The Krell Brain Booster: How AI Improves Thinking and Work

AI can act like the Krell brain booster from the 1956 film Forbidden Planet — a machine that amplifies a mind far beyond its natural limits, at work and in everyday thinking. But the film’s warning still applies: the machine amplifies whatever’s already there, blind spots included. 2026 research from Brookings, the OECD, and university labs on three continents shows the difference between a boosted mind and a surrendered one comes down to one learnable skill: metacognition, the ability to think about your own thinking, before, during, and after every AI interaction.

Here’s a simpler way to picture the failure mode. Imagine buying a textbook for a subject you need to master, opening it once, skimming a page, and putting it back on the shelf — then telling people you know the subject because you own the book. Nobody would accept that as learning. Yet that’s exactly what it looks like when we ask AI a question, glance at the answer, and move on: the answer exists, on a screen instead of a shelf, but nothing has actually moved into our own head.

The Textbook Problem: Having an Answer Isn’t the Same as Knowing It

The textbook-on-the-shelf problem shows up in the data with uncomfortable precision. A 2026 research brief synthesising recent studies on AI in learning found that people using unrestricted ChatGPT scored 17% lower on independent exams than those working under AI guardrails — despite scoring higher on the AI-assisted practice questions themselves (Kharbach, “When AI Supports Learning—and When It Replaces Thinking”). Another study found that once AI help was removed, participants’ own accuracy on the same maths tasks dropped from 73% to 57%. A third found that people who learned a topic with ChatGPT’s help retained just 57.5% of it after 45 days, compared with 68.5% for people who learned it the traditional way.

In every case, the AI-assisted group looked more capable in the moment. The book was open, the answer was right there. But close the book — take the AI away — and the knowledge wasn’t actually in their heads. That’s the textbook problem, quantified.

The Krell Machine: A Brain Booster, or an Amplifier of Your Blind Spots

In Forbidden Planet, an ancient, vanished civilisation called the Krell built a machine that could amplify a mind’s intelligence to godlike levels — the Krell laboratory scene is worth the four minutes if you haven’t seen it. But the Krell never accounted for their own unconscious. The machine amplified everything, including the parts of the mind they hadn’t examined, and it destroyed them in a single night.

That’s a strikingly literal preview of what researchers Steven Shaw and Gideon Nave found when they studied how people actually reason with AI (Shaw & Nave, “Thinking—Fast, Slow, and Artificial”). They coined the term “cognitive surrender” for what happens when people adopt AI’s answers with minimal scrutiny — and in their experiments, when AI deliberately gave wrong answers to structured problems, participants went along with it roughly 80% of the time, and felt more confident afterwards regardless of whether the AI had actually been right. A separate 2025 study found the same pattern in the wild: of 23 people interviewed about their use of ChatGPT for factual questions, only one bothered to fact-check what they were told, even when the AI’s accuracy on those questions was as low as 12% (Jacob, Kerrigan & Bastos, “The chat-chamber effect: Trusting the AI hallucination”). The term has since travelled well beyond one lab: MIT’s own 2026 committee report on AI in education independently warns of student “cognitive surrender” — abandoning the intellectual struggle the moment AI offers a quick answer — and argues that preserving what it calls “productive struggle” is now central to genuine learning (MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, 2026).

This isn’t confined to the lab, either. A 2025 Microsoft Research and Carnegie Mellon study surveyed 319 knowledge workers about 936 real examples of using generative AI at work, and found the same confidence dynamic driving actual on-the-job behaviour: higher confidence in the AI predicted less critical thinking, while higher confidence in one’s own ability predicted more (Lee et al., “The Impact of Generative AI on Critical Thinking,” CHI 2025). Tellingly, the study found GenAI doesn’t switch critical thinking off so much as relocate it — away from generating ideas, and toward verifying, integrating, and overseeing whatever the AI produced.

None of this is stopping adoption. Pew Research found that the share of American adults using AI chatbots jumped from 33% to 49% between 2024 and 2026 — even as trust moved the opposite direction, with 63% now saying AI is advancing too fast and only 16% expecting it to be good for society over the next 20 years. People are reaching for the amplifier faster than they’re learning to watch it.

AI, in other words, is a Krell machine. It amplifies your thinking — but only what’s already there, blind spots included, unless you deliberately supply the self-awareness the Krell never did. Used reflexively, it’s a genuine brain booster. Used passively, it just makes your existing gaps louder and more confident-sounding.

