AFTER THE DRAGON IS GONE

Humanity, AI, and the Architecture of the Threshold

AXYNAO

The threshold has disappeared. The question is what disappeared with it. — AXYNAO

I. The Guardian at the Boundary

Human beings have always imagined boundaries. Between the village and the wilderness. Between the living and the dead. Between ignorance and knowledge. Between ordinary existence and extraordinary power.

And remarkably often, people placed something at that boundary. A serpent. A sphinx. A ferryman. A trial.

In Greek myth, Ladon coils around the golden apples of the Hesperides; another serpent keeps the Golden Fleece at Colchis. In the Norse tradition, Fafnir becomes the monstrous keeper of a hoard that brings ruin rather than uncomplicated prosperity. The pattern is older and wider than the medieval European dragon. In Neo-Assyrian palaces, colossal human-headed winged bulls stood at the gateways to throne rooms. Babylon had the mushhushshu, the serpent-dragon of Marduk. Protective figures clustered at doorways, at gates, at the vulnerable seams of built space.

We should be careful not to make this tidier than it was. It would be convenient to claim that every civilization imagined the dragon as a monster hoarding forbidden treasure, and the evidence does not support it. The Chinese long was associated with water, rain, cosmic order, good fortune and legitimate imperial authority — forces that sustained human flourishing rather than obstructing access to it.

So the deeper pattern is not that dragons guard treasure. Across several traditions, human cultures assigned extraordinary beings responsibility for boundaries between ordinary human existence and forces understood as larger than ourselves. Sometimes the guardian prevented entry. Sometimes it protected the entrant. Sometimes it was the power itself.

What recurs is not the monster. It is the conviction that approaching certain kinds of power should not be casual.

A threshold is not simply a line between two spaces. It is a place where the conditions change. Cross it, and you are somewhere else. Sometimes you become someone else.

II. Talos

There is one guardian who matters more than the rest for our purposes, and he is not a dragon.

Talos was made. Hephaestus, the god of craft, forged him in bronze, and he was set to circle the coast of Crete. He was not hoarding a prize. He was patrolling a perimeter. A manufactured guardian, built to hold a boundary.

This is worth pausing on. We tend to speak as though the idea of an artificial guardian belongs to our century — that we are the first to build systems and station them at the edge of things. We are not. The oldest material already contains the thought. Antiquity did not only imagine guardians at thresholds. It imagined making one.

And then it imagined how he ended.

Talos was not defeated by a hero who proved worthy. In the Argonautica, Medea works on him and the single vein running through his body is opened, and the divine fluid drains out. He was found out. He had one vulnerability and it was exploited.

Jason took the Fleece. He crossed the threshold. And he acquired, in the crossing, precisely nothing — no judgment, no restraint, no transformation. The rest of his story is a long demonstration of the fact.

A threshold defeated by exploit teaches nothing to the one who crosses it.

This is the hinge of everything that follows, and it should immediately complicate any nostalgia we might be tempted toward. We are not comparing a wise ancient architecture with a careless modern one. The ancient architecture failed too, and it failed in the same way ours is failing: the guardian was bypassed rather than satisfied, and the crossing produced no formation in the one who made it.

What we are looking at is not a lost golden age. It is a recurring architectural problem.

III. The Threshold as Human Technology

We usually place myth in opposition to knowledge. Myth belongs to an earlier age; knowledge to the modern one. Myth explains what people did not yet understand, and science eventually replaces it.

There is truth in that history, but it describes only what mythology said. It says little about what mythology did.

Stories established boundaries between permissible and forbidden behaviour. They encoded obligations between generations, identified sacred places, preserved collective memory, communicated danger. They gave communities the means to interpret birth, adulthood, leadership, death, violence, knowledge and the unknown.

In that sense mythology can be understood as a form of cultural technology. Not made of silicon or steel. Made of narrative.

And one of its recurring products was distance. Something wanted; something in the way. Fire, immortality, treasure, a kingdom, passage into another world — rarely presented as lying on a table waiting to be collected. There is a journey, a prohibition, a sacrifice, a guardian.

Something separates wanting from having.

Modern architecture — commercial, computational, technological — is largely designed to remove exactly that. One click. One prompt. Instant publication, distribution, computation, access.

At its idealized mythological form, the sequence was:

Desire → Threshold → Learning → Trial → Transformation → Access

The architecture we are now building offers:

Desire → Interface → Access

The middle is disappearing. The question is what the middle was doing.

IV. What We Are Not Arguing

Before going further, two concessions. Neither is decorative. If the reader cannot state both of them back to us, this essay has failed.

The first is that the oldest version of our own argument was wrong.

In the Phaedrus, Socrates tells the story of Theuth, who brings the invention of writing to King Thamus and presents it as a remedy for memory. Thamus refuses the claim.

