The Olorin: The Human That AI Was Built to Reward
Why the Most Dangerous Myth in AI Is That Anyone Can Use It Equally Well
There is a quiet lie circulating through boardrooms, classrooms, and Twitter threads: that AI is the great equalizer. That it flattens the gap between the novice and the expert. That a teenager with ChatGPT is now equivalent to a seasoned strategist with thirty years of pattern recognition burned into their nervous system.
This is wrong. And understanding why it’s wrong is the key to understanding who will actually capture the asymmetric value that AI creates.
AI does not replace human depth. It rewards it. The returns on AI are not distributed equally — they concentrate violently in the hands of a specific kind of person. Not the most technical. Not the youngest. Not the most “plugged in.” The person who extracts maximum value from AI is the one who brings the richest internal world to the conversation.
We need a name for this person. Not “power user.” Not “prompt engineer.” Something that captures the full dimensionality of what makes them different.
In Tolkien’s mythology, before Gandalf ever arrived in Middle-earth — before he was the Grey Pilgrim, before he was the wizard who changed the fate of kingdoms — he was Olórin. A being who, as Tolkien wrote, “walked among the Elves unseen, or in form as one of them, and they did not know whence came the fair visions or the promptings of wisdom that he put into their hearts.”
Olórin didn’t command. He didn’t teach from above. He moved among others as a peer, and the quality of the thinking around him simply elevated through proximity. He was the wisest of his order — not because he held rank, but because he carried the deepest knowledge, the broadest compassion, and the most refined perception. He was a wanderer, a thinker, an observer. Self-directed. Knowledge-loaded. Polymathic. Unhurried.
That is the archetype.
The Olorin is the human who makes AI transformative — not through technical skill, but through the accumulated depth of a fully lived intellectual life. They are the wanderers between domains who carry hard-won wisdom into every prompt, every evaluation, every collaboration with a machine that is powerful but directionless without them.
What follows is not speculation. Each of the five dimensions that define the Olorin is grounded in research — from cognitive science, linguistics, expertise studies, and the emerging field of human-AI collaboration. The evidence is clear: the humans who get the most from AI are the ones who brought the most to the table before AI existed.
The Five Dimensions of the Olorin
1. Linguistic Precision — The Deep Vocabulary
The first and most underestimated advantage is language itself.
AI operates on language. It parses meaning through tokens, but meaning is not evenly distributed across words. The person who says “make this better” and the person who says “tighten the cadence, cut the Latinate abstractions, and let the Anglo-Saxon verbs carry the weight” are not making the same request. They will not get the same output. They are playing different games entirely.
This is not merely intuition. Research in prompt engineering has demonstrated that LLM performance is remarkably sensitive to how instructions are phrased. A comprehensive review published in Nature Partner Journals found that detailed, precise prompts enable models to generate content far more aligned with specific requirements by reducing the model’s internal uncertainty and guiding it toward the correct response space (Sahoo et al., 2025, PMC). Meanwhile, a 2023 study from the ATLAS Lab demonstrated that well-engineered prompts can increase model accuracy by 57% on LLaMA models and 67% on GPT-4, with the quality of improvement hinging on the specificity and structural clarity of the language used (Bsharat et al., 2023).
The theoretical backing runs even deeper. The weak version of the Sapir-Whorf hypothesis — the principle of linguistic relativity — proposes that the structure and vocabulary of a person’s language influences their perception and cognition. While the strong deterministic version has largely fallen out of favor, modern research has produced substantial empirical evidence that linguistic categories shape how speakers perceive and categorize experience, without strictly limiting them (Gumperz & Levinson, 1996). As Edward Sapir himself argued, language does not merely reflect reality — it participates in shaping how we apprehend it.
Applied to AI interaction, this means something profound: the Olorin with a richer vocabulary doesn’t just communicate more efficiently with a language model. They think in higher resolution. When a materials engineer says “anisotropic” instead of “it behaves differently depending on direction,” they are compressing an enormous amount of information into a single term that the AI can unpack with total fidelity. The vague prompt produces vague output. The precise prompt produces something that looks like magic — but it is not magic. It is vocabulary doing its ancient job: carrying thought.
The research on ExpertPrompting (Xu et al., 2023) demonstrated this principle at the system level: when LLMs were given detailed, domain-specific expert identity descriptions — essentially, when the prompt itself contained the kind of precise language and contextual framing that an Olorin would use — the quality of outputs improved dramatically, achieving 96% of ChatGPT’s capability with open-source models. The expert’s vocabulary, encoded in the prompt, was the active ingredient.
