FIELD NOTES / RESEARCH UPDATE
October 2026
WORKING CLAIM — HOLD THIS LIGHTLY
AI is not making us dumber.
It's changing where the thinking happens.
That is not the same thing.
And I need to be careful not to turn that into a story before the evidence earns it.
THE OLD PATTERN
Humans have always moved thinking out side of the skull.
Writing → memory onto paper.
Maps → navigation into the environment.
Calculators → arithmetic into a tool.
Search → retrieval into a system. So this isn't new, nothing scandalous about this.
It's one of our best tricks.
Don’t carry everything.
Build something that carries some of it for you.
That's basically the logic already sitting underneath the “good tools pay rent” idea:
a tool should take something the brain would otherwise have to remember / track / sequence / check / coordinate and carry it outside the skull.
When that works:
less clutter on the workbench.
more attention left for judgment + action.
Good trade.
AI is the same move, except it reaches further up the stack.
Not just:
store/ retrieve/ calculate/ navigate. Now:
summarise/ compare/ synthesise/ draft/ generate/ frame/ evaluate/ recommend. Potentially even:
decide what matters.
That's the bit I care about.
[IMAGE 1 — WHERE THE THINKING MOVES]
RECEIPT #1 — “AI MAKES THE BRAIN SWITCH OFF” IS TOO SIMPLE
The MIT “Your Brain on ChatGPT” study is the one everybody quotes. Useful. Interesting. Also: preprint. Not peer reviewed. 54 people in the first sessions, 18 in the final crossover. It found lower EEG connectivity in the LLM condition during essay writing. That does NOT equal: less brain less intelligence brain rot damage.
Peer-reviewed work since then makes the clean headline fall apart. Zhang et al., Acta Psychologica, Aug 2026. 34 undergrads. ChatGPT teacher vs human teacher. Retention: roughly similar. Transfer to a new problem: worse with ChatGPT. EEG: broadly HIGHER theta activity with ChatGPT. So: more measured activity ≠ better learning.
Another 2026 creativity study — Jin et al. 95 people. AI teaching vs human teaching vs control. AI improved originality + flexibility, including transfer. Different neural signatures again. So stop treating “brain activity” like the rev counter on intelligence. More can mean: more useful processing or more struggle or a clunky interface or compensation. Less can mean: efficiency or disengagement or offloading or a different route. Brain scan alone doesn’t answer the question.
Better question: WHAT COULD THE PERSON DO AFTERWARDS?
RECEIPT #2 — OUTPUT AND CAPABILITY ARE SPLITTING APART
This might be the biggest thing. Dell’Acqua et al. / BCG / Organization Science, 2026. 758 knowledge workers. Inside GPT-4’s capability frontier: +12.2% more tasks ~25% faster higher quality. Then a task just outside the frontier: AI users were ~19 percentage points less likely to get it right. Same tool. Same smart people. Different task. Different outcome. => “uses AI” is basically a useless category by itself. Need to know: for what? where? how? with what verification? inside whose capability frontier? Then the learning work. Melumad + Yun, PNAS Nexus, 2025. 7 experiments. 10,462 people. LLM answers vs normal web search. LLM group got the information. But later: shallower reported knowledge sparser advice less original advice less likely to be adopted. Likely mechanism: search makes you FIND + CHOOSE + COMPARE + SYNTHESISE. LLM can hand you the synthesis already assembled. The information did not disappear. Part of the cognitive construction did. Important distinction.
Then Wong + Qiu, Educational Psychology Review, Apr 2026. 196 students. 3 groups: 1) human only 2) free ChatGPT 3) THINK FIRST, ChatGPT second Immediate creative output: free ChatGPT looked best. Take AI away: free-ChatGPT advantage basically disappears. Think-first group: better later independent creativity. Same AI. Different division of cognitive labour. Different human left behind afterwards. That is the line. Performance-with-AI ≠ capability-in-human.