Writing, he says, will produce forgetfulness in the souls of those who learn it, because they will rely on external marks instead of remembering from within. Learners will receive a quantity of information without instruction, and will seem to know much while knowing nothing. They will have the appearance of wisdom rather than wisdom itself.

That is a threshold-collapse argument about an information technology, made roughly 2,400 years ago, in terms almost identical to those now used about artificial intelligence. And it is generally judged to have been mistaken. Writing did not hollow out human thought. It became the substrate of nearly everything we subsequently learned to do with it — including the preservation of Socrates' objection, which survives only because someone wrote it down.

Anyone making an argument of this shape in 2026 is obliged to notice that the argument has been made before, confidently, by a first-rate mind, and has not aged well.

The second concession is that threshold collapse is measurably good for the people the old thresholds excluded.

A randomized experiment with 1,174 adults, conducted outside firm settings, found that generative AI raised performance for all participants, with substantially larger gains for those with less education. The gains were not merely borrowed: participants did not perform worse than controls once the assistant was removed, and lower-education participants retained part of their advantage — although a sizable gap re-emerged.

This is not a small point offered in fairness. It is the strongest single argument against the position this essay is about to develop, and it deserves to be stated at full strength. The democratization of capability is one of the genuine goods of this technology, and the people it most helps are the people the old thresholds most reliably kept out.

We are not arguing that the thresholds should return. We are asking a narrower question, and it only becomes interesting once both concessions have been granted.

V. The Thresholds That Are Actually Collapsing

Some of them are collapsing measurably, and in ways the evidence lets us describe with precision.

In a study of a generative assistant deployed to customer-support agents, published in the Quarterly Journal of Economics, access to the tool raised productivity by about 15% overall, and by roughly 30% among less-skilled and less-experienced workers. The most striking figure is not the percentage. It is the tenure: agents with two months of experience, using the tool, performed about as well as agents with more than six months who were working without it.

Four months of formation, compressed into an interface.

That is a threshold collapsing, visible in operational data. And for the new agent — and for the customers she serves — it is very good news.

The second finding is less comfortable. In a pre-registered field experiment with 758 consultants at Boston Consulting Group, published in Organization Science, participants working inside the model's competence completed 12.2% more tasks, 25.1% faster, at more than 40% higher quality. Consultants below the median performance level gained 43%; those above gained 17%. Skill compression again, and again largely benign.

But the experiment also included a task deliberately positioned outside the model's competence. There, consultants using AI were 19 percentage points less likely to reach a correct solution than those working without it. The researchers describe a jagged frontier: capability that is excellent on one side of an invisible line and misleading on the other, with no signal marking where the line runs.

The detail that matters most for this essay is who these people were. They were highly specialized knowledge workers, drawn from a global consulting firm, working within their own domain of expertise. And the study reports that they tended to over-rely on the system precisely where closer supervision was required.

Hold that finding. We will need it shortly.

None of this is unprecedented in kind. The question long predates generative AI, and one of the clearer examinations of it concerns a capability most of us have already surrendered without discussion. Studying fifty regular drivers, Louisa Dahmani and Véronique Bohbot found that those with greater lifetime GPS experience had worse spatial memory when required to navigate without it — and, in a small three-year follow-up, that heavier GPS use was associated with a steeper decline. What makes the finding useful is a detail the authors were careful to establish: heavier GPS users were not people who already felt they had a poor sense of direction. The direction of the effect ran the other way.

The sample is small, the longitudinal group smaller still, and one study of navigation cannot be stretched to cover judgment in general. But it describes the shape of the thing exactly. A capability that was formed by the difficulty of doing something without help does not announce its own departure. Nothing goes wrong. You arrive.

What is new is the range of capabilities to which the question now applies, and the speed.

VI. Borrowed Expertise

There is a systemic version of this question, and it has been stated well by someone else.

Writing for Brookings in July 2026, Niam Yaraghi argued that the productivity gains now being recorded are produced in large part by people whose expertise was formed before these tools existed. They direct the systems skillfully because they spent careers building the judgment that knows what to ask, what a good answer looks like, and where a confident-sounding model is likely to be wrong. They are extracting value, in his formulation, because they paid a developmental cost that the technology now allows others to avoid.

His diagnosis is that this expert class is a stock rather than a flow — and that the flow which would replenish it is being slowed in ways the current data are not built to detect. He takes the policy question up in a companion piece.

We want to take the same observation somewhere different.

Yaraghi's concern is economic and institutional: the pipeline of experts. The concern that belongs to AXYNAO is architectural and older. It is not only expertise that was being produced by the developmental journey. It was judgment about when not to proceed. It was the felt sense of a domain's edges. It was the ability to recognize that an answer is wrong before being able to articulate why. These are not credentials. They are formations — and they were formed bythe difficulty, not merely certified after it.