This is why writers, lawyers, scientists, and anyone trained in a tradition of careful naming tend to get disproportionate value from AI. They have spent years sharpening the instrument — language — that AI runs on.
2. Historical Context — The Weight of Experience
A twenty-three-year-old can ask an AI to draft a go-to-market strategy. A fifty-year-old who has watched three market cycles, two recessions, and the slow death of a company she loved can ask an AI to draft a go-to-market strategy and then interrogate every assumption in the output against a library of lived consequence.
Experience is not nostalgia. It is a pattern library. Every failure, every surprise, every “I’ve seen this before” moment is a data point that the human carries but the AI does not. The AI has read about recessions. The experienced operator has felt the room change when a CFO realizes the burn rate is unsustainable. These are different kinds of knowledge, and the second kind is what allows a human to steer AI output away from the theoretically elegant and toward the practically survivable.
K. Anders Ericsson’s landmark research on expert performance provides the theoretical scaffold here. In his 1993 paper with Krampe and Tesch-Römer in Psychological Review, Ericsson demonstrated that expert performance is not a function of raw talent or even simple experience accumulation, but rather the product of prolonged engagement in deliberate practice — focused training aimed at improving specific tasks, with immediate feedback, problem-solving, and opportunities for repeated refinement (Ericsson, Krampe & Tesch-Römer, 1993). Crucially, Ericsson later showed that observed performance does not necessarily correlate with greater professional experience alone. What matters is the quality of the engagement — whether the individual actively processed feedback, confronted failures, and built mental representations of complex situations (Ericsson, 2008).
This finding maps directly onto AI interaction. The Olorin who has done deliberate practice — who has not merely accumulated years, but has actively processed those years through cycles of hypothesis, failure, and correction — brings three irreplaceable powers to the human-AI collaboration:
Anomaly detection. They notice when AI output pattern-matches to a common template but misses the specific wrinkle that makes this situation different. A 2024 study published in Scientific Reports demonstrated this clearly: domain experts augmented with explainable AI significantly outperformed those working with opaque “black-box” AI, precisely because their expertise allowed them to validate or overrule AI predictions against their own knowledge. Factory workers with an average of 13.8 years of experience were able to follow accurate AI predictions and catch the AI’s errors — but only when the system was designed to let their expertise participate in the process (Feuerriegel et al., 2024).
Consequence modeling. They can trace a recommendation forward through second and third-order effects because they have watched those chains play out before. AI generates outputs that are internally consistent. The Olorin knows that internal consistency is cheap — external consistency with reality is what matters.
Calibrated trust. They know when to accept AI output at face value and when to probe, because they have been fooled by confident-sounding nonsense before — from humans and now from machines. Recent research from Aalto University (Fernandes et al., 2026, Computers in Human Behavior) has revealed a troubling “reverse Dunning-Kruger effect” in AI use: participants who considered themselves more AI-literate were actually more likely to overestimate their performance when using ChatGPT. Everyone overestimated, but the supposedly savvy users overestimated the most. The Olorin’s hard-won skepticism — their calibrated trust, born of being wrong in consequential ways — is a direct defense against this trap.
The AI does not know what it does not know. The Olorin has a catalog of what tends to go wrong, and that catalog is irreplaceable.
3. The Polymathic Knowledge Base — Breadth and Depth Together
The most powerful AI users are not specialists. They are not generalists either. They are something rarer: deep generalists — people who have gone far enough into multiple domains to understand the structural grammar of each one.
This matters because AI is phenomenally good at operating within a single domain but struggles at the seams between domains. It can write excellent legal language. It can write excellent marketing copy. But knowing when a legal constraint should reshape a marketing strategy — that is a human judgment call that requires competence in both worlds simultaneously.
David Epstein’s research, synthesized in his #1 New York Times bestseller Range: Why Generalists Triumph in a Specialized World (2019), provides the most compelling evidence for this dimension. Epstein examined the careers of the world’s most successful athletes, scientists, artists, and inventors and found that in most fields — especially those characterized by complexity and unpredictability — generalists, not specialists, are primed to excel. They find their paths late, juggle many interests rather than focusing on one, and are more creative, more agile, and able to make connections their specialized peers cannot see.
Epstein draws a critical distinction between “kind” learning environments — where rules are clear, patterns repeat, and feedback is immediate, like chess — and “wicked” environments, where rules are unclear, feedback is delayed or misleading, and novel situations are the norm. Most real-world problems, and certainly most business problems, are wicked. In wicked environments, the person with range outperforms the deep specialist because they can import mental models from adjacent fields, recognize structural analogies across domains, and avoid the “model blindness” that comes from having only one lens.