[IMAGE 2 — PERFORMANCE NOW VS CAPABILITY LATER]
WHERE IS THE THINKING MOVING?
Need a better map than “offloading”. At least four layers.
1. STORAGE / RETRIEVAL
Old territory. Remember less. Find more. Probably fine in a lot of cases.
2. CONSTRUCTION
Who turns pieces into a model? Who does the comparison? Who resolves contradictions? Who builds the explanation? LLMs can do this before I ever see the raw material. Potential upside: massive time saved. Potential cost: I receive the finished structure without building enough internal structure of my own.
3. GENERATION/CREATION
The blank page moves. First idea. First hypothesis. First frame. First outline. First interpretation. First code. First answer. Wong/Qiu matters here: human generates first → AI extends = different outcome from AI generates first → human edits. Need more research here.
4. EVALUATION / CALIBRATION
Most important layer? Fernandes et al., Computers in Human Behavior, 2026. 2 studies. 246 + 452 people. AI improved LSAT-style reasoning performance. But users were poor at judging how well they had actually done. In study 1: ~3 more questions correct but ~4 questions of overestimation. And weirdly: higher AI literacy linked with worse self-assessment accuracy. This is the danger zone. Not: “AI answered for me.” But: “I no longer know when the answer is dodgy.” That is steering. If we outsource generation but retain judgment, maybe okay. If we outsource judgment as well, who is driving the fucking thing?
COGNITIVE LOAD — BE CAREFUL WITH THE WORD “LESS”
Everyone likes: “AI reduces cognitive load.” Okay. WHICH load? Qian et al., Frontiers in Psychology, 2026. Systematic review. 39 empirical studies. 28 experimental. Results all over the place: some lower load some no difference some higher some nonlinear Main problem: most studies do not cleanly separate useless load from learning-related load. This matters. Searching the wrong portal for 20 minutes = bullshit load. Working through why two explanations conflict = maybe exactly the load I need. Formatting citations = bullshit. Building an internal causal model = probably not bullshit. So: LESS EFFORT ≠ AUTOMATICALLY BETTER.
The right question: did we remove friction, or did we remove practice? Jain et al., Cognitive Processing, 2026. 120 young adults after exclusions. Logical-fallacy task. AI group: more correct answers less reported mental effort Manual group: more independent reasoning more verification more metacognitive monitoring AI group itself varied: some collaborated some delegated. Again: the interaction pattern matters. Not “AI good/bad”. Not even “AI use”.
WHERE DID THE COGNITIVE WORK GO? BIG META-POINT — THERE MAY BE NO USEFUL “AVERAGE AI EFFECT” Boolzen et al., Artificial Intelligence Review, 2026. 85 eligible STEM studies. 49 studies / 59 effects in meta-analysis. At first: AI looks broadly positive. Then: wild variation between studies. Correct for likely publication bias → overall average benefit basically collapses. More useful signal: whether AI augmented the learner’s cognitive activity, substituted for it, or changed the task architecture. Good. That fits everything else. “Used ChatGPT” can mean: retrieve these papers explain this concept challenge my argument find the holes give me three counterfactuals invent the whole argument write the answer tell me if your answer is right I copied it because I’m tired Those are not one intervention. Calling all of that “AI use” is like asking whether “using machinery” makes you fit. Depends what the machine is doing.
ARE WE GETTING DUMBER?
Current answer: NOT PROVEN.
Could specific skills weaken through disuse? Yes. Very plausible. Some early AI evidence points that way. Automation + cognitive-offloading research gives us a long historical warning. Are basic human cognitive abilities broadly declining because of AI? No evidence strong enough to say that. Important distinction. Cash et al., Trends in Cognitive Sciences, 2026. Actual paper title: “Is AI making us stupid?” Their answer is basically: offloading can interfere with skill acquisition and contribute to skill decay BUT effects depend on use and core cognitive abilities may be more resilient. That seems right.
