The Amoeba in the Graph

Why AI doesn’t rescue the knowledge-less knowledge worker — it exposes them

In the companion piece to this one, I argued that the augmented hyperscaler — the older promiscuous learner with a wide, weathered, multi-domain corpus — is the most leveraged human in the AI economy. Today the inverse claim, which is harder to make politely but more important.

The 25-year-old who shows up to the augmented economy without a bedrock of knowledge — without forty years of accidental cross-domain reading, without the calibrated judgment that comes from being repeatedly wrong in different fields, without a stocked long-term memory that the AI can amplify — is not on a level playing field with the hyperscaler. They’re not even on the same field. In the graph of knowledge that AI now makes navigable, the hyperscaler is a high-centrality hub connecting distant clusters. The 25-year-old without a corpus is a peripheral node with two or three thin edges, drifting across a topology they cannot see, dependent on a tool that confidently tells them which way is up.

They look like an amoeba feeling its way across a graph the hyperscaler is reading like a map.

This is not a generational complaint. It is a structural one, with neuroscience and economics behind it.

The graph problem: AI multiplies your connectivity, not your existence

In any network, a node’s value is a function of its connectivity. High-betweenness nodes — the ones that link otherwise distant regions — are disproportionately consequential; sparsely connected peripheral nodes are not. The augmented hyperscaler is consequential because forty years of cross-domain learning has given them edges into immunology and shipping history and Borges and semiconductors and Mandarin and welding. AI doesn’t create those edges. It accelerates traversal across edges that already exist.

This matters because the dominant cultural story about AI is that it democratizes expertise. The headline finding behind that story comes from Brynjolfsson, Li, and Raymond’s “Generative AI at Work” (NBER 2023, revised 2024), which studied 5,179 customer support agents and found that AI raised average productivity 14%, with novices gaining 34% and experienced workers gaining almost nothing. Read carelessly, this looks like a flattening: AI lifts the bottom and compresses the distribution.

But customer support is the wrong reference task for what hyperscalers do. It is bounded, codifiable, and has a ceiling. AI in that setting is doing the synthesis the senior agent had memorized and handing it to the junior. It’s not asking the junior to navigate a jagged frontier of partly-correct outputs across multiple domains.

The relevant study for that work is Dell’Acqua, Mollick, Lifshitz-Assaf, Kellogg and colleagues’ “Navigating the Jagged Technological Frontier” (2023, Organization Science 2025). On 18 realistic consulting tasks with 700+ BCG consultants, AI raised quality 40% within its capability frontier. But on tasks designed to fall just outside what GPT-4 reliably handles, AI users produced worse results than the no-AI control group. The reason was simple: the AI was confidently wrong, and the user couldn’t tell.

That capacity to tell — to feel the wrongness of an answer in a domain you understand — is exactly what the 25-year-old without a bedrock does not have. Not because they’re not smart. Because the calibration takes a long time to build, in many places, by being wrong over and over, and there is no synthetic version.

You can’t validate what you don’t know — and the data shows you’ll stop trying

The Microsoft Research / Carnegie Mellon study by Lee et al. (“The Impact of Generative AI on Critical Thinking,” CHI 2025) surveyed 319 knowledge workers using ChatGPT, Copilot, or similar tools at least weekly, with 936 first-hand task examples. Two findings, in tension and both important:

Higher confidence in the AI was associated with less critical thinking. Higher confidence in one’s own skills was associated with more.

In other words: the workers most likely to scrutinize AI output were the ones who already knew enough to do the task themselves. The workers who trusted the AI most — disproportionately the less-experienced — engaged in the least cross-checking. The researchers also found, qualitatively, that workers refrained from critical thinking on AI output specifically when they lacked the domain skills to inspect, improve, or guide it.

This is the amoeba problem made empirical. The 25-year-old without a bedrock has nothing to check against, so they don’t check. The hyperscaler has decades of pattern memory to check against, so they do. Same tool, opposite behavior, opposite outcomes.

A counter-argument here is reasonable: well, at least the AI gives the novice a starting point they wouldn’t have otherwise. True. But a confidently wrong starting point is worse than no starting point if you can’t audit it — and the same Microsoft study found that workers using GenAI produced less diverse task outcomes than non-AI counterparts. The convergent, plausible-sounding default of the model becomes the floor of the work, and it stays there unprodded.

Cognitive debt: the brain you don’t use atrophies, and the data is now on EEG

In June 2025, researchers at the MIT Media Lab — Kosmyna, Hauptmann, Yuan and colleagues — published “Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Task.” They divided 54 participants into three groups — LLM users, search engine users, and brain-only writers — across three sessions over four months, with EEG monitoring throughout. A fourth session crossed conditions: LLM users were forced to write without AI; brain-only users were given AI for the first time.

