नेति नेति (Neti Neti)

Tag: CriticalThinking

  • In the Era of Cognitive Crutch: Would We Have the Next Aryabhatta?

    *Image Reference: Generated using ChatGPT

    Disclaimer:
    The views expressed in this blog are personal. Research papers, institutional studies, and reference links have been included wherever applicable to support technical concepts, statistics, and scientific discussions presented in the blog.

    Somewhere in fifth-century India, a mathematician was wrestling with an idea civilization had never fully formalized before- representing nothingness itself! He wasn’t prompting a tool. He was doing something rarer and harder- holding an unresolved question in his mind long enough, uncomfortably enough, that an entirely new idea had room to emerge. We call it zero. Zero was not discovered by Aryabhatta, it was endured. We built the modern world on it. The question worth asking now is whether that kind of thinking- born entirely from struggle, silence, and sustained discomfort is still possible. Or whether we are quietly engineering the conditions that make it less and less likely.

    Human breakthroughs were never purely about intelligence. Humans learned to control fire, invented the wheel, developed astronomy, created music systems, Sushruta performed remarkably advanced surgical procedures centuries before modern medicine. Visionaries like Van Gogh, Monet and many others in the era shattered centuries-old artistic doctrines in realism and brought revolution in art. Socrates transformed philosophical inquiry by questioning accepted truth itself. What they all shared was not access to tools, but the willingness to sit with difficulty long enough to see what others had missed.

    For the first time in human history, we have created a technology that does not merely store information or automate physical labour. Large Language Models (LLMs) participate directly in reasoning, articulation, synthesis, brainstorming, communication, and ideation itself. The printing press changed how humans distributed knowledge. Search engines changed how humans retrieved knowledge. Now, Generative AI may fundamentally change how humans produce thought.

    That distinction matters. Because The central question is no longer, “Will AI improve productivity?”, the deeper question is: What happens to human cognition when thinking itself becomes partially outsourced?, What happens when impatience for quick answers replaces the habit of sitting with struggle, uncertainty, and deep thought? When speed begins to earn more recognition than depth, and getting work done becomes more important than truly understanding or creating it ourselves?

    An Aryabhatta-level breakthrough requires someone to reject the prevailing consensus entirely. Will offloading thinking restrict human innovation or might AI actually accelerate the next “Aryabhatta”?

    How LLMs Work?
    At its core, a Large Language Model (LLM) is a highly sophisticated statistical pattern recognition machine. It possesses no consciousness, biological understanding, or internal grasp of factual reality. The fluency of the outputs produced by generative AI/LLMs is so convincing that humans naturally anthropomorphize the system and assume understanding exists underneath the surface. It does not.

    Imagine someone says to a child, “Once upon a…”. The word follows in the child’s mind is “time!”. When someone asks AI to complete the same sentence, AI would also most likely say “time”. But what’s happening inside is fundamentally different.
    The AI is a gigantic math machine. It has read billions of sentences before. So, when it hears “Once upon a…”, it calculates the next word based on probability: “time” at 80%, “road” at 30%, “mango” at 10% (numbers illustrative). Then it picks one. That’s mostly what an LLM does- pattern prediction using artificial neural networks trained on mammoth data.

    The model continuously predicts the most probable next word based on the sequence of words that came before it. Those probabilities are not fixed — they shift dynamically based on context, known as conditional probability. “The astronaut stepped onto the…” makes “Moon” highly probable. Change it to “The chef carefully placed the…” and now “cake” or “plate” dominate. This constant reshaping of probability based on context is why LLMs feel remarkably conversational.

    One more layer matters, that is, the temperature. Temperature acts like a creativity dial for LLMs. At low settings, the LLM model chooses the safest, most probable words — stable and predictable. At higher settings, it explores lower-probability choices. Ask it to complete “The pirate opened the treasure chest and found…” at low temperature: “gold coins” At high temperature: “a dancing chicken wearing sunglasses”. The knowledge is identical. What changes is the model’s willingness to deviate from the statistical average. This creates the illusion of a creative spark but it is actually just manufactured randomness. The machine is never truly inventing. It is rolling a “weighted die” based on the mathematical average of human data it has already seen.