The data on innovation reinforces this. Epstein cites research showing that the most impactful inventors crossed domains rather than deepening their knowledge in a single area. A study of comic book creators found that the key predictor of creative quality was not years of experience, not workload, and not resources — it was how many different genres a creator had worked in. The polymaths in patent data had depth in a core area but also breadth across dozens of technology classes, and they consistently produced the most innovative work.
Applied to AI interaction, the implications are direct. The Olorin sees connections that the AI will not volunteer. They say things like: “This financial model assumes linear adoption, but the behavioral economics literature on loss aversion suggests the curve will flatten earlier — adjust the projection and show me the sensitivity analysis.” That prompt bridges three disciplines in a single sentence. The AI can execute it brilliantly, but only a human with footing in all three could construct it.
Knowledge breadth also expands the solution space. A narrow specialist asks AI to solve the problem within their domain. An Olorin asks AI to solve the problem using tools borrowed from adjacent or even distant fields. As physicist Freeman Dyson wrote, we need both focused frogs and visionary birds — frogs who see the details of particular objects and birds who survey broad vistas and bring together diverse problems from different parts of the landscape. The Olorin is the bird who can also become the frog when needed.
Research on human-AI complementarity (Hemmer et al., 2022; Fügener et al., 2021) further supports this: the value of human-AI collaboration emerges from information and capability asymmetries between the two agents. The human who brings the broadest knowledge base creates the largest asymmetry — and therefore the most value — in the collaboration.
4. Trial and Error — The Scar Tissue of Real-World Iteration
There is a category of knowledge that cannot be acquired through reading, instruction, or even observation. It can only be acquired by doing the thing and getting it wrong enough times that the wrongness itself becomes information.
Olorin tend to be people who have built things, shipped things, broken things, and repaired things. They carry what we might call operational intuition — a sense for what will actually work when it meets the friction of reality. This is the difference between a recipe and a cook. The recipe is information. The cook has burned enough sauces to know that the recipe’s instruction to “stir occasionally” actually means “stir constantly for this particular pan on this particular stove.”
Ericsson’s deliberate practice framework is again relevant here, but with an important nuance. Ericsson’s research showed that the critical differentiator in expert performance was not passive experience but active engagement with feedback loops — what he called the cycle of “cognitive” and “associative” phases before reaching genuine expertise (Ericsson, 2008). The Dreyfus model of skill acquisition (Dreyfus & Dreyfus, 1980) describes a parallel journey from novice to expert, with the critical inflection point being where the practitioner stops relying on abstract rules and begins operating from pattern recognition grounded in extensive situational experience.
This kind of knowledge transforms how a person interacts with AI in a critical way: it gives them the ability to evaluate outputs against reality rather than against internal coherence. AI is extraordinarily good at producing outputs that are internally consistent, well-structured, and logically sound — but completely wrong in practice. The Aalto University research on AI and the Dunning-Kruger effect powerfully illustrates the danger: when 240 participants used ChatGPT to solve logical reasoning problems from the Law School Admission Test, they consistently overestimated their performance. Most participants simply copied the question into the AI, accepted the output, and moved on — what researchers described as “cognitive offloading” without verification (Fernandes et al., 2026).
The person who has tried and failed in the real world has a filter that catches these beautiful errors. The person who has not is defenseless against them. As the researchers noted, the AI interaction eliminated the traditional Dunning-Kruger pattern where experts self-correct through metacognitive awareness. Instead, everyone — novice and experienced alike — overestimated. The only defense is the kind of hard-won skepticism that comes from having watched confident-sounding plans fail in practice.
Trial and error also builds prompt iteration instinct. Experienced builders know that the first attempt rarely works. They are comfortable with rapid cycles of try, evaluate, adjust. They treat AI not as an oracle that should produce the right answer on the first attempt, but as a collaborator that requires steering — and they are patient and skilled enough to do the steering. This mirrors the iterative refinement process that prompt engineering researchers have found essential: the most effective AI interactions are not single-shot queries but structured conversations where the human progressively narrows the output toward the desired result through skilled feedback (Sahoo et al., 2025).
5. Reference Dimensions — The Ability to Navigate Conceptual Space
This is the most abstract of the five dimensions, and arguably the most powerful.
Every problem exists in a space defined by multiple axes — dimensions along which the solution can vary. A branding problem has aesthetic dimensions, cultural dimensions, competitive dimensions, linguistic dimensions, and emotional dimensions. A technical architecture problem has dimensions of scalability, cost, latency, maintainability, and team capability.
