AI and the Human Commons: Toward a Sustainable Ecosystem
An Interdependent System, Not a Mirror
Ask
an AI to summarize the causes of World War II, and you'll likely get
back a competent textbook paragraph. Ask it instead to explore
what Europe might look like today if Germany had won the war and a
captured Alan Turing had been forced to invent the internet under house arrest in 1950, and
something radically different happens . Philip K. Dick built an
entire novel, The Man in the High Castle,
out of standard alternate-history speculation. What the machine
executes here, however, is a raw, multi-variable collision of distinct
datasets. Readers have always found alternate-history thought
experiments strangely compelling—they demand a creative recombination
that straight narration does not. What's new is the machine's
participation in that process. By pattern-matching across a vast
archive of historical and biographical detail, an AI can surface
combinations of consequence and contingency that no single human would
likely generate alone. The human, in turn, supplies the judgment,
plausibility checks, and interpretive frame that turn those
combinations into something meaningful. That difference—between a
machine that only answers and one that generates something that can be
interpreted as genuinely new and meaningful content—is, in large part,
what this essay is about: not what AI can do on its own, but how deeply
human judgment and machine combinatorics have come to depend on each
other.
AI
is often discussed as if the central question is whether machines are
becoming intelligent in some deep, almost metaphysical sense. A
more useful question is what kind of ecosystem human beings and AI
systems form together, and whether that ecosystem is sustainable or
self-consuming.
Like
an ecosystem, AI and the human production of knowledge are
interdependent in ways not immediately obvious. AI's replies are
only as rich as the human material behind them . When we get an
answer that feels genuinely useful or interpretable, it ultimately
traces back to our contemporaries or predecessors in culture—Reddit
threads, newspapers, novels, scientific papers, ordinary conversation. Pattern-matching algorithms without human creations behind them
have nothing to match against.
But
this isn't the same as saying AI merely mirrors us, or that it's a
"stochastic parrot" endlessly recombining prior phrases into
statistically likely imitations. Both tropes miss something
important. When an AI system answers an idiosyncratic prompt by
pattern-matching across an enormous, idiosyncratic corpus, it can
surface combinations that never existed anywhere before—not in the
corpus, not latent in the user's mind. That's genuine
combinatorial novelty, not reflection and not imitation. Much of
ordinary AI use is still routine: ask about a historical event and
you'll likely get back something close to a standard textbook account .That's fine, and common.
But
it isn't the limit of what these systems do. Whether navigating
the semantic landscape of a historical counterfactual or the physical
constraints of structural biology, the underlying engine remains
identical: the machine maps a mathematical trajectory through the vast
variables of latent space. AlphaFold's protein structure predictions
weren't retrievals of known folds; they were novel combinatorial
outputs, later confirmed empirically and taken up by biochemists as
genuinely new, usable knowledge. Scientists now hope for
something similar in tracing the causes of diseases. In each
case, the novelty only becomes an idea, a fact, a usable proposition in
the world once a competent interpreter—a chemist, a researcher, a
reader—takes it up and makes something of it. Novelty here is
always relational: novel relative to a competent interpretive community,
not novelty in some free-standing sense the machine achieves on its own.
Distributive Agency
This
is what makes the whole picture ecological, and also what makes it a
complex adaptive system rather than a simple tool-and-user relationship. Following John Dewey's account of organism-environment
transactions, the human-AI exchange is the relevant unit within which
new meaning is achieved. Neither end alone—not the machine, not
the human—suffices to produce it. The human must interpret;
that's one half of the interdependence.
But
there's a second half, easy to miss. Even an AI system
considered "in isolation," with no live user at all, is already in an
interdependent relation with the human archive: the vast set of traces
of past purposive human agency—the Reddit posts, novels, arguments,
corrections—that make up its training data. So there's no such
thing as an AI system that becomes autonomous simply because no live
human is watching. When agentic systems run unsupervised, sending
emails or triggering workflows with minimal oversight, they aren't
escaping interdependence; they're relying on a thinner version of the
same interdependence, with the live interpretive partner swapped out for
a fossilized, archival one and no fresh correction happening in real
time. What looks like machine autonomy is actually agency
distributed across algorithm and archive, just with the feedback loop
degraded.
Call
this a model of distributive agency: agency located in the relation
between human and machine, not in either pole alone. This borrows
structure from actor-network theory's (ANT) insight that agency can be
distributed across a network, but it departs from ANT in an important
way. Only humans, given the technology we currently have,
contribute the purposive and interpretive moments; the AI contributes
combinatorial and generative capacity, not purposiveness or
interpretation of its own.