    This is worth holding onto as we go further. Because the question of what AI can and cannot do for us begins here- not with its outputs, but with its fundamental nature.

    How Human Brains Work?
    When a child hears “Once upon a…”, she too is recognizing a pattern from stories she has heard before. But along with that, something entirely different happens. Her brain may remember grandma telling stories. She may already imagine what this story will be- the one with fairies, forests, or dragons. She feels curiosity, connects emotions and memories, asks questions, experiences fear, invents something completely new.

    Both humans and LLMs recognize patterns, predict what comes next, learn from examples, and improve with more data. But the human experience of pattern recognition creates meaning. It associates data with experience, goals, fear, curiosity, survival instincts, emotions, physical sensations, consciousness, and self-awareness. Your brain also uses probability, but it assigns meaning to that probability, and then it cares about the outcome.

    Take an example of Van Gogh’s impasto technique — paint laid so thick it became almost sculptural. It did not come from studying what was already accepted or data already available. It came from something felt. Whether rebellion, exclusion, or simply a compulsion to express the world exactly as he experienced it, the origin was very human and interior. I won’t be precise about his history- that’s not the point. The point is that something original emerged from something felt. An LLM generates by pattern. It does not feel. And so far, nothing it produces has needed it.

    Humans bring grounded understanding we learn from physical reality, consequences, pain, risk, and reward. We bring intent and judgment; we decide what matters and take responsibility for outcomes. We reason causally in messy, incomplete environments where the rules keep changing.
    No LLM does any of that. It responds and we decide. AI is probabilistic, not deterministic.

    What Is Needed for Human Brains to Be Creative?
    What is fascinating is that thoughts and breakthroughs rarely emerged from comfort or efficiency. They emerged from unresolved problems, intellectual obsession, necessity, prolonged uncertainty, deep observation, and intense conceptual struggle.
    But here’s the catch. The human brain is an energy-conserving system that runs on just 20 watts of power. It naturally minimizes heavy cognitive processing and constantly seeks the path of least resistance, a phenomenon called cognitive offloading. The brain operates on the “law of least effort”. When faced with a complex task, the neural network adapts by instinctively defaulting to easiest option. Do you experience your brain taking an “I’ll just google it!” or “Let me ask AI!” pathway more and more frequently? Over time, this makes independent problem-solving feel more difficult and mentally draining, creating a dependency loop where the AI-consulting reflex becomes deeply ingrained. However, if used and challenged continuously, the brain thrives.

    While AI tools save time, if not used for right reasons, the risk is that we outsource our deep thinking, planning, and memory to them. Studies from MIT and Harvard suggest that overusing generative AI for critical thinking or writing can reduce brain activity, erode long-term learning, and make it harder to retain information.
    The concern is not hypothetical. Researchers are increasingly investigating whether excessive cognitive offloading to AI may contribute to forms of cognitive atrophy over time.

    Breakthrough ideas require intense conceptual friction, synthesis, and deep incubation. Current generative systems risk homogenizing human thought and fundamentally alter how we construct an entirely new idea.
    Cognitive friction is the mental resistance or struggle you experience when working through a difficult problem — the discomfort of not knowing, the effort of figuring something out, the slow grind before clarity arrives. Cognitive friction matters. Modern society tends to treat friction as inefficiency. We celebrate speed, optimization, automation, and convenience. But not all friction is waste. Some friction is psychologically necessary. A musician struggling unsuccessfully with melodies for months may eventually discover an entirely new sound. A mathematician wrestling with incomplete structures may suddenly see a hidden pattern. A writer spending weeks unable to articulate an idea may slowly arrive at genuine clarity.

    Historically, deep insight often emerged from prolonged unresolved tension. AI removes much of that friction. That is simultaneously its greatest strength in some professions and potentially its greatest cognitive risk in others. Because while AI may optimize the production of answers, it may unintentionally weaken the mental conditions under which originality historically emerged. The danger may not be that humans become unintelligent. The danger may be that humans stop practicing difficult thought long enough to develop deep insight.