The Olorin has the ability to name these dimensions explicitly and direct the AI along specific vectors within the solution space. They say things like:
- “Move this in the direction of more formality but less distance.”
- “Optimize for latency at the expense of storage cost, but only up to a 20% increase.”
- “This needs to feel authoritative to a technical audience but accessible to a board member.”
Each of these instructions is a vector — a direction and magnitude within a multi-dimensional space. The person issuing these instructions is performing a cognitive act that AI cannot do for itself: they are choosing which tradeoffs matter and how far to push in each direction. This is not prompt engineering in the mechanical sense. It is navigation. And it requires having a mental map of the space in the first place.
The cognitive science literature on mental models supports this framework. Johnson-Laird’s mental models theory (1983) posits that humans reason by constructing and manipulating internal representations of situations — essentially, internal simulations of how things work. Experts in any domain possess more numerous, more detailed, and more accurate mental models than novices. When an expert designer says “increase the visual weight of the CTA without making it feel aggressive,” they are navigating within a mental model of visual design that contains dozens of interacting dimensions — weight, color temperature, whitespace, proportion, typographic hierarchy — each of which they can independently adjust.
Research on human-AI collaboration from the distributed cognition perspective reinforces why this matters. A 2025 study published in Computers in Human Behavior on “Collaborative AI Metacognition” found that effective human-AI collaboration depends heavily on the human’s ability to manage what the researchers call the “metacognitive tasks” of the collaboration: planning what to delegate, coordinating information sharing, and evaluating the quality of joint outputs. The Olorin who can articulate the dimensional structure of a problem is performing precisely this metacognitive coordination — they are telling the AI what the problem looks like from above so that the AI can operate effectively within it.
People develop reference dimensions through deep expertise, cross-domain experience, and — critically — through having seen many instances of what good and bad look like along each axis. A designer who has reviewed ten thousand interfaces has internalized the dimensions of visual design so deeply that they can articulate micro-adjustments that a novice could not even perceive, let alone name.
This connects back to the linguistic precision dimension in an important way. Reference dimensions must be nameable to be useful in AI interaction. The Olorin who can feel that something is wrong but cannot articulate which dimension is off-target cannot communicate that insight to an AI. The true Olorin is the person who has both the perceptual sophistication to detect dimensional misalignment and the vocabulary to express it.
The Compound Effect: Why These Dimensions Multiply
These five dimensions do not add. They multiply.
An Olorin with deep vocabulary and historical context writes prompts that are both precise and grounded. An Olorin with a broad knowledge base and real-world trial and error spots cross-domain opportunities and knows which ones will survive contact with reality. An Olorin who can navigate reference dimensions and has the language to articulate those dimensions gives the AI instructions of extraordinary resolution.
Remove any one dimension and the output degrades — not linearly, but geometrically. A person with vocabulary but no experience produces prompts that are precise but naive. A person with experience but no breadth solves problems within their silo and misses the lateral move. A person with breadth but no trial-and-error knowledge generates interesting ideas that collapse on implementation.
The research on human-AI complementarity supports this multiplicative model. Hemmer and colleagues’ work on information and capability asymmetries (2022) demonstrates that the value of human-AI teaming comes specifically from the differences between what each agent brings. The human who brings depth in only one dimension creates limited asymmetry. The Olorin who brings all five creates a collaboration partner that the AI cannot replicate or substitute.
The Olorin is the person in whom all five dimensions have developed to meaningful depth. And this reveals an uncomfortable truth about the AI revolution.
The Uncomfortable Implication
AI does not democratize expertise. It amplifies the gap between the prepared and the unprepared.
The Aalto University research made this visible in a laboratory setting: AI improved everyone’s raw performance, but it destroyed everyone’s ability to accurately assess that performance. The classic Dunning-Kruger pattern — where experts self-correct and novices overestimate — vanished entirely under AI use (Fernandes et al., 2026). In the age of AI, the defense against overconfidence is not AI literacy (that made things worse). It is the pre-existing depth of knowledge, experience, and self-awareness that the individual brings to the interaction.
Meanwhile, the ExpertPrompting research showed the flip side: when expert-level knowledge is encoded into the interaction — through precise vocabulary, domain-specific context, and detailed framing — the outputs improve dramatically (Xu et al., 2023). The gap between a vague prompt and an Olorin’s prompt is not 10% or 20%. It can be the difference between useless output and transformative output.