The Metacognitive Cycle: Four Steps to Using AI Well

So what does “reflexively” actually mean in practice? It comes down to four steps — and, importantly, they don’t run once and stop. They form a cycle, where each pass leaves you better equipped for the next one.

1. Know your domain. You can’t spot a wrong answer in a field you don’t understand. A 2026 University of Technology Sydney report on AI and cognitive offloading found an equity split forming in classrooms: learners with strong existing domain knowledge use AI as a genuine accelerant, while learners without that foundation are far more likely to offload the very thinking they still need to build (UTS, “Artificial intelligence, cognitive offloading and implications for education”). Domain knowledge isn’t optional background — it’s what makes step 3 possible at all. A separate 2026 study of 468 university students found the same mechanism at the level of individual psychology: confidence in one’s own ability to think a problem through — not the tool itself — was what predicted stronger critical thinking, task persistence, and learning depth from using digital tools at all (Wang, Frontiers in Psychology, 2026).
2. Know your own intellectual weaknesses. Before you ask AI anything, know where your own understanding is thin. This is precisely the “metacognitive engagement” the OECD’s Digital Education Outlook 2026 found declines when people lean on AI too heavily — the quiet loss of noticing what you don’t understand. If you don’t know your own weak points, you won’t know which of AI’s answers deserve the most scrutiny.
3. Critically evaluate the AI’s response. Don’t accept the answer because it’s fluent and confident — that fluency is exactly the mechanism behind the chat-chamber effect described above. A controlled study of 116 middle-schoolers found that a two-hour training in how to interrogate AI output changed behaviour measurably: trained students asked follow-up questions after a weak AI answer 59.2% of the time, versus 27.9% for untrained peers (Teaching Students to Question the Machine). Tellingly, how confident people felt about their own AI literacy had no relationship to how well they actually evaluated AI output (r = 0.01) — this is a trained behaviour, not a feeling.
4. Lean in and learn. Use whatever gap the AI just exposed as the reason to actually build your domain knowledge, rather than letting the AI quietly paper over it. Do this well, and you don’t end the cycle where you started — you return to step 1 with more domain knowledge than you had before, ready to evaluate AI more sharply next time. That’s what makes this a cycle rather than a checklist: an investigative mindset in step 3 generates the domain-knowledge growth in step 4, which then upgrades step 1 for good.

The cycle above assumes the AI is already reasonably set up for the job — that setup is worth a moment of its own. Give it a persona or role suited to the task, load it with your own reference material or a trained skill if it hasn’t seen your domain before, and keep refining that setup as you learn what it gets wrong. Skip this and you’re running the cycle against a generic assistant instead of one primed for your specific problem. It’s a big enough topic to deserve its own article, but worth flagging here: preparing the machine is part of the process, not a separate step from it.

The Cost of Skipping the Cycle: Bad Work and Missing Understanding

Skip the cycle — accept AI’s first answer without running it through your own domain knowledge — and the failure shows up in two distinct, measurable ways: in the work itself, and in you.

The first is bad work product hiding behind a well-formatted answer. The chat-chamber effect study found that a chatbot’s confident phrasing led people to skip verification even when the underlying information was wrong — on the questions researchers tested, ChatGPT’s accuracy ran as low as 0–22% against comparison sources, yet only 1 of 23 users checked it before relying on the answer. Handed to a real audience — a client, an employer, a marker — the polish doesn’t hold up; the errors do. That’s step 3, critical evaluation, skipped.

A 2026 comparative study split young adults into an AI-assisted group and a manual group on a logical-fallacy task, and found the outcome hinged on approach rather than the tool itself: AI users scored higher on accuracy and reported less mental effort, but only those who engaged with it as a strategic partner kept their own independent reasoning and verification intact — the rest showed what the researchers called “complete delegation” (Jain et al., Cognitive Processing, 2026). Accuracy went up either way; critical thinking only did for the group still running the cycle.

The second failure only surfaces once the AI isn’t there to lean on. Exams, interviews, and follow-up questions all do what the AI conversation didn’t: force you to answer from your own head. This is exactly what the exam-score research showed — people who used AI without guardrails scored 17% lower on independent assessments than those who’d been made to verify and reconstruct their work, and a separate study found retention dropping to 57.5% versus 68.5% for people who’d actually learned the material rather than borrowed the answer. That’s step 4, lean in and learn, skipped.

Skip both steps at once and you get the worst version of AI use: confident, unverified output that also failed to teach you anything — the exact opposite of a brain booster. It’s the Krell machine with nobody watching the amplifier.