The findings are not ambiguous.

EEG showed brain-only participants exhibiting the strongest, most distributed neural networks. Search engine users showed moderate engagement. LLM users displayed the weakest connectivity across alpha and beta bands. Self-reported essay ownership was lowest in the LLM group and highest in the brain-only group. Most striking: the LLM group struggled to quote from their own essays — 83% could not produce a quote from work they had finished minutes earlier.

The session-four results matter most. When habituated LLM users were forced to write without the assistant, they did not perform like the brain-only group; they performed worse than the brain-only group, with reduced alpha and beta connectivity indicating sustained under-engagement. The cognitive muscles weren’t dormant; they were further atrophied than baseline.

The researchers coined the term “cognitive debt” to describe the pattern: short-term ease, long-term diminishment of critical thinking, creativity, and independent reasoning. The debt accrues silently. The user feels productive. The EEG disagrees.

This is the engine of the amoeba state. A 25-year-old who reaches for AI before reaching for their own mind on every task does not build the substrate that would let them eventually use AI as leverage rather than as substitute. They reach 30 with the corpus of a 22-year-old. They reach 40 with the corpus of a 25-year-old. The compounding the hyperscaler enjoyed runs in reverse.

The chunks problem: pattern recognition without a corpus is impossible

In 1946, the Dutch chess master and psychologist Adriaan de Groot ran a study that would define expertise research for the next eighty years. He found, contrary to lore, that grandmasters did not think further ahead than weaker players. They simply saw the board differently. Briefly shown a position from a real game, masters could reconstruct it almost perfectly. Weaker players could not.

In 1973, William Chase and Herbert Simon (Carnegie Mellon, where Simon would soon win the Nobel) formalized this as chunking theory. Experts, they showed, store thousands of meaningful patterns — “chunks” — in long-term memory through years of exposure. When they look at a chessboard, they don’t see 32 individual pieces; they see a small number of recognizable structures. Novices, lacking chunks, see the pieces. The crucial control: when chess pieces were placed randomly on the board, expert advantage collapsed. The advantage isn’t in the eye. It’s in the library.

Subsequent work by Fernand Gobet, Neil Charness, and others (reviewed in The Cambridge Handbook of Expertise and Expert Performance) has replicated and extended this finding across domain after domain. Pattern recognition is not a free-floating cognitive skill. It is a function of stored chunks acquired through long, structured exposure. Without the chunks, the patterns are simply not visible.

Now consider what cross-disciplinary synthesis actually is. It is the recognition of a chunk in domain A that rhymes with a chunk in domain B, where neither domain alone would have produced the link. The hyperscaler has chunks across dozens of fields. They look at a problem and see structure that the 25-year-old, looking at the same problem, literally cannot perceive — not because of intelligence, but because the chunks are not in their long-term memory to be activated.

AI does not provide the chunks. It provides retrieval across the externalized record. But retrieval is downstream of knowing what to retrieve, why, and whether what comes back fits. Those are chunk-dependent operations. The amoeba metaphor is not insulting — it is mechanically accurate. Without an internalized library of patterns, motion across the graph is undirected. The 25-year-old without a corpus is doing a random walk through information space, and AI is making the random walk faster.

You have to be wrong, in many places, before you can tell when you’re being lied to

There is one more thing that cannot be shortcut, and it is the most important.

Calibrated judgment — the visceral feeling that an answer is wrong even before you can articulate why — is built only by being wrong, repeatedly, in different domains, and remembering it. The hyperscaler has bet on something stupid in 1998 and remembers exactly why. They have made the same category error in three different fields and noticed the third time that it was the same error. They have watched three full hype cycles, lost money on at least one, and now feel a hype cycle in their stomach before their head catches up.

This is the substrate underneath the Dell’Acqua jagged-frontier finding. It is the substrate underneath the Microsoft/CMU critical-thinking finding. It is what allows the hyperscaler to read AI output and feel, in the way you feel a wrong note in a song, that something is off — and to investigate. The 25-year-old does not have this substrate, not because of any failure on their part, but because forty years have not happened yet.

There is no synthetic version. There is no course. There is no certification. The wrongness has to land in your own life enough times, in enough different places, that your nervous system encodes it. AI accelerates the rate at which the unprepared can produce confident-sounding output, but it cannot accelerate the wrongness-detection that has to be earned the slow way.