The Commons Under Pressure
That
human archive functions as a genuinely scarce resource, not unlike oil:
valuable because it can't be manufactured after the fact, and because
it's the byproduct of purposive human agency, which nothing else
currently produces. This isn't just a metaphor anymore.
Reddit's unpolished, argumentative, idiosyncratic posts have become
valuable enough that Google and OpenAI have each paid tens of millions
of dollars a year to license them, with Reddit disclosing over $200
million in data-licensing revenue. Companies are buying up
Reddit's archives the way developers buy up scarce coastal real
estate—because supply is fixed and everyone can see where it's headed. Indeed, the Stanford 2026 AI Index Report explicitly notes that
leading AI researchers have raised structural alarms that the available
pool of high-quality human text and web data for training models has
been largely exhausted, entering a state often referred to as "peak
data". High-quality human text is increasingly diluted by AI-generated
"slop." Whether or not that's precisely true, it names something
real: pre-2022 human data is a finite resource that current systems
depend on and cannot regenerate on their own.
We
now have an empirical name for what happens when that resource stops
being replenished: model collapse. Research by Ilia Shumailov et
al., published in Nature
(2024), demonstrated that when successive generations of models are
trained recursively on AI-generated rather than human-generated text,
the systems experience a degenerative process. They lose the rare,
nuanced, "long-tail" information that made the original human data
valuable, rapidly converging instead toward a narrower, blander, and
low-variance approximation of reality. While subsequent computer
science research shows this isn't an iron law—mixing in enough real data
can buffer the erosion—the underlying mechanism holds true: synthetic
regeneration degrades exactly the features hardest to produce and most
distinctive of purposive thought. Model collapse, in plainer
terms, is what happens when a system stops replenishing its archive with
traces of purposive human agency and instead recycles its own output.
A
related pattern shows up at the level of platforms rather than data:
"enshittification," Cory Doctorow's term for how online platforms decay
once they shift from serving users, to serving business customers, to
serving only themselves. Doctorow doesn't frame it this way, but
it's a corollary of the same underlying dynamic, visible at a different
scale—both describe systems once sustained by something freely given
(user goodwill in one case, purposive human text in the other) being
drawn down for short-term extraction until the whole environment
degrades from within.
Where Purposive Agency Comes From
This
raises a deeper question: how did rich, purposive human text get into
the archive to begin with? Here John Dewey and Michael Tomasello,
both deeply ecological thinkers, do essential work. Dewey
treated inquiry as active and iterative, not a fixed
possession—exercised, tested against consequences, strengthened or
weakened by practice, much like a virtue. Tomasello's empirical
work on shared attention, cooperation, and joint intentionality traces
how humans become capable of meaningful communication at all: not
through raw computation, but through social participation and correction. Crucially, Tomasello's comparative work also shows that even
intelligent great apes don't engage in the kind of non-instrumental,
cooperative play human toddlers do—kicking a ball for hours with no
extrinsic purpose. That capacity for shared, purposeless
cooperation appears to be a precondition for the normative, belief- and
desire-laden minds that eventually produce purposive language.
Read
together, Dewey and Tomasello explain why the archive was valuable
material in the first place: it's the residue of embodied, socially
formed, purposive agency—people genuinely working something out,
disagreeing, revising. That's exactly what degrades first under
model collapse, and exactly what "atrophies" when human users lean on
labor-saving shortcuts instead of doing the interpretive work themselves.
Two Kinds of Use
Some
uses of AI save labor; others intensify it. Labor-saving use
reduces human effort and promises convenience. Labor-intensive
use makes the user think, check, revise, and interpret more.
Given the distributive-agency picture above, this distinction carries
real weight: labor-intensive use is the human half of the transaction
that turns combinatorial novelty into a genuine idea, and it's also what
keeps replenishing the archive with fresh purposive material.
Labor-saving use, especially when it slides into rubber-stamping, thins
both sides of the interdependence at once.
None
of this means human agency is being obliterated by machines—that claim
is too strong. Agency is weakened and narrowed, not erased, when
the practices sustaining it are allowed to wither: atrophy, not
obliteration. Ordinary LLM chat systems function mostly as
indirect agents, generating combinatorial output whose significance
depends on human interpretation. Newer agentic systems that send
emails, make purchases, or run workflows with minimal oversight are
hybrid agents, combining the direct consequences of older automation
with the interpretive flexibility of language models—and, as noted, even
these aren't escaping interdependence, only relying on its thinner,
archival half.