Climbing Bloom’s Taxonomy: Why the Investigative Mindset Matters

That last link in the cycle — investigative questioning turning into real domain knowledge — has a name in the education research: Bloom’s Taxonomy. It’s a well-established hierarchy of cognitive skills, running from lower-order thinking (Remember, Understand, Apply) up to higher-order thinking (Analyze, Evaluate, Create). Most single-shot AI use never leaves the bottom rung: ask a question, receive a fact, done — pure “Remember,” with none of the higher-order skills engaged.

A study of 665 pharmacy students measured exactly where AI-supported learning tips over into genuine critical thinking, using this taxonomy as the yardstick (Does ChatGPT-enhanced collaborative learning foster critical thinking in education? A Bloom’s Taxonomy perspective). The model explained 78.3% of the variation in students’ critical-thinking outcomes — and the strongest predictor wasn’t simply using the tool, it was Understanding (β = 0.812), the level where students actively interpret and re-explain material rather than just receive it. Remembering barely mattered at all (β = 0.098). The researchers’ conclusion, in one line: AI “does not inherently enhance critical thinking but must be integrated into structured [active] learning.”

That’s the investigative mindset in mechanical terms. Curiosity and follow-up questions are what push an AI interaction up off the bottom rung of Bloom’s Taxonomy — toward Understand, Analyze, and Evaluate — and it’s only at those higher rungs that anything durable gets built. Without that mindset, you can use AI for hours and never leave “Remember.” With it, every AI interaction becomes a small rep at climbing the taxonomy, and the domain knowledge you build along the way makes the next climb faster.

The same research base backs this up at scale. The OECD found that AI tutoring built around Socratic questioning — prompting reasoning, not supplying answers — produced measurable critical-thinking gains, and that AI used to support peer discussion, rather than replace it, produced “small-to-medium gains in subject learning and more substantial improvements in critical thinking.” The Brookings Global Task Force on AI in Education — a year-long study spanning 500+ stakeholders in 50 countries — reached a similar dividing line, distinguishing tools and habits “that teach, not tell” from ones that simply hand over finished answers. The pattern is consistent: the benefit isn’t in the tool, it’s in whether a metacognitive cycle is running around it.

Three Fields, One Pattern: What This Looks Like in Practice

Theory aside, this is what the cycle actually looks like in three very different fields.

Building trading tools. When I set out to build a set of trading tools with AI’s help, the value AI added was real — but it was never a shortcut around the four steps above, it was entirely dependent on running them. My own expertise is in IT and hosting, and that domain knowledge was what let me evaluate whether the technical logic AI proposed actually made sense, and where it quietly didn’t (steps 1 and 3). But the project still pushed me further than that existing expertise reached — I was pushed into learning virtual machines specifically, in more depth than my general IT and hosting background had covered, because the tools needed VMs to run on. AI could generate configuration and suggest fixes, but it made mistakes there too, and following its suggestions blindly would have meant shipping something broken. The only way through was to actually build that VM knowledge up well enough to judge whether AI’s proposed fix was right, wrong, or half-right (step 4) — not optional once real infrastructure is on the line.

What stood out most was how constant the decision-making was. AI didn’t get things right and then step back — it needed a person weighing in on trade-offs, catching mistakes, and solving problems it couldn’t reliably solve itself, over and over, for the length of the build. That’s Bloom’s Taxonomy in practice: AI handled a lot of the Remember and Apply work, but the Analyze, Evaluate, and Create work — the higher-order judgment that actually shipped a working product — stayed squarely with me.

Conducting an orchestra. A friend of mine is a composer, and his experience shows the same pattern in a completely different domain. Using AI, he’s gone from writing compositions to effectively conducting a whole virtual orchestra of skilled musicians performing his work — often better than he could play any single instrument in it himself. What makes that possible isn’t the AI alone; it’s his own domain knowledge of composition and arrangement, which lets him direct the result rather than just accept it — hearing when a phrase is wrong, when an arrangement doesn’t serve the piece, when to reject a take and ask for another. His role has effectively moved up a level: from composer to producer, directing a higher-order creative process that AI made possible but that his musical judgment is what makes actually good.

Writing a book. The same pattern holds when using AI to research or draft a book. AI can surface facts, sources, and leads at a scale no one person could match alone — but it has no domain judgment of its own, so it can’t reliably tell you which of those leads actually matters to your subject. That recognition is entirely yours, and it depends on how much domain knowledge and inquisitiveness you bring to the exercise: an inquisitive mind chases the throwaway detail AI mentioned in passing and turns it into the strongest part of the book; a passive one takes AI’s summary at face value and never notices what it missed. AI did the finding; the writer still has to do the recognising — and often has to go and build the specific knowledge needed to know the find was worth chasing in the first place.