The person who spent twenty years developing vocabulary, accumulating experience, building cross-domain knowledge, learning from failure, and developing the ability to think in multiple dimensions simultaneously — that person was building the exact cognitive infrastructure that AI now rewards exponentially. They did not know they were becoming an Olorin. But they were.
The person who skipped the hard years, who never developed precision in language, who never accumulated the scar tissue of real-world iteration, who never ventured outside their single domain — that person will find AI useful. But they will never find it transformative. They lack the raw material that AI needs to produce extraordinary output.
As David Epstein’s research concludes: in a world of wicked problems — complex, unpredictable, with changing rules — the person with range outperforms the narrow specialist. AI has made every problem wickeder. It has expanded the solution space, accelerated the rate of change, and increased the premium on the kind of thinking that can operate across domains, under uncertainty, with calibrated trust in imperfect information. That kind of thinking is not a product of AI. It is a product of a fully developed human life.
The Path Forward: Becoming an Olorin
The good news is that every dimension of the Olorin can be cultivated. None of them are innate talents. They are all, in Ericsson’s framework, products of deliberate practice:
Linguistic precision improves through reading widely, writing regularly, and studying the specialized vocabularies of multiple fields. Every new domain-specific term you internalize is a new instruction you can give an AI.
Historical context deepens through reflective experience — not merely logging years, but actively processing what happened, why it happened, and what it teaches about the structure of similar situations.
Polymathic knowledge expands through what Epstein calls “sampling” — deliberate exploration of unfamiliar domains, not to become a specialist in each, but to understand their structural grammar well enough to import their tools and frameworks.
Trial-and-error wisdom accumulates only by doing — building, shipping, failing, and rebuilding. There is no shortcut. The scar tissue is the knowledge.
Reference dimensions develop through exposure to many instances of quality — reviewing thousands of examples of good and bad work across multiple domains until the dimensions of evaluation become instinctive and nameable.
The investment thesis of a human life has not changed. The returns on depth, breadth, experience, precision, and hard-won judgment have not decreased in the age of AI. They have increased. Dramatically.
The Olorin is, in the end, simply the most fully developed human — the one who took the long road, did the hard work, accumulated the knowledge, and refined the language. AI was always going to reward the prepared mind.
The only thing that has changed is the magnitude of that reward.
References
Bsharat, S. M., Myrzakhan, A., & Shen, Z. (2023). Principled instructions are all you need for questioning LLaMA-1/2, GPT-3.5/4. arXiv preprint, arXiv:2312.16171.
Dreyfus, S. E., & Dreyfus, H. L. (1980). A five-stage model of the mental activities involved in directed skill acquisition. Operations Research Center, University of California, Berkeley.
Epstein, D. (2019). Range: Why generalists triumph in a specialized world. Riverhead Books.
Ericsson, K. A. (2008). Deliberate practice and acquisition of expert performance: A general overview. Academic Emergency Medicine, 15(11), 988–994.
Ericsson, K. A., Krampe, R. T., & Tesch-Römer, C. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100(3), 363–406.
Fernandes, D., Villa, S., Nicholls, S., Haavisto, O., Buschek, D., Schmidt, A., & Welsch, R. (2026). AI makes you smarter, but none the wiser: The disconnect between performance and metacognition. Computers in Human Behavior, 175, 108779.
Feuerriegel, S., et al. (2024). Explainable AI improves task performance in human–AI collaboration. Scientific Reports, 14, Article 82501.
Fügener, A., et al. (2021). Cognitive challenges in human-artificial intelligence collaboration: Investigating the path toward productive delegation. Information Systems Research, 32(4), 1287–1305.
Gumperz, J. J., & Levinson, S. C. (Eds.). (1996). Rethinking linguistic relativity. Cambridge University Press.
Hemmer, P., et al. (2022). Forming effective human-AI teams: Building machine learning models that complement the capabilities of multiple experts. Proceedings of the AAAI Conference on Artificial Intelligence.
Johnson-Laird, P. N. (1983). Mental models: Towards a cognitive science of language, inference, and consciousness. Harvard University Press.
Sahoo, P., et al. (2025). Unleashing the potential of prompt engineering for large language models. Nature Partner Journals — Science of Learning (PMC/12191768).
Sapir, E. (1929). The status of linguistics as a science. Language, 5(4), 207–214.
Tolkien, J. R. R. (1977). The Silmarillion. Edited by Christopher Tolkien. George Allen & Unwin.
Xu, B., et al. (2023). ExpertPrompting: Instructing large language models to be distinguished experts. arXiv preprint, arXiv:2305.14688.