    To be fair, there is a counter-argument worth taking seriously. Some researchers argue that AI actually helps expands creative range- that having a tireless brainstorming partner lowers the activation energy for ideas and exposes thinkers to combinations they would never have encountered alone.

    The honest answer is that it depends entirely on how AI is used. AI when used as a sparring partner builds cognition. But AI used as a ghostwriter quietly erodes it. The tool is the same. The posture is everything. While this blog acknowledges optimistic possibilities, it focuses on the growing trends and challenges within the broader AI consumer base.

    How AI-Led Content Creation Could Limit Us?
    Fresh AI trains on new human-made content on the internet. Generates content based on it. Eventually, content created by AI increases. AI now generates content based on content created by AI. Creativity limits. Experts have named this phenomenon as “model collapse”.

    While AI tools expand the overall volume of generated ideas, the structural novelty and variety of that content decreases over time. Model collapse is what happens when AI begins feeding on its own output each generation of training data slightly less human, slightly less varied, slightly less real than the last. The loop tightens. The range of ideas narrows. The machine begins to dream only of itself because they are trained recursively on data generated by preceding AI models rather than on organic, human-created data. The idea is, the variety of ideas and originality will shrink by addition of less and less human produced data in the training data set.
    As the internet becomes flooded with LLM-generated text, future AI models will inevitably scrape this synthetic content. This forms an intellectual closed-loop system. If human innovators outsource their brainstorming to these models, they subject their brains to a narrowing intellectual field. This drastically stunts the cognitive deviation required to invent a truly new idea.

    The irony is sharp: the more we rely on AI to think for us, the less original the thinking AI has left to learn from.

    What Areas Is AI Really Good For?
    Let’s be honest about where AI genuinely earns its place — because it does earn it, in ways that matter enormously.

    Think about what it means for a doctor to detect cancer from a biopsy image in minutes- a process that would take a human pathologist several hours. Or for a neurologist studying Alzheimer’s to get a risk prediction from something as simple and non-invasive as a recorded conversation- no expensive scan, no months of waiting. Researchers at the University of Sheffield built exactly that: a tool called CognoSpeak that identified dementia with 90% accuracy just by analyzing how a person spoke. A meta-analysis of 83 studies published in Nature Digital Medicine found AI diagnostic accuracy already on par with non-expert physicians not yet at the level of specialists, but closing the gap in ways that would have seemed implausible a decade ago. These are not incremental improvements. They are compressions of time and scale that free human experts to do what only humans can make the judgment call, sit with the patient, carry the responsibility.
    Now zoom out further. Astronomers at Oxford built an AI tool that filtered over 30,000 alerts from space telescopes, missing fewer than 0.08% of real supernova signals reducing the load on human researchers by 85% while losing almost nothing. As Stanford’s KIPAC puts it, upcoming surveys like the Rubin Observatory will generate datasets so vast that traditional analysis techniques simply cannot exploit them, AI is no longer optional in astronomy, it is the only viable path forward. The same story plays out in banking, where AI spots fraud patterns buried in millions of transactions that no human analyst could track in real time, and in cybersecurity, where systems monitor network behavior continuously in ways no team of people could sustain.

    The pattern is consistent: wherever the bottleneck is volume, repetition, or finding signals buried in overwhelming amounts of data- AI wins. Not because it understands, but because it is fast, it doesn’t get tired, doesn’t get bored, and doesn’t need the work to mean something.

    How Are We Adopting AI in Our Organizations?
    Leaving the exceptions aside, walk into an organization today and you will find the same scene. Dashboards tracking AI usage. Leadership decks celebrating adoption milestones. Someone in every meeting mentioning their team’s token consumption like it’s a fitness score.
    There is AI utopia everywhere. But scratch beneath the surface and the picture gets uncomfortable.