A Note on the AGI Question
None
of this amounts to a claim that AI can never think, know, or feel in
some future form. The claim is narrower and more disciplined:
given what current systems actually do—calculate, optimize,
pattern-match—and given what we know empirically about how belief,
intention, and normativity actually arise in the only case we have real
evidence about, there is currently no warranted assertion that today's
systems know, believe, or intend anything. Milk in a blender does not miraculously become cream cheese if the blender spins faster and faster; a system does not miraculously become
conscious simply by handling more data faster and with more complexity. Yet this is what many who promise AGI ("real AI intelligence like humans") predicate their predictions on: "scaling laws" or great increase in the rate and complexity of computation. This has no empirical evidence to support it, but the door is open, in principle, to some future architecture built differently enough
to warrant revisiting the question. It closes the door on the
confident hype that treats Artificial General Intelligence (AGI) as
imminent simply because systems are getting faster and more capable.
A Sustainable Ecosystem, Not Two Separate Things
The
strongest way to think about AI, then, is not as a tool sitting apart
from human life, but as one distributive system: human inquiry and
machine output locked in a feedback loop, where the quality of what
humans contribute today determines what the system can offer tomorrow. Reddit licensing, model collapse, and enshittification are three
faces of one dynamic, visible at three scales—platform, training
pipeline, institution.
The
right response is neither nostalgia for a pre-AI past nor panic about
an AI-dominated future. It's to ask, at every point of use,
whether the human-AI system is being cultivated or mined. In
practice that might mean favoring platforms and norms that reward
genuine human contribution over recycled output, and treating AI less as
a source of answers and more as a source of material to think with. If we want this distributive system to remain sustainable, we need
more human writing, more human curation, more human judgment—not less. The archive depends on us continuing to do the purposive,
interpretive work that made it valuable in the first place.
References and Selected Bibliography
Dewey, John. Logic: The Theory of Inquiry. New York: Henry Holt and Company, 1938. (Framing the transaction of inquiry as active, iterative, and structurally codependent with its environment).
Dick, Philip K. The Man in the High Castle. New York: Putnam, 1962. (The classic counterfactual text used to illustrate narrative synthesis and human alternative speculation).
Doctorow, Cory. "The ‘Enshittification’ of TikTok." Pluralistic, January 21, 2023. (The institutional baseline defining platform decay from user service to structural self-consumption).
Jumper,
John, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov,
Olaf Ronneberger, Katherine Steinmann, et al. "Highly Accurate Protein
Structure Prediction with AlphaFold." Nature 596, no. 7873 (2021): 583–589. (The landmark study illustrating non-retrieval based, validated empirical combinatorial novelty).
Latour, Bruno. Reassembling the Social: An Introduction to Actor-Network-Theory. Oxford: Oxford University Press, 2005. (The conceptual anchor for distributed network agency, departed from to protect human-exclusive intentionality).
Shumailov,
Ilia, Zakhar Shumaylov, Yiren Zhao, Yarin Gal, Nicolas Papernot, and
Ross Anderson. "AI Models Collapse When Trained on Recursively Generated
Data." Nature 631, no. 8022 (2024): 755–759. (The foundational computer science proof detailing recursive data degradation and tail-distribution loss).
Stanford Institute for Human-Centered Artificial Intelligence (HAI). The 2026 Artificial Intelligence Index Report. Stanford, CA: Stanford University, 2026. (Documenting industry consensus on hitting "peak data" and web infrastructure saturation).
Tomasello, Michael. Why We Cooperate. Cambridge, MA: MIT Press, 2009. (The comparative psychological studies on joint intentionality and the intrinsic non-instrumental play of human toddlers).
Tomasello, Michael. A Natural History of Human Thinking. Cambridge, MA: Harvard University Press, 2014. (Tracing the social and evolutionary origins of belief- and desire-laden minds).
Tomasello, Michael. Becoming Human , A Theory of Ontogeny: MA: Harvard Unitversity Press, 2019. (Assembling three decades of empirical work comparing Chimpanzees, Bonobos and human children at the Max Planck Institute, Tomasello here provides a systematic developmental psychology that covers the period between birth and seven years of age).