Need to separate: SKILL DECAY from GENERAL COGNITIVE DECLINE. Not the same thing.
I don’t care that I can't remember 40 phone numbers anymore. I care if I lose the ability to: spot bullshit/ build a causal model/ notice the missing variable/ know when I don’t know/ challenge a fluent answer/ act when the automation breaks. Those are different assets.
OLD AUTOMATION WARNING — STILL VERY RELEVANT
Bainbridge, “Ironies of Automation”, 1983. Seminal for a reason. Automation does the routine stuff. Human gets left with the weird 2%. Problem: the routine stuff was also the practice. So we automate the normal decisions, then expect a human who no longer makes normal decisions every day to suddenly be brilliant when the machine hits an edge case. That is structurally stupid. This is exactly the kind of setup that looks efficient until the day it matters. AI makes this problem more serious because it can automate cognitive work, not just physical or procedural work.
LONG TERM
Need to stay brutally honest here. What does 10–20 years of normal AI use do to cognition? NO FUCKING IDEA YET. Anyone speaking with certainty is ahead of the data. We have: short experiments weeks-long interventions survey waves early longitudinal signals decades of adjacent automation / offloading evidence We do NOT have: a generation of adults who have used generative AI for 20 years. Obvious because the technology has not existed that long. So: don’t pretend mechanism = outcome.
Possible long-run futures:
A) we offload rubbish → humans have more capacity for judgment / creativity / relationships / hard problems.
B) we offload the exact practice that builds judgment → output rises, capability hollows out.
C) both happen at once in different skills.
Probably C. Also probably different by: person role existing expertise task pressure interface incentives how AI is introduced what the system rewards after the time saving appears. Need research on that.
SYSTEMS LENS — WHERE THIS MEETS MY EXISTING WORK
This bit fits almost too neatly, so watch confirmation bias. My existing tool test: good tool = pays rent. Takes cognitive work off the bench. Returns clarity + leverage. Keeps reality visible. Makes the right action cheaper. Bad tool: toll booth. Looks efficient. Quietly burns human attention. Makes people do translation / checking / glue / recovery forever. AI adds a second ledger.
OLD RENT TEST: How much bandwidth does this tool give back?
NEW RENT TEST: What capability does it stop the human practising?
A tool can pass test #1 and fail test #2. That is new / important. Example: AI drafts the minutes. Great. No important skill lost. AI identifies every strategic issue, frames all options, recommends the call, explains why, and writes the decision paper. Also very efficient. But what exactly is the human now practising? Reading? Approving? Prompting? Maybe fine. Maybe not. Need to price that honestly.
“BRAIN-FRIENDLY AI” — WORKING DEFINITION
Not a scientific term. My synthesis. Brain-friendly AI = move cognitive work into the machine WITHOUT quietly removing the human capabilities needed to steer, verify, update, or recover the whole system. Not maximum automation. Best allocation of cognition.
Potential rules:
OFFLOAD HARD: - retrieval - admin - reformatting - repetitive checking - context reconstruction - low-value switching - tedious comparison where the comparison itself is not the learning goal
BE CAREFUL OFFLOADING: - problem framing - first-principles model building - causal reasoning - deciding which evidence matters - error detection - uncertainty calibration - final high-consequence judgment - the practice required to stay competent when the machine fails Human does NOT need to do everything. That would be idiotic.
Goal is not “protect effort”. Goal is: protect useful cognition.
[
IMAGE 3 — BRAIN-FRIENDLY AI: MOVE FRICTION, KEEP STEERING]
POSSIBLE DESIGN PRINCIPLES
1. THINK FIRST WHEN LEARNING MATTERS
Not every time. But if the task is meant to build capability: make the human generate / predict / frame before AI answers. Receipt: Wong + Qiu.
2. MAKE VERIFICATION CHEAP Sources. Provenance. Easy path back to primary evidence. Show what changed. Show uncertainty properly. But warning: confidence widgets are not automatically helpful. Ojewale et al. suggests uncertainty cues can still create overreliance.