    Part of the problem is what we are choosing to measure. Many organizations have settled on token usage as their proxy for AI maturity- how much are people using it, how often, at what volume. It is not entirely wrong, but it is dangerously incomplete. A metric that can be gamed will be gamed! And so, a quiet, troubling trend has emerged: people inflating usage counts to meet mandates and hit performance targets optimizing the scoreboard rather than developing the skill. Microsoft’s 2025 Future of Work research captures why this happens: workers resist top-down AI mandates that prioritize efficiency above quality and creativity. When adoption is imposed rather than understood, compliance replaces genuine engagement.

    The cultural debate around all this mirrors a very human pattern. In some teams, using AI is quietly frowned upon- seen as laziness, as cheating, as a shortcut that real professionals don’t need. In others, not using it is frowned upon- seen as resistance, as falling behind, as naivety. People get criticized no matter what they choose. But this debate is a distraction from the question that actually matters: is any of this building genuine capability, or just generating activity?
    The cultural pressure lands hardest on the people who can least afford it- the youngest people in the workforce.

    A fresh graduate joining a team today is unlikely to be told: here is a problem, work through it, struggle with it, come back when you have tried. They are more likely to be handed an AI tool, pointed at a token target, and measured on throughput. The uncomfortable question nobody is asking in these organizations is: what exactly is being learned? Because what looks like productivity is often something else entirely- the fluent reproduction of AI output by someone who does not yet have the foundation to know whether it is right, relevant, or dangerously incomplete.
    This is how a generation develops what might be called AI dependency before expertise. Not through laziness- these are motivated, capable people but through a system that optimizes for output before capability is built. They learn to prompt before they learn to think. They learn to review AI-generated code before they have written enough of their own to understand what they are reviewing. They learn to ask the machine before they have developed the instinct to ask themselves. Do you recognize that ‘Let me google or I’ll ask AI to write that quick email’ feeling?

    The numbers quietly confirm what is happening on the ground. Early-career workers aged 22–25 in AI-exposed roles have seen a 13% relative decline in employment. Entry-level postings have been cut in some sectors. Organizations are eliminating the very roles that used to be where expertise was assembled- the unglamorous, repetitive, foundational work that turned a graduate into someone who actually understood how things functioned. Are we removing the on-ramp and preparing ourselves to wonder in future why people can’t navigate the highway.

    What makes this particularly dangerous is that the loss is invisible in the short term. A junior analyst who has never learned to structure an argument can still produce a well-structured document with AI assistance. A developer who has never debugged their own code can still ship working software. The output looks fine. The capability gap underneath it doesn’t show up until the moment it matters most- when the AI is wrong, when the problem is novel, when nobody in the room actually understands what the system is doing or why.

    We are, in effect, building a generation of very capable operators of AI and a shrinking number of people who can think independently of it. That is not a workforce problem for five years from now. It is being quietly assembled today, one unguided adoption mandate at a time.

    What Happens Next And What Governance Do We Need?
    “To AI or not to AI?” that’s not the question here. It’s not a “mind or machine” fight. We are standing at a “mind aided by machine” juncture. But that juncture requires deliberate choices because the default path, left unmanaged, does not lead somewhere good.

    Here is a paradox worth sitting with. Employers anticipate that 39% of core skills will change by 2030. In response, 82% of enterprise leaders say they are already providing AI training. Yet 59% still report a significant and growing skills gap. More training, same gap. The reason is not hard to find: teaching someone to use an AI tool is not the same as building the underlying capability the tool is meant to augment. We are handing people a calculator and calling it a mathematics education. Teach how to ask ChatGPT to write code, but don’t teach what are unique ways in which AI writes code and how the loopholes can be found in it.
    The organizational risk that follows is not abstract. ISACA describes it plainly a company that routinely relies on AI to complete daily tasks may gradually lose its own understanding of how those tasks actually work. Personnel stop knowing how the system functions. Decisions get made by people who can evaluate AI output but no longer fully understand what lies beneath it. That is not a productivity gain. That is institutional fragility dressed up as efficiency.