Same amplifier, same Krell machine, three entirely different fields — a trader, a composer, and a writer, each with an improved, more capable domain expertise, not a technician, a musician, or an author replaced by one.

None of this is effortless, and it shouldn’t be mistaken for the easy path. Working well with AI is cognitively hard — arguably harder than doing the task without it, because you’re running two demanding processes at once: the task itself, and a constant, active supervision of whether the AI is helping you do it or quietly doing it badly on your behalf. The convenience is real. The cognitive load doesn’t disappear; it just relocates, exactly as the Microsoft/Carnegie Mellon knowledge-worker study found — from generating the work to verifying, integrating, and overseeing it, which is a different and, for most people, less familiar kind of effort.

AI as a Workplace Improver: Why It Raises the Ceiling, Not Just the Floor

The Krell in Forbidden Planet built exactly one amplifier and buried it on one planet. AI is the opposite: the same amplifier is now open in a browser tab for anyone with an internet connection, free or near enough to it. That difference is exactly why the metacognitive cycle matters at the scale of a whole labour market, not just for one person’s learning — when the amplifier is universally available, the thing that stops being universal is who has the approach to use it well.

This is also why AI’s effect on work isn’t a flat, across-the-board productivity boost. Mapped onto Bloom’s Taxonomy, low-order AI use — a fact, a first-draft paragraph, a boilerplate answer — is trivially easy to automate and easy for anyone to replicate, so it commands very little premium. Higher-order AI use — interrogating an answer, applying it to a judgment call only you have the context for, synthesising it into something new — is exactly the work that still needs a person with domain knowledge and an investigative mindset running the loop. The likely outcome isn’t “AI replaces jobs” so much as “AI raises the floor on low-order tasks and raises the ceiling on high-order ones” — a genuine human improver for the people supplying the judgment, and a genuine threat to anyone hoping to skip straight to the finished answer.

The labour-market data is already showing this split. PwC’s 2026 Global AI Jobs Barometer found firms in the most AI-exposed sectors achieved 34% productivity growth from 2018–2025, against 24% elsewhere — and AI-skilled workers now command a 62% wage premium. The more telling number is the split PwC calls the “two-track” labour market: “professionalised” roles, where AI handles routine tasks while elevating human expertise (their examples include radiologists and recruiters), are growing jobs twice as fast and delivering 42% higher salary increases than “democratised” roles, where AI just makes the existing task easier for anyone to do. It shows up at entry level too — AI-exposed junior roles now demand traditionally senior “human-intensive” skills like leadership, creativity, and judgement seven times more often than before, and those roles grew 35% since 2019 while traditional entry-level postings fell 10%. That’s the metacognitive cycle showing up as a hiring statistic: the roles rewarding domain knowledge and judgment are pulling away from the roles that just reward AI access.

Independent research points the same way. An analysis of over 10 million UK job postings found candidates with AI-related skills command an advertised salary 23% higher than otherwise-comparable candidates without them — nearly double the premium of a master’s degree — and in a hiring experiment with 1,700 recruiting professionals, identical résumés listing AI skills were 8–15% more likely to get an interview, an advantage large enough to offset age and education disadvantages entirely (World Economic Forum, 2026).

For Skillset Centre’s audience specifically, that shows up in very concrete places:

• Writing a resume or cover letter with AI. Accept the first generic draft — “Remember”-level use — and you get a resume that reads like everyone else’s, because the AI’s default output is everyone else’s. Challenge it to surface achievements you’d forgotten, push back on generic phrasing, cross-check every claim against what you actually did — “Analyze/Evaluate”-level use — and you get a stronger, truer document. Our 10 Things Not to Do When Using AI to Write Your Resume article covers the specific traps that come from stopping at the first draft.
• Reading AI-heavy job ads. You can’t critically evaluate a role you don’t understand — step 1 of the cycle, applied to your job search. The AI Jargon Decoder gives you the domain knowledge to read past the buzzwords and judge whether a role is actually right for you.
• Interview preparation. Used well, AI is a sparring partner that stress-tests your answers and asks the follow-up questions a real interviewer would — the same “critically evaluate, then lean in and learn” behaviour the AI-literacy intervention study measured. Used passively, it just hands you a script that falls apart the moment a real interviewer deviates from it.
• Assessing your own employability. The skills employers increasingly screen for — judgment, verification, the ability to work with AI rather than defer to it — are exactly the higher-order skills this cycle builds. The Employability Quiz is a useful gut-check on where you stand.