3. TEST THE HUMAN WITHOUT THE TOOL If a system claims it is making people better: occasionally remove the tool. Can they: explain? transfer? spot the error? make the call? recover? If not, maybe the SYSTEM got better while the HUMAN didn’t. Could still be a valid trade. Just price it honestly.
4. KEEP THE HUMAN IN THE STEERING LAYER Not because “humans are special”. Because somebody has to own: goals values exceptions trade-offs reality when the model is wrong.
5. FIT, NOT ONE-SIZE-FITS-ALL Same AI can be extension for one brain / tax for another. No Normal Brain. No Normal AI User. Design needs: different levels of scaffolding different modes different defaults different points where automation hands control back.
6. SAVED BANDWIDTH IS NOT AUTOMATICALLY REINVESTED Important. If AI saves me 30% of my cognitive effort, the system might not give me 30% back for deeper thinking. It might simply give me 30% more work. This is the organization-level risk. Efficiency saving becomes throughput demand. Then AI did not buy cognition back. It just raised the quota. That fits the existing “under pressure we do what the system makes easiest” problem. Very important thread to develop.
CURRENT POSITION
Not: AI makes us stupid.
Not: AI makes us smarter.
Not: AI reduces brain activity.
Not: AI frees us for higher-order thinking.
Those are all too clean.
Current defensible version:
AI changes the distribution of cognitive work.
That can increase output/ lower effort/ improve some kinds of learning/ reduce others/ improve reasoning/ damage calibration/ remove useless friction/ remove useful practice/ change which neural routes get recruited/ sometimes all in the same person.
The important variable is not “AI use”. It is:
WHAT THINKING MOVED?
WHERE DID IT MOVE?
WHAT DID THE HUMAN KEEP PRACTISING?
WHAT DID THE SYSTEM DO WITH THE CAPACITY THAT CAME BACK?
NOTE TO SELF — THE LINE
Maybe: AI is not making us dumber. It is making some thinking cheaper. That means we will stop doing some of it ourselves. The question is whether we stop doing the thinking we no longer need, or the thinking we still need to stay in control.
Or: AI is changing where thinking happens. The opportunity is to move the cognitive tax. The risk is that we accidentally move the judgment with it.
Or: The goal is not to keep thinking inside the skull. The goal is to keep the right thinking there.
QUESTIONS I STILL WANT ANSWERED
- Which skills decay fastest once LLMs take over the operation?
- Does existing expertise protect people from bad offloading?
- Do novices + experts need completely different AI?
- Does “think first” still work outside education / creativity?
- What happens to causal reasoning over 1 / 5 / 10 years?
- What happens to error detection?
- Does AI improve judgment if it frees bandwidth, OR does saved bandwidth just get filled with more volume?
- What tasks genuinely show neural efficiency versus disengagement?
- Does AI widen or narrow cognitive differences between people?
- Can interfaces actively preserve calibration?
- What should always have a “manual mode” because the human remains the fallback?
- At organization level: who gets the saved attention?
- Is AI compounding capability or just compounding output?
That last one feels like the real test.
WHAT WOULD CHANGE MY MIND?
Evidence that long-term heavy AI users:
- retain / improve independent transfer,
- retain calibration,
- become better at framing + judgment,
- and use saved capacity on higher-level cognition, would push me toward “AI is mostly cognitive leverage.”
Evidence that:
- independent performance declines,
- error detection weakens,
- foundational skills stop developing,
- calibration drops,
- and organizations simply turn saved effort into higher workload, would push toward “we are building dependency and calling it productivity.”
Need both. Don’t go hunting get receipts.