    So, what does responsible adoption actually look like? Here are not abstract principles but concrete shifts organizations need to make most of which cost nothing except the willingness to think beyond the quarterly adoption scorecard.
    Sequence AI access by experience, not just by role- The most consequential governance decision an organization can make is deciding when in someone’s development journey AI tools become available for a given task. A junior developer should write code, break it, debug it, and understand it before AI scaffolds the solution for them. A new analyst should structure their own argument before AI polishes it. Junior professionals should be able to demonstrate and articulate a thought process before AI aids their document writing. This is not about withholding tools, it is about sequencing capability before convenience. Practically, this means defining, for each role and each skill domain, what foundational experience must be demonstrated before AI assistance is unlocked for that task. Think of it like a driving license- you learn to control the vehicle before you get cruise control.

    Redefine what you measure- Token counts and completion rates measure activity, not capability. Organizations serious about sustainable AI adoption need a second set of metrics alongside usage data: are people making fewer errors over time, or just better-looking ones? Are junior staff developing independent judgment, or deepening dependence? Are code reviewers able to explain and challenge AI-generated output, or simply approving it? Simple interventions help here- periodic “no AI” tasks where individuals solve problems independently, used not as a punishment but as a calibration tool to surface where genuine skill exists and where it has quietly eroded.

    Redesign the learning pyramid, not just the org chart- The traditional pyramid assumed that volume of work is done at the bottom to build the expertise that rose to the top. AI has hollowed out that bottom layer. Now the work done by the junior most workforce is getting replaced by AI, you don’t need developers, you need reviewers! Isn’t the org pyramid already becoming a rhombus? It’s not a stable structure that can sit on the ground. Skill building was happening organically, organizations now need to deliberately build it through structured apprenticeship programs, mandatory code-writing rotations before code-review roles, writing workshops before AI-assisted documentation becomes standard, and design-from-scratch exercises before templates and generators are handed over. This is not nostalgia. It is pipeline maintenance. The senior reviewers, architects, and decision-makers of 2035 are the junior staff of today and they need the foundational friction now, while there is still time to build it in.

    Make “AI as sparring partner” a cultural norm, not an exception- The most valuable shift in how AI gets used day-to-day is not in the tools themselves but in the posture. Organizations should actively train and reward the habit of challenging AI output rather than consuming it. This means asking: what is the AI missing here? What assumption is it making? What would I have done differently and why? Some forward-thinking teams have already introduced a simple rule: before submitting any AI-assisted work, the person must be able to explain and defend every substantive claim or decision in it. If they cannot, it goes back. This single norm, consistently applied, transforms AI from a shortcut into a genuine learning accelerator.

    Build governance architecture before the debt accumulates- Governance retrofitted onto embedded systems is far harder than governance designed in from the start. At minimum, every organization deploying AI at scale needs three things clearly defined: who owns the quality of AI output in each domain, what the human review standard is before that output is acted upon, and how capability development is tracked independently of AI usage metrics. Without these three, adoption is just exposure and exposure without structure builds dependency, not skill.

    It is very well recognized that AI is powerful and meant to amplify skills. But, if there are no skills, it amplifies nothing. Like a powerful steam engine attached to nothing. It just gives you answers and quietly builds your future risk profile.
    Do not let AI create an illusion of false mastery. In my opinion, the organizations that will lead in ten years are not the ones that adopted AI fastest. They are the ones that kept their people genuinely capable while doing it. And if they did, we know AI will create new kind of genius and in fact accelerate an “Aryabhatta“
    Every major technology in history triggered fears about cognitive decline. Writing was once criticized for weakening memory. Calculators were accused of destroying arithmetic ability. Search engines changed how humans remember information.

    AI may alter how we think- but alteration is not destruction. Governed with vision, it redefines where human genius gets to focus rather than replacing it. A modern Aryabhatta won’t compete with AI. They will use it to clear the path, so their mind is free for the question nobody has thought to ask yet. That is still a human job. Probability distributions don’t ask questions. They answer them.
    The real risk is not that AI thinks for us. It is that we raise a generation that never learns to think without it. The thinkers who change the world are not the most efficient ones. They are the ones who chose difficulty when they didn’t have to. That person is still possible. Whether the world we are building still produces them- that is the question worth sitting with.

    References: