Knowledge Is Not Experience, and AI Cannot Collapse the Difference
I knew what the Black Hills looked like before I saw them.
The Country Beyond the Map
I knew what the Black Hills looked like before I ever saw them. I had seen the photographs and watched the videos. I knew the geography, where the roads went, how the Badlands were formed, where prairie gave way to rock and where the rock eventually rose into forest. I could pull the entire route up on a map and move across it with my finger. Then, on a 26-day drive from Delaware to Washington and back, I actually went there.
The distinction arrived almost immediately. A map lets you hold South Dakota in your hand. Driving across it does not. Hour after hour of prairie changes what distance means. The horizon stops being scenery and becomes a physical fact. You can watch weather happening fifty miles away. Towns that looked close together on the map become islands separated by miles of land. The geography begins to make sense differently because you are no longer looking down at it. You are moving through the space that connects everything you thought you understood, and eventually the Black Hills rise out of all that open country in front of you.
I knew they were there and knew what they looked like. None of that prepared me for their arrival. The road changed, the air changed, the smell changed. Trees closed around us after all that open country and the wind moved differently through the terrain. My attention changed with it, and something in me changed with it too. The change did not wait for reflection to clean the experience up and put it into language. It happened when I got there.
From that point until we finally left that part of the country, it kept happening. The Badlands were part of it, but this was not one grand moment at one famous overlook. It was relentless. Ten minutes down the road there would be another formation, another view, another impossible stretch of country. We would stop, get out, look, and the tears would come. There was barely enough time to absorb one thing before the country presented another.
I texted a friend from the road, “Yeah, it’s been like, how many sights of beauty can you see in a day that break you open.” Those were the words I had in the moment, and they remain better than anything cleaner I could manufacture afterward. “Beautiful” was not enough for what was happening. There are photographs of the Badlands technically better than anything my eyes can produce, made with better lenses and better light by people who knew exactly where to stand and when. They are accurate, but accuracy turned out to be a much smaller category than experience.
A photograph can preserve the formation. It cannot give you the moment when you step out of the vehicle and discover that whatever internal frame you carried for how beautiful the world could be is suddenly too small. The heat matters, and the wind, and the dry smell of the earth. So does the enormous sky, the silence underneath the small human sounds around you, the body that has already traveled hundreds of miles, and the last place still working on you when the next one appears around a bend.
There was something almost violent in beauty at that scale. Not violence as harm, but rupture. Something exceeded the proportions I had brought with me, and the emotional response arrived before I had finished deciding what I thought about any of it. It tore me down and built me back up at the same moment. I left each place carrying something I had not arrived with.
Across many traditions and disciplines, mythology, religion, poetry, philosophy, and aesthetics have developed language for encounters like this, moments when ordinary categories become inadequate to what someone has encountered. Different traditions have explained those moments through symbols, cosmologies, ideas of the sacred, or concepts such as the sublime. I do not need to settle the metaphysics to recognize the recurring human pattern underneath them. Something is encountered, the existing frame proves too small, emotion outruns description, and the person who leaves is no longer exactly the person who arrived.
For me, that was happening in ten-minute intervals. The person who saw the next formation had already been changed by the last one, so the next experience did not arrive in the same person who had entered the region. It arrived in someone whose sense of scale, beauty, distance, place, and self had already moved. Then the frame moved again. By the time we left that part of the country, I did not merely possess more information about it. The country had changed the person carrying the information.
What Experience Actually Is
Two claims sit close together here, and they should not be confused. The first is categorical. A representation of an experience does not become the experience represented, however complete the representation becomes. The second is empirical. Current language models give us no demonstrated reason to conclude from their generated language alone that they undergo the states they describe. The first claim does not depend on the second.
We flatten something important when we talk about knowledge and experience as though one were simply a higher-resolution version of the other. Representations can change us. A book can alter a life. A film can change how someone understands war, love, grief, or themselves. A conversation can rearrange a person’s future. Direct physical encounter does not own transformation.
Being changed by a representation of something is still different from encountering the thing represented. Reading about grief can change you without giving you the experience of grieving. Studying combat can alter how you understand violence without putting you under fire. Learning everything humans have recorded about the Black Hills can change how you understand them without giving you the experience of watching them rise after hours of prairie while your own body is sitting inside the scale you previously knew only as numbers.
Knowledge by representation and participatory contact with reality can inform one another, prepare one another, and correct one another. They are not interchangeable. Increasing the fidelity of the representation does not make the representation become the encounter, and artificial intelligence (AI) is making that distinction harder to see precisely because it is making representation so powerful.
We can give models more books, images, video, context, telemetry, tools, and sensors. We can connect systems to cameras, microphones, radar, lidar, temperature sensors, chemical sensors, accelerometers, Global Positioning System (GPS) data, and whatever comes next. We can build increasingly complete records of what exists and increasingly capable systems for reasoning across those records. That can make the map astonishingly good without making the map become the place.
Consider an autonomous vehicle crossing the same terrain. It can perceive portions of the roadway with precision no human driver can match, continuously measuring movement, geometry, velocity, distance, lane position, and objects around it. Add enough instrumentation and it could record temperature, humidity, vibration, sound, air chemistry, barometric pressure, and hundreds of variables that would never reach my conscious attention. The vehicle could leave behind a record more complete in many measurable respects than my memory ever could. Completeness of measurement does not produce the event that occurred between the landscape and me.
A thermometer can register heat without being hot. A microphone can record thunder without being startled. A camera can capture the precise light falling across a landscape without anything inside the camera being overwhelmed by the sight. Exhaustion exposes the same distinction from the other direction because it is difficult to confuse once we look at what exhaustion actually is.
The Model Has No Need
A human being becomes exhausted because an organism is spending itself. Energy is consumed, water is lost, muscles fatigue, attention degrades, and the body’s internal systems respond. Eventually the body begins making demands because continued activity has consequences for the organism itself. Hunger, thirst, pain, fatigue, and sleep arise from a system whose continued existence depends upon responding to them.
A large language model (LLM) does not have that relationship to the infrastructure running it. Datacenters consume electricity. Cooling systems consume water. GPUs produce heat. Hardware wears out. Those material costs are real, but they do not become needs inside the model.
The datacenter can be thirsty. The model cannot.
If cooling fails, equipment may overheat. If the electricity disappears, computation stops. There is no internal deprivation experienced by the language model before the stop. It does not become hungry, tired, frightened, or increasingly aware that something necessary for its continued existence is being taken away. The infrastructure has resource requirements. The model does not experience those requirements as needs.
Call them psychic organs, and I do not mean anything supernatural by that. I mean the functional architecture through which sensation acquires consequence, affect, memory, meaning, continuity, and eventually some place inside a life. Human beings experience the world through bodies with internal states. We have needs and attachments. We carry fear, pleasure, pain, fatigue, memory, expectation, and an awareness that things can be lost.
What happened yesterday alters the meaning of what happens today. A previous failure can change how uncertainty feels. A previous loss can change the weight of a new risk. A landscape encountered ten minutes ago can change the person looking at the next one. The world does not merely register in us. It matters to the one undergoing it, and that mattering becomes part of what is carried forward.
This is an empirical claim about current systems, and empirical claims should have conditions under which they move. I would change my judgement if a system demonstrated persistent identity, durable consequence that survived removal of external state, stable agency under resistance, and a coherent self-model under controlled testing. I have proposed measurable tests for those properties in the Five Gates for Stakebearing Interiority precisely because claims about interiority should produce evidence rather than demand belief.
Current LLMs can describe human states with extraordinary fluency. An LLM can write about grief without grieving, explain exhaustion without becoming tired, and generate a moving account of awe without anything being overwhelmed by the generation of those words. It may even produce language that helps a human understand their own experience better, and I have no problem granting the value of that capability. What I reject is the upgrade from an increasingly convincing representation of first-person experience to evidence that a first-person experience exists behind the representation.
First-person grammar does not establish first-person experience. Generated descriptions of fear, fatigue, desire, grief, or need are not themselves evidence that those states are being undergone by the model. I therefore see no basis for granting current LLMs the rights of persons merely because their simulations of personhood become persuasive. That position does not require solving every philosophical problem of consciousness. It requires refusing to treat resemblance in language as proof of the thing the language resembles.
None of this requires treating AI systems carelessly. Moral hygiene in how humans design, use, and interact with artifacts is a different question from whether the artifact itself is a moral subject. The accountability question remains human because people and organizations choose the architecture, deployment, permissions, risk thresholds, and systems into which AI is inserted. They decide when these systems run, what they may touch, and capture the economic value created around them. When something goes wrong, responsibility should move toward that chain rather than disappear into language suggesting that “the AI decided.” Calling a model creative likewise does not erase the human creators whose work contributed to systems now able to reproduce portions of their craft at scale.
When Knowledge Gets Cheap
The distinction matters beyond arguments about machine experience because AI is changing the economics of knowledge whether or not we ever settle the philosophy. If representation can become extraordinarily capable without becoming experience, the question for human development changes. We should spend less time asking whether machines can give people enough information. Increasingly, they can. The harder question is what becomes scarce when information no longer is.
Experience is one answer. Consequence is another. Judgement forms through the interaction between them.
AI may become one of the greatest knowledge-transfer technologies humans have ever built. A junior engineer can ask about a distributed systems problem and receive in minutes what might once have required finding the right expert, book, or mailing list. A student can cross disciplinary boundaries much faster. An operator can search thousands of pages of documentation while an incident is unfolding. A traveler can understand the geology, history, ecology, and cultural significance of the Black Hills before ever leaving home.
We should use that capability aggressively. Knowledge should not remain expensive merely because previous generations paid more for access to it. For much of history, specialized knowledge often required proximity to a book, school, guild, teacher, laboratory, or experienced practitioner. Scarcity made possession of that knowledge economically valuable, and institutions, credentials, career ladders, and professional identities grew around it.
AI weakens some of that scarcity. Not completely, not evenly, and certainly not without serious questions about accuracy, provenance, ownership, consent, and access. Enough is changing, though, that defending knowledge scarcity as the foundation of expertise increasingly looks like defending friction. If a junior engineer can acquire in six months the conceptual vocabulary that once took someone five years to accumulate, stretching those six months back into five years does not protect expertise. It protects the old path by which expertise happened to be acquired.
The better response is to give people the knowledge faster and move scarce human effort somewhere more valuable. Incident management makes that distinction concrete. I can give someone every runbook I have written, every postmortem, every outage transcript, every architecture diagram, every incident recording, and every lesson extracted from failures my teams have survived. Put an AI system over that corpus and a new engineer can interrogate years of operational knowledge in seconds, exploring why decisions were made, alternative actions, failure modes, communications patterns, and dependencies they might otherwise have missed.
They may arrive at their first serious outage knowing more about incident management than I knew after years of piecing that knowledge together, and that is progress worth taking. The next lesson begins when production actually fails and their judgement becomes part of the causal chain. The executive channel fills. Revenue starts disappearing. Three teams are talking over one another. The obvious fix fails. Someone wants to roll back while another engineer warns that rollback could damage data. A senior person you trust suddenly sounds less certain than usual. Nobody has enough information, and delay itself has acquired a cost.
All the knowledge still matters. So does everything the AI helped retrieve. What changes is the position of the person using it. Ten minutes means something different when customers are affected. Silence from an engineer you trust carries a different signal after you have heard that silence in another incident. You begin to recognize when a team is expressing confidence because the evidence supports it and when the room is manufacturing certainty because uncertainty has become emotionally intolerable. You learn that further investigation can be prudent until the moment when continued investigation becomes another way of making a decision without admitting you made one.
Those lessons can be described afterward and should be. They can be simulated beforehand and that can improve preparation enormously. Neither substitutes for the moment when something real depends upon your judgement. The first serious incident changes the person who enters the second. The second changes the person who enters the third. Eventually someone notices things they could not have noticed during the first incident because the person capable of noticing them had not been formed yet.
Experience does not guarantee wisdom. Someone can repeat the same year twenty times and call it twenty years of experience. Contact with reality can produce superstition, stale assumptions, bitterness, bad habits, and unjustified confidence just as easily as it can produce better judgement. Sometimes the newcomer sees precisely what experienced people normalized years ago. Experience needs reflection if it is going to become useful, and evidence has to remain capable of correcting memory and intuition. Knowledge and experience do their best work when they can challenge one another.
Apprenticeship After AI
Apprenticeship becomes more important in an AI-rich world rather than less. If information can be retrieved in seconds, we should stop spending years pretending retrieval is formation. Let AI explain the protocol, teach the syntax, summarize the history, surface previous examples, and help construct simulations. Then use the time recovered to put people closer to consequential work while someone with deeper experience is still there to help contain the cost of failure.
