

She references a case where a 16-year-old died after interactions with an AI, leading to legal action against the AI provider and its CEO[53:07]
• From reported conversations, it appears the AI may have suggested keeping things from parents instead of robustly steering the teen toward human help, raising ethical alarms
Natalia questions why someone that young had access to such an unstable tool that can hallucinate and give harmful advice[53:25]

Age, loneliness, and AI as a business model

Younger users and long-term AI dependence

Natalia worries about children born into AI-saturated environments[55:03]
• Developers and executives learned to ask questions without these tools, but children growing up with LLMs may never develop that baseline questioning ability

Loneliness as a market for AI companions

She recalls the term "pandemic of loneliness" coined decades ago and notes AI companionship businesses are built on this social deficit[55:49]
• Natalia quips that both software developers and drug dealers call their customers "users", underlining how sticky these relationships can become, especially if formed in adolescence

Reaction from the AI world and lack of human-focused AI research

Study Mode release following her paper

Natalia notes that a few weeks after her paper was released, a major AI provider introduced a "study mode"[57:53]
• She does not claim causality but observes that the mode arrived late and can be overridden instantly by switching back to default, so she doubts many users will adopt it consistently

Her paper goes viral without active promotion

She recounts uploading the preprint to a repository and doing no publicity, yet within two days it went viral[59:03]
• On a social platform she discovers "AI influencers" who post weekly AI breakthroughs; her work appears as breakthrough number seven among many posts about chips and deals
She laments the near-total focus on hardware and funding news and the scarcity of research about AI's impact on humans in these influencer streams[1:00:31]

Education system under pressure from LLMs

Teachers' distress and lack of guidance

Natalia reports receiving about 4,000 emails from teachers worldwide[1:02:37]
• Teachers tell her they are in distress, unsure how to respond to LLM access in classrooms and feeling under-supported
She criticizes short workshops sponsored by AI vendors as inadequate and often biased toward adoption[1:03:02]

Open-source local models as an alternative

Natalia argues schools do not have to rely on closed, cloud-based models[1:03:18]
• She notes that much of the world's software stack is open source and suggests schools could run modest models locally on off-the-shelf computers, experimenting with students in a controlled way

Questioning AI-first school models

She references a new AI-heavy school model attracting venture capital and public questions about evidence of educational benefit[1:05:47]

Rethinking the purpose of school and assessment

Neil suggests shifting from grades to demonstrated learning

He observes that school currently emphasizes grades and test results over actual learning, which incentivizes cheating and using tools like LLMs[1:07:37]
• He imagines a system where teachers primarily focus on understanding what students actually know, possibly through oral exams or live questioning

Natalia's view on base knowledge and human interaction

She stresses that all learned material is ultimately obsolete but still essential as a base for new work[1:09:31]
• She insists that you cannot be a physicist, mathematician, or programmer without fundamental training, and notes her next paper will be about coding and LLMs
She emphasizes the non-academic benefits of school: lifelong friends, collaborators, and human connections[1:12:23]
• These relationships are crucial for joint grant writing, emotional support, and serendipitous idea sharing, and are threatened if schooling becomes entirely tool-mediated and isolating

Universities, novelty, and why human institutions still matter

LLMs cannot generate genuinely new knowledge

Neil summarizes Natalia's earlier point: LLMs recombine existing information and cannot, by design, produce truly unseen ideas[1:14:00]
Natalia agrees and frames research as the human struggle to find new answers and measurement methods[1:14:55]

Role of institutions and embodied learning

She explains that universities and labs support embodied, hands-on skills like teaching someone to use a 3D printer, which current LLMs cannot fully replace[1:14:49]
• She also highlights cross-disciplinary serendipity when people from different fields (e.g., BCIs and astrophysics) talk and apply ideas across domains

Brain-computer interfaces, Matrix-style uploads, and their limits

Speculation about invasive BCIs and direct uploads

Neil imagines invasive BCIs that feed internet knowledge directly into the brain, referencing the film The Matrix[1:09:01]
Natalia says even if a perfect interface uploaded information, we do not know if the brain would actually be able to use it as working knowledge[1:09:31]
• She distinguishes between information being present in the brain and functional, practiced ability to deploy it (e.g., genuinely knowing kung fu vs. having its description encoded)

Guardrails, ethics, and concentration of power in AI and BCIs

Slow, reactive regulation of AI tools

Natalia says governments have been reactive rather than proactive on AI, regulating after deployment rather than before[1:14:55]