The amplifier is available to everyone now. The approach isn’t — yet. That gap is where the real advantage sits, and where the actual improvement — to your work, and to you — gets made.

Key Takeaways: How to Actually Do This

Reading about the cycle isn’t the same as running it — here’s the checklist version to act on the next time you open an AI tool.

1. Name your domain knowledge before you ask. Spend ten seconds stating what you already know about the question. If the honest answer is “not much,” treat every part of the AI’s response as unverified until you check it elsewhere.
2. Name your weak spot, specifically. Don’t just ask AI a question — identify which part of your own understanding is thinnest first. That’s the part of the answer to scrutinise hardest.
3. Ask at least one follow-up question, every time. Don’t accept the first answer. Push back, ask “why,” or ask for the counterargument — this single habit is what separated trained AI users from untrained ones in the research above.
4. Run the “book test” before calling anything done. Ask yourself: if the AI disappeared right now, could I still explain or reproduce this? If not, you have an answer, not knowledge — go back and close the gap yourself.
5. Chase the throwaway detail. When AI mentions something in passing that seems minor, don’t skip it — that’s often exactly where an inquisitive domain expert finds something the AI itself didn’t recognise as important.
6. Never submit AI output you haven’t personally verified. Resumes, reports, code, client work — if you can’t stand behind every claim in it yourself, it’s not ready, no matter how polished it reads.
7. Expect the effort to move, not disappear. If using AI feels completely effortless, that’s a signal you may have stopped at “Remember.” The real work is in judging and directing the output, not just receiving it.
8. Treat every AI interaction as a rep, not just a task. Each time you push an answer up to Analyze or Evaluate instead of stopping at Remember, you leave the exchange with more domain knowledge than you started with — bank it, so the next question you ask is sharper.

FAQ

The Bottom Line

None of this research says “stop using AI.” It says the amplifier is real, and it’s now sitting in everyone’s browser — which means the old advantage of simply having access to information or computing power is gone. What’s left as the actual differentiator is a process and a mindset, not a tool: the four-step cycle above is the process; approaching every AI answer with domain knowledge, self-awareness, and genuine curiosity is the mindset. Underneath both is the single idea this whole article has been circling: recognising the importance of metacognition itself — supervising your own thinking rather than quietly handing it over — as the master skill AI has made more valuable, not less relevant. Neither the process nor the mindset works without the other — a good process run with a passive mindset just produces faster, more confident-sounding mistakes, and an inquisitive mindset with no domain knowledge to apply it to has nothing to bite into.

Run both together, and AI is the Krell machine working as intended — a genuine boost to a mind that’s already supervising itself, and a genuine improver of the work that mind produces. Skip them, and you get a confident-sounding version of the textbook nobody opened: the answer was right there, and somehow nothing was learned. The skill worth building in 2026 isn’t avoiding AI. It’s the process and the mindset that keep you the one doing the thinking.

References

• Jacob, S., Kerrigan, P., & Bastos, M. (2025). The chat-chamber effect: Trusting the AI hallucination. New Media & Society (SAGE).
• Kharbach, M. (2026). When AI Supports Learning—and When It Replaces Thinking. Research brief.
• Brookings Global Task Force on AI in Education (2026). A New Direction for Students in an AI World.
• OECD (2026). Digital Education Outlook 2026.
• University of Technology Sydney (2026). Artificial intelligence, cognitive offloading and implications for education.
• Teaching Students to Question the Machine (2026). arXiv:2604.01955.
• Shaw, S., & Nave, G. (2026). Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender. SSRN.
• Does ChatGPT-enhanced collaborative learning foster critical thinking in education? A Bloom’s Taxonomy perspective (2025). ScienceDirect.
• Lee et al. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. Microsoft Research & Carnegie Mellon, CHI 2025.
• PwC (2026). 2026 Global AI Jobs Barometer.
• Forbidden Planet (1956). The Intelligence Indicator (Krell Lab Scene), 4K.
• Pew Research Center (2026). Americans and AI 2026: Chatbots, Smart Devices and Views on Impact.
• MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training (2026). Report on AI and Education.
• 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: a comparative study of AI-assisted and manual task performance in young adults. Cognitive Processing (Springer).
• Wang, J. (2026). Cognitive offloading through digital tools and its relationship with critical thinking, task persistence, and learning depth. Frontiers in Psychology, 17.
• World Economic Forum (2026). How AI skills and experience are transforming the workplace.