There is a related cost at the collective level. In a study published in Science Advances, writers given access to AI-generated story ideas produced work rated more creative, better written and more enjoyable — especially the less creative writers. Yet the AI-assisted stories were measurably more similar to one another than the stories written without assistance. The authors describe it as resembling a social dilemma: individually better, collectively narrower.

Lift the average, compress the variance. For most purposes that is a straightforwardly good trade. But the tail is where the genuinely new arrives, and the tail is what compression costs.

VII. The Wrong Answer, and Why It Is Wrong

There are two available responses to all of this, and both are wrong.

The first is to rebuild the gate. If thresholds mattered, guard the knowledge. Let experts determine access. Let institutions certify worthiness. Let capability belong to those who have completed approved initiations.

No. Human history gives abundant reason to distrust that answer. Gatekeepers preserve knowledge and also monopolize it. Priesthoods preserve wisdom and also preserve privilege. Institutions certify competence and also defend themselves. Elites exercise stewardship and also call exclusion stewardship. And the honest reading of the historical record is harder still: for most people in most societies, the threshold produced no transformation whatsoever, because they never reached it. It filtered by birth, by sex, by wealth, by proximity to a city. The initiatory path we are tempted to admire was, statistically, a wall.

Those thresholds should stay down.

The second wrong answer is subtler, and it is the one this essay was originally going to give. If the gate cannot return, then let the guardian move inward. Let the external threshold become an internal discipline. Let each person cultivate the judgment necessary to hold extraordinary capability without being deformed by it.

This is attractive. It is also insufficient, and the evidence says so plainly.

Return to the consultants. Elite professionals, working inside their own field of expertise, with every reason to be careful, over-relied on the system exactly where care was most needed. Personal judgment, individually held and individually exercised, did not detect the jagged edge. Not because the people were careless — because the edge is invisible from inside the interface, and nothing in the environment marked it.

A capability that fails in a way its holder cannot perceive is not a problem that personal virtue solves. Asking individuals to supply, through character, what the surrounding architecture no longer supplies is the same error as asking them to be careful around an unguarded machine. It is the error of putting the entire load on the person standing closest to the risk.

AXYNAO's first principle runs the other way: conditions determine what becomes possible, and architecture precedes agency. If the threshold's function cannot be restored as a gate, and cannot be discharged by individual character, then there is only one place left for it.

It has to be built.

VIII. The Replenishment Test

Which brings us to the question this essay exists to ask.

When a technology removes the difficulty through which a human capability was once formed, what now replenishes that capability?

Notice what this question does not assume. It does not assume that removing difficulty is bad. Much difficulty was pure obstruction and its removal is one of the great achievements of the modern world. Nobody should mourn the barriers that stood between a capable person and a library.

The question is narrower and harder. It asks whether the vanished difficulty was producing something — and if it was, what produces it now.

Sometimes the honest answer is: nothing was being produced, and nothing needs replacing. Sometimes the answer is: something was being produced, and something else already produces it. And sometimes the answer is that something was being produced, nothing replaces it, and no one has noticed, because the output looks the same as it always did while the capacity behind it quietly stops being made.

Five questions follow. They are not a score, and they do not produce a rating. They are questions to sit with — about a practice, a profession, an institution, a discipline, a household.

Capability. What can we now do that previously required significant learning, experience or judgment?

Displacement. Which developmental process disappeared when that threshold collapsed?

Dependency. What knowledge or capability can we no longer reproduce without the system?

Stewardship. Which judgment must remain human, institutional or independently recoverable?

Architecture. What form of friction, verification, provenance, apprenticeship or human intervention should deliberately remain?

The last question is the one that carries the argument, and it is worth saying clearly that friction is only one of its possible answers. Deliberate difficulty is a real mechanism, and there is now serious design research on cognitive forcing functions and on interfaces built to slow a decision rather than accelerate it. But provenance is a different mechanism. So is apprenticeship. So is the requirement that a human being be able to reconstruct a result independently before it is relied upon. So is the deliberate preservation of a slower path alongside the fast one, kept alive not for efficiency but for formation.

Which mechanism fits depends entirely on what the vanished threshold was making.

IX. After the Dragon Is Gone

The answer is not to bring the dragon back.

It is to understand what work the dragon was doing.

Some thresholds protected privilege. They should remain dismantled. Others created time for learning, verification, responsibility and judgment to develop. Their disappearance creates a different problem.

The task before us is therefore neither gatekeeping nor frictionless access. It is to decide which functions of the threshold humanity must deliberately preserve — and to build architectures capable of preserving them without rebuilding the gates.

For thousands of years, in tradition after tradition, people imagined extraordinary beings standing close to extraordinary power. Perhaps our ancestors were not simply frightened of monsters. Perhaps they understood something about us: that possession changes the possessor, that knowledge creates obligation, and that some crossings change what is possible afterward.