The amoeba on the graph

Put the strands together.

The 25-year-old without a bedrock arrives at the augmented economy with three structural disadvantages, each backed by a different body of research:

1. Sparse connectivity in the knowledge graph (the chunking literature, from de Groot through Chase, Simon, and Gobet). Cross-domain pattern recognition requires stored chunks across many domains. AI accelerates traversal but cannot supply chunks.
2. No critical-thinking substrate to evaluate AI output (Dell’Acqua et al. on the jagged frontier; Lee et al. on confidence and critical thinking). When the AI is wrong, the unprepared user does not know, does not check, and propagates the error.
3. Active cognitive debt accumulation (Kosmyna et al.’s EEG study). The 25-year-old who outsources thinking before building the substrate doesn’t just fail to gain — they actively lose neural engagement that does not return cleanly even when AI is removed.

Put plainly: AI is a multiplier on whatever cognitive substrate you bring to it. The hyperscaler is multiplied by forty years of cross-domain compounding, which is a large positive number. The 25-year-old without a corpus is multiplied by a small one — and may, on the cognitive-debt evidence, be multiplied by a quantity that is degrading rather than growing.

That is what the amoeba metaphor is pointing at. The amoeba moves — it has motion, even speed — but the motion is reactive, undirected, and driven by gradients in its immediate environment rather than by any internal map of the space. Place an amoeba on a vast graph and watch it for an hour and it will appear busy. It will not arrive anywhere meaningful. The hyperscaler placed on the same graph traverses with intent, because the graph is already in their head in compressed form, and the AI is helping them externalize and traverse what they already understand.

What the 25-year-old should actually do

The honest version of this argument is not that young knowledge workers are doomed. It is that the path forward is exactly opposite to what the technology seems to invite.

The invitation is: outsource thinking, get more done, look productive. The actual path is: build the corpus the slow way, deliberately, in many directions, while resisting the cognitive debt the tools accumulate by default. Read widely, in domains that don’t pay you. Get things wrong on purpose. Argue with the model rather than accepting its first plausible output. Write without it sometimes — the Kosmyna data is unambiguous that brain-only writing builds neural networks that AI-only writing degrades. Treat AI as a peer to interrogate, not a parent to obey.

The forty years are coming whether or not you do the work. The only question is whether you arrive at 65 as a hyperscaler with a stocked mind that AI now amplifies — or as a 65-year-old amoeba, having spent four decades letting the tool do the chunks for you, with nothing in your own long-term memory for the next iteration of the tool to leverage.

The augmented hyperscaler is the upside case for AI. The empty-node 25-year-old is the downside case. Both are visible in the data right now. Which one any given young person becomes is a function of what they do with the next decade — and specifically of whether they understand that the tool they are using rewards depth and punishes its absence, however productive it makes the absence look.

About the Author

Brian Curry is a Kansas City-based Data Scientist, AI expert, and founder of Vector1 Research, where he explores the frontier of Artificial Intelligence, Causal Inference, Agentic Systems, and Economic Modeling.

His work focuses on building intelligent systems that reveal the causal structures driving performance across marketing, content, and economics. He is the creator of PyCausalSim (causal discovery through simulation), Papilon (complex system optimization), Daedalus (causal and economic analysis), and Memory-Node Encapsulation (data structures for artificial episodic memory).

Sources

• Brynjolfsson, E., Li, D., & Raymond, L. (2023, rev. 2024). Generative AI at Work. NBER Working Paper №31161.
• Dell’Acqua, F., McFowland III, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023, pub. 2025). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Organization Science.
• Lee, H., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. CHI 2025, Yokohama.
• Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv:2506.08872. MIT Media Lab.
• de Groot, A. D. (1946 / 1965). Het denken van den schaker / Thought and Choice in Chess.
• Chase, W. G., & Simon, H. A. (1973). Perception in Chess. Cognitive Psychology, 4(1), 55–81.
• Gobet, F., & Simon, H. A. (1996, 2000). Templates in Chess Memory: A Mechanism for Recalling Several Boards. Cognitive Psychology and follow-ups; reviewed in The Cambridge Handbook of Expertise and Expert Performance.
• Hartshorne, J. K., & Germine, L. T. (2015). When Does Cognitive Functioning Peak? The Asynchronous Rise and Fall of Different Cognitive Abilities Across the Life Span. Psychological Science, 26(4), 433–443.
• Jones, B. F. (2009). The Burden of Knowledge and the “Death of the Renaissance Man”: Is Innovation Getting Harder? Review of Economic Studies, 76(1), 283–317.