REFERENCES / RECEIPTS RECENT CORE Cash, T.N., Kelly, M.O., Macnamara, B.N. & Risko, E.F. (2026). “Is AI making us stupid?” Trends in Cognitive Sciences, 30(8), 673–675. https://doi.org/10.1016/j.tics.2026.06.004 https://pubmed.ncbi.nlm.nih.gov/42425850/ Boolzen, C., Kuhn, J., Flegr, S. et al. (2026). “Evidence of impact and interpretational limits of generative AI in STEM education: a systematic review and meta-analysis on cognitive learning outcomes.” Artificial Intelligence Review, 59, 223. https://doi.org/10.1007/s10462-026-11665-9 Qian, W., Yang, F., Cao, Y. et al. (2026). “Generative AI and cognitive load in education: a systematic review of WoS/SSCI-indexed studies through the lens of cognitive load theory.” Frontiers in Psychology, 17. https://doi.org/10.3389/fpsyg.2026.1921504 Wong, S.S.H. & Qiu, S.X. (2026). “Think First, ChatGPT Later: Guiding Human–AI Collaboration for Learning Gains in Independent Human Creativity.” Educational Psychology Review, 38, 45. https://doi.org/10.1007/s10648-026-10118-7 Fernandes, D., Villa, S., Nicholls, S. et al. (2026). “AI makes you smarter but none the wiser: The disconnect between performance and metacognition.” Computers in Human Behavior, 175, 108779. https://doi.org/10.1016/j.chb.2025.108779 Jain, P., Shajit, D., Maini, A., Shahin, Y. & Panda, A. (2026). “Cognitive offloading, critical thinking and attitudes towards artificial intelligence in the era of ChatGPT.” Cognitive Processing. https://doi.org/10.1007/s10339-026-01375-z https://pubmed.ncbi.nlm.nih.gov/42377671/ Dell’Acqua, F., McFowland, E. III, Mollick, E. et al. (2026). “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality.” Organization Science, 37(2), 403–423. https://doi.org/10.1287/orsc.2025.21838 Melumad, S. & Yun, J.H. (2025). “Experimental evidence of the effects of large language models versus web search on depth of learning.” PNAS Nexus, 4(10), pgaf316. https://doi.org/10.1093/pnasnexus/pgaf316 https://pubmed.ncbi.nlm.nih.gov/41163786/ Zhang, J., Xu, B., Liu, Y. et al. (2026). “Exploring ChatGPT’s potential in dialogic teaching: A comparative neurocognitive study of AI and human instruction.” Acta Psychologica, 268, 107268. https://doi.org/10.1016/j.actpsy.2026.107268 Jin, Z., Yin, J., Tian, Q. et al. (2026). “Can large language models teach creativity? Cognitive gains, affective gaps, and distinct neural patterns.” Computers in Human Behavior: Artificial Humans, 8, 100288. https://doi.org/10.1016/j.chbah.2026.100288 Ojewale, V., Ryan, J., Venkatasubramanian, S. & Boykin, C.M. (2026). “More is not better: Visual uncertainty cues and the fragility of trust calibration in LLM-assisted decision making.” Computers in Human Behavior: Artificial Humans, 8, 100307. https://doi.org/10.1016/j.chbah.2026.100307 PUBLIC CONTEXT / NOT PEER-REVIEWED CORE Kosmyna, N. et al. (2025). “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.” arXiv preprint 2506.08872. https://arxiv.org/abs/2506.08872 https://www.media.mit.edu/projects/your-brain-on-chatgpt/overview/ SEMINAL / OLDER BUT ACTUALLY USEFUL Risko, E.F. & Gilbert, S.J. (2016). “Cognitive Offloading.” Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002 Clark, A. & Chalmers, D. (1998). “The Extended Mind.” Analysis, 58(1), 7–19. https://doi.org/10.1093/analys/58.1.7 Sparrow, B., Liu, J. & Wegner, D.M. (2011). “Google effects on memory: cognitive consequences of having information at our fingertips.” Science, 333(6043), 776–778. https://doi.org/10.1126/science.1207745 Bainbridge, L. (1983). “Ironies of automation.” Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098(83)90046-8