They also understood, in the story of the bronze guardian on the Cretan shore, that a guardian can be found out. Talos did not test Jason. He was drained, and the way lay open, and the man who walked through was exactly the man who had arrived.

The dragon may be gone.

The function of the threshold remains.

Stoic Reflection

For even sheep do not vomit up their grass and show to the shepherds how much they have eaten; but when they have internally digested the pasture, they produce externally wool and milk. — Epictetus, Enchiridion 46 (trans. George Long)

Epictetus wrote this against ostentation — against the student who parades doctrine he has not lived. But the image reaches further than his warning did.

Access is not digestion. What we can retrieve is not yet what we have become.

Sources

Myth and the ancient guardian

Apollonius of Rhodes, Argonautica, Book IV — the death of Talos: the single vein, the draining of the ichor, and the fact that the guardian was undone by vulnerability rather than by trial. The basis for this essay's central claim about crossings that teach nothing.

Adrienne Mayor, Gods and Robots: Myths, Machines, and Ancient Dreams of Technology (Princeton University Press, 2018) — the ancient imagination of manufactured and self-moving beings, including Talos as a made guardian. Supports the argument that artificial guardianship is an old thought, not a new one.

The oldest objection

Plato, Phaedrus, 274c–275b — Thamus's refusal of Theuth's gift: writing will produce forgetfulness, and the appearance of wisdom in place of wisdom. Cited here against our own thesis, as the earliest and most distinguished version of the argument this essay makes.

What the evidence shows

Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond, "Generative AI at Work," Quarterly Journal of Economics140(2), 2025 — the ~15% overall and ~30% novice productivity gains, and the finding that two-month agents with the tool matched six-month agents without it. Supports the claim that formation time is being compressed, measurably.

Fabrizio Dell'Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Saran Rajendran, Lisa Krayer, François Candelon and Karim Lakhani, "Navigating the Jagged Technological Frontier," Organization Science, 2025 — the 758-consultant field experiment: gains inside the frontier, a 19-percentage-point accuracy loss outside it, and over-reliance among domain experts. Supports both the skill-compression claim and the essay's refusal of individual judgment as a sufficient answer.

"Does Generative AI Narrow Education-Based Productivity Gaps? Evidence from a Randomized Experiment," NBER Working Paper 34851, 2026 — the 1,174-participant experiment showing larger gains for lower-education participants and partial retention after the tool was withdrawn. The essay's second concession rests on this.

Anil R. Doshi and Oliver P. Hauser, "Generative AI enhances individual creativity but reduces the collective diversity of novel content," Science Advances 10(28), 12 July 2024 — AI-assisted stories rated more creative individually while resembling one another more closely in aggregate. Supports the argument about compressed variance.

Niam Yaraghi, "Borrowed expertise: Why AI's productivity boom may not survive the generation that built it," Brookings Institution, 10 July 2026, and "Repaying the inheritance: How education and research policy can address AI's borrowed expertise," 20 July 2026 — the stock-versus-flow diagnosis of expertise formed under earlier developmental conditions. The starting point for this essay's central question, taken here in an architectural rather than a policy direction.

Louisa Dahmani and Véronique D. Bohbot, "Habitual use of GPS negatively impacts spatial memory during self-guided navigation," Scientific Reports 10:6310, 14 April 2020 — fifty drivers assessed for lifetime GPS experience and spatial memory, with a three-year follow-up of thirteen participants, and the finding that heavier users were not those who already reported a poor sense of direction. Supports the argument that a capability formed by difficulty can decline without any visible failure.

The threshold as a question about formation

Epictetus, Enchiridion 46, translated by George Long — the sheep who do not display their fodder but produce wool and milk. The source of this essay's closing distinction between access and digestion, extended here beyond Epictetus's own warning against ostentation.

Friction as design

Chiara Natali, Mohammad Naiseh, Federico Cabitza and Brett Frischmann, "Better AI with Designed Friction: Theories, Applications and Research Agenda," Proceedings of the 4th International Conference on Hybrid Human-Artificial Intelligence (HHAI 2025), 2025 — develops designed friction in human-AI interaction through mechanisms including cognitive forcing functions, seamful design and programmed inefficiencies. Supports friction as one possible architectural response in the Replenishment Test, rather than as the essay's proposed universal solution.

AXYNAO Reflection

Knowledge is not only information that survives. It is a relationship between what is known, who holds it, what obligations it creates, and what had to happen to a person before they could carry it.

An architecture that preserves the first while quietly dissolving the last has not preserved knowledge. It has preserved the record of knowledge, and left the question of who is capable of receiving it to be settled somewhere else, by no one in particular.

AXYNAO Erik Kling, "Diplom Oekonom" University Of Hohenheim

• Speaking & Advisory AXYNAO: "Architecture Determines Meaning"