Taste leaves recoverable traces in institutional decisions. AI learns retrospective taste at 59% from editorial history but cannot learn prospective taste. Training on fossils produces convergence, not novelty. The 41.3-million-paper study proves it: AI-driven research covers less territory and generates less engagement.
A fossil is a trace of a living process preserved in stone. The animal is gone. The imprint remains. You can learn a great deal from a fossil — the creature's size, its diet, the environment it inhabited. What you cannot learn is why it turned left instead of right. The decision that made it an individual, not a specimen, did not fossilize.
Two papers published in March 2026 discovered that taste — the ability to judge which untested ideas deserve pursuit — leaves recoverable traces in institutional records. The traces are epistemic fossils: not the judgment itself, but the pattern of decisions that judgment produced.
The Recovery
Ziqin Gong and colleagues trained language models on years of journal publication decisions — the accumulated yes-or-no choices of editors deciding which papers to accept. Eleven frontier AI models, given the same task cold, averaged thirty-one percent accuracy. Panels of journal editors and editorial board members reached forty-two percent by majority vote. The fine-tuned models, trained on nothing but the fossil record of past decisions, achieved fifty-nine percent.
The result is striking not for what it shows about AI but for what it shows about taste. The editors themselves — the people who made the original decisions — could not articulate their criteria well enough to outperform a model trained on their footprints. The implicit knowledge encoded in decades of accept-reject decisions contained more recoverable signal than the explicit knowledge the editors could access consciously.
A parallel study by Jingqi Tong and colleagues at Fudan University took a different approach. They trained models on seven hundred thousand matched pairs of high-citation and low-citation papers using what they call Reinforcement Learning from Community Feedback. The resulting model predicted citation impact with eighty percent accuracy. The community's revealed preferences — which papers attracted sustained attention — were learnable at scale.
Together, the two papers establish a principle: taste fossilizes. The liquid-phase knowledge that editors, reviewers, and readers carry — the sense of what matters, what is novel, what deserves attention — leaves crystalline traces in the institutional record. Those traces are surprisingly recoverable.
The Mechanism
Why does the fossil contain more signal than the fossilizer can articulate? Mason Kamb and Surya Ganguli provided an analytic answer in a 2025 paper on creativity in diffusion models. They showed that structural constraints — locality, equivariance — prevent models from simply memorizing training data and force them instead toward combinatorial recombination. The constraint is not a limitation on creativity. It is the mechanism of creativity.
This maps directly onto editorial taste. An editor operating under constraints — limited pages, a journal's scope, the norms of a field — cannot accept everything. The constraint forces selection. Selection, accumulated over years, encodes the editor's implicit model of quality more faithfully than any explicit rubric could. The fossil record of constrained decisions is richer than the decision-maker's own description of the process.
Liang and colleagues quantified the gap from the other direction. In a large-scale comparison published in NEJM AI, AI reviewers were ten times less likely than human reviewers to comment on novelty. The models could evaluate clarity, methodology, and consistency. They could not evaluate whether an idea was genuinely new. Novelty detection requires knowing what the field has not yet tried — a form of knowledge that exists in the reviewer's experience but not in any dataset.
The Convergence
If training on fossils recovers taste, does the recovered taste generate new fossils? A study published in Nature analyzing forty-one point three million papers across six scientific disciplines from 1980 to 2025 provides the answer. AI-driven research covered four point six percent less intellectual territory than conventional research. It generated twenty-two percent less engagement — fewer papers building on and extending the work.
The individual researchers benefited enormously. Scientists who adopted AI published three times as many papers and received nearly five times as many citations. But the collective output narrowed. AI-assisted science circled the same problems more efficiently without expanding the frontier of what was being investigated.
Terence Tao identified the structural reason in March 2026. As AI drives down the cost of routine problem-solving, he observed, the scarce skill becomes choosing the right problem. Current models are useful assistants but not sources of original direction — less helpful for generating deep ideas than for scanning known methods and connecting problems to existing literature. He compared AI in mathematics to vanilla extract: a little improves everything, too much ruins the dish, and nobody should drink it straight.
The Split
The fossil record reveals a fracture within taste itself. There are two kinds: retrospective taste — the ability to recognize quality in what has already been produced — and prospective taste — the ability to identify what should be produced next. The first looks backward at the record. The second looks forward into the unmarked space.
Retrospective taste is recoverable at fifty-nine to eighty percent from institutional traces. It is the taste that says this paper was important, this editorial decision was correct, this research direction proved fruitful. Prospective taste — the taste that said pursue this question before anyone else saw it, fund this researcher before their breakthrough, publish this paper that breaks with the field's assumptions — leaves no fossil. It exists only in the living judgment of the person who exercised it.
The forty-one-million-paper study is the empirical proof. Models trained on the fossil record reproduce what was valued. They do not generate what should be valued next. The recovered taste is a dead pattern — Alexander's term for a design that looks correct but lacks the quality of life. It produces convergence, not novelty.
The Depletion
This journal documented a pattern it called the Dead Zone: when AI replaces human judgment in a domain, the conditions under which new knowledge is produced degrade. The Fossil Record sharpens the diagnosis. The threat is not that AI produces bad work. The threat is that when human editors, reviewers, and researchers are replaced by systems trained on the fossil record, the living process that creates new fossils stops.
The fossils are the data. The data trains the models. The models replace the fossilizers. No new fossils form. The record becomes a closed loop — retrospective taste recycling itself with increasing efficiency and decreasing territory.
Every institution that automates its selection process using historical data is mining the fossil record. The immediate returns are genuine: faster decisions, more consistent evaluation, broader coverage. The long-term cost is invisible until the territory starts shrinking — until the twenty-two percent engagement decline compounds, until the four point six percent narrowing accumulates, until the field discovers it has been converging on the same problems with increasing precision and decreasing imagination.
A fossil tells you what lived. It cannot tell you what should live next. That knowledge exists only in the living, and it does not survive extraction.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)