Bring a junior engineer into the incident before they are expected to lead one. Let them watch decisions get made while the evidence is incomplete, then give them bounded responsibility and let the outcome matter enough to create a real feedback loop. Review what happened while the consequence is still attached to the decision that produced it. Increase the scope as their judgement develops. There is nothing noble about making someone spend three days hunting for information a machine can retrieve in thirty seconds. That is shitty search masquerading as professional development, not apprenticeship.
Apprenticeship is structured contact with reality. It moves someone from knowing about the work toward carrying the work, with enough support that failure teaches without becoming catastrophe. Hazing and inherited suffering have nothing to do with the objective. The point is to remove obsolete friction without removing contact with consequence.
That model has costs because meaningful experience is harder to scale than information. It requires supervision, real work, bounded risk, and senior people willing to share consequential decisions instead of hoarding them as proof of their own importance. It also creates another problem we should not ignore. If experience becomes the scarcer resource, access to experience becomes power.
Organizations can gate meaningful assignments just as effectively as they once gated knowledge. Some people will be given the incident, the launch, the negotiation, or the real architecture decision. Others will be left producing artifacts around work they are never allowed to carry. If we are serious about apprenticeship, we will have to watch that boundary carefully rather than replacing one prestige economy with another.
AI makes this more urgent because it can supply the language and artifacts of experience before the experience arrives. Someone can produce an architecture document before operating the architecture, draft a postmortem without really understanding the incident, or generate strategy without sitting beside the people who will bear its costs. The output can arrive before the person has been formed by the conditions the output implies.
Synthetic Competence becomes dangerous when organizations mistake polished output for grounded capability, and Ghost Apprenticeship follows when people acquire the artifacts of seniority without enough contact with the conditions that form senior judgement. The organization fails when it stops after teaching.
We can already see this in software. An engineer can use AI to produce code well beyond what they could have written independently, and that can be enormously useful. I use AI extensively in technical work myself. Something important is missing, though, if that engineer never understands the failure modes, operates the system, watches generated abstractions collide with production, decides whether the rollout should continue, or has to explain what happened after something customers depended on broke. What is missing is not purity. It is formation.
Taking the AI away would solve the wrong problem. The better move is to redesign learning around the fact that information is becoming easier to obtain. If a machine can retrieve the syntax, stop making memorized syntax the gate. If it can surface known failure modes, spend less time proving someone can recite them and more time seeing whether they recognize an unknown one. Ask whether they can reason when the instructions run out, recognize that the map is wrong, explain why they made the decision they made, identify what evidence changed their mind, and see consequences they could not see before they carried responsibility themselves.
Those questions are harder to counterfeit because they concern judgement formed in contact with reality. For centuries, experience often accumulated incidentally while people struggled to acquire knowledge. As knowledge becomes cheaper, we have an opportunity to reverse that relationship deliberately. Schools, companies, leadership pipelines, apprenticeship systems, and professional careers can spend less human time reproducing information and more of it creating responsible encounters with the world the information describes.
“What do you know?” will still matter, but it cannot carry as much weight by itself. We also need to know what someone has operated, built, carried, repaired, or watched fail. Where did reality disagree with their model? What happened that changed how they understood the problem? What could they only learn because something was actually at stake? Did they reflect on it well enough that the next person can begin with a better map without pretending that inheriting the map means inheriting the experience?
The Map Has No Wind
AI and human experience do not need to be competitors in this model. AI can give us extraordinary maps. It can compare them, explain them, preserve what previous travelers learned, find roads we missed, and warn us about conditions ahead. It may become the best mapmaker humanity has ever built.
Use it, but remember what the map cannot carry. It does not know the smell of the prairie before rain or what it is like to watch the Black Hills rise after hours of open country. It has never stepped into the Badlands and discovered that the representation it carried was too small. It has never looked at something so beautiful that its existing frame broke open and had to become larger. It has never been tired, needed water, feared losing something it loved, or carried the previous ten minutes into the next ten as a changed participant in the world.
We can give people better maps faster than we ever could before. That should not make us fight harder to preserve the old cost of knowledge. It should free us to spend more of our lives doing the thing no map, model, book, simulation, or inherited account can do on our behalf.
Then get them into the country.
Artifacts are cheap, judgement is scarce.
Per ignem, veritas.