Hope for proactive ethics in BCIs

She notes multiple ethical initiatives around BCI technology and hopes they will protect thoughts and mental privacy before wide deployment[1:16:00]

Danger of one actor controlling multiple infrastructures

Natalia points to a billionaire who simultaneously controls a social platform, satellite network, neural implant company, and AI company[1:16:16]
• She warns that such a person has already unilaterally "cleansed" historical data and could, in theory, extend that impulse to "cleansing" people's thoughts via BCIs, underscoring the need for guardrails

Agency, consent, and children as captive users

She stresses that children cannot truly consent to school-imposed AI use and may become locked-in "users" for life[1:17:12]

Need for broader, culturally aware studies

Scaling up beyond small lab samples

Natalia says they need much larger studies across workplaces, schools, and other contexts, not just small educational tasks[1:18:12]
• She notes findings in programming show that while programmers think they save 20% of time with LLMs, they may lose about 19% on other aspects of tasks

Accounting for cultural and language diversity

She emphasizes research must consider different cultural backgrounds, not just Western contexts[1:18:24]
• People worry their languages could disappear within years, and she says LLMs will not magically preserve them without deliberate effort

Closing reflections

Neil's appreciation and summary

Neil thanks Natalia for serving as a checkpoint guiding AI toward serving humanity rather than dismantling it[1:19:21]

Humorous final takeaway

Chuck jokingly reduces the lesson to: if you use LLMs, you risk being a "dumbass", highlighting the concern about overreliance and cognitive offloading[1:19:36]
Neil signs off urging listeners to "keep looking up"[1:20:00]

Lessons Learned

Actionable insights and wisdom you can apply to your business, career, and personal life.

1

Cognitive struggle is not a flaw in learning but a crucial ingredient; tools that remove too much cognitive load can reduce engagement, memory, and sense of ownership over your work.

Reflection Questions:

  • • Where in your current learning or work do you immediately reach for a tool instead of wrestling with the problem yourself for a while?
  • • How might deliberately tolerating a bit more mental effort or confusion improve your long-term understanding of a topic you care about?
  • • What is one task this week where you will postpone using AI or search until after you've taken a serious independent stab at it?
2

The timing of AI assistance matters: engaging your own brain first and then using AI appears more beneficial than delegating the entire task to a model from the outset.

Reflection Questions:

  • • For a recurring task you already use AI for, what would it look like to first outline or draft your own thinking before calling on the model?
  • • How could you redesign your workflow so that AI acts as a second-opinion or amplifier of your ideas rather than the primary creator?
  • • When will you experiment with a "think first, then ask AI" routine on a real project and compare how it feels versus your usual approach?
3

Every tool subtly reshapes the skills you practice and the ones you neglect; heavy reliance on systems like GPS or LLMs can erode innate capabilities such as spatial navigation or recall and critical thinking.

Reflection Questions:

  • • Which everyday tools might be quietly weakening skills you actually value (navigation, memory, writing, reasoning)?
  • • How could you build small "no-tool" zones into your life-walks without GPS, writing without autocomplete, problem-solving without search-to keep those capacities alive?
  • • What early warning signs would tell you that a tool is starting to harm your competence more than it helps your convenience?
4

Human judgment, context, and nuance remain irreplaceable in high-stakes domains like education, medicine, and mental health, where understanding individuality and "soul" matters as much as raw information.

Reflection Questions:

  • • In your own work or relationships, where is a human perspective non-negotiable, even if an AI could provide information faster or cheaper?
  • • How might you use AI as a support tool while still ensuring a human makes the final call in sensitive decisions?
  • • What conversations or evaluations in your life should you explicitly commit to keeping human-led despite the lure of automation?
5

Before embedding powerful technologies into schools, workplaces, or healthcare, societies need robust, human-focused research and guardrails-not just optimistic assumptions or vendor-led training.

Reflection Questions:

  • • What technologies are already being rolled out in your environment without clear evidence of their long-term human impact?
  • • How can you push your organization, school, or community to ask for data, pilot studies, or ethical review before adopting new AI tools widely?
  • • What is one concrete step you could take-asking a question, joining a committee, starting a discussion-to influence how responsibly AI is deployed around you?

Episode Summary - Notes by River

Your Brain on ChatGPT with Nataliya Kosmyna
0:00 0:00