The Second Renaissance

In July 2007, a pair of astronomers at Oxford put a simple question on a website: is this galaxy a spiral or an ellipse? Visitors clicked an answer and saw another image. Within a year, more than a hundred thousand volunteers had looked at nearly a million galaxies and generated over forty million individual classifications. When the results were compared against galaxies that professional astronomers had classified by hand, the crowd’s judgments agreed (Lintott et al., MNRAS, 2008).

That project, Galaxy Zoo, is the cleanest existing piece of evidence for the question this article takes up. It shows that a large number of non-specialists, given a well-designed tool and a narrow task, can produce data that professionals accept. A movement from print to cheap instruments to digital networks has repeatedly widened who can participate in serious intellectual work. The question is whether the current wave of cheap, automated knowledge tools is another widening of that kind — a second Renaissance of amateur science, art, and scholarship — or whether something different is happening, because this time the tool can generate the work as well as assist it.

Answering that requires going back to the first time a technology of knowledge abundance actually reshaped who could think and discover. The historical record is more specific than the word “Renaissance” suggests, and it contains a warning that the optimistic reading usually omits.

What print actually changed

Elizabeth Eisenstein’s The Printing Press as an Agent of Change, published in two volumes in 1979, set out to answer a deceptively narrow question: what did the printing press do that manuscript copying did not already do? Her answer was not that print spread new ideas. It was that print changed the conditions under which knowledge could be stored, corrected, and accumulated (Eisenstein, The Printing Press as an Agent of Change, Cambridge University Press, 1979).

Three of her observations matter here. First, print made texts stable. A manuscript copy drifts as it is copied; a printed edition is identical across hundreds of copies, so an error introduced in one place can be found and corrected everywhere. Second, print made texts cumulative. An author could build on a fixed edition instead of a moving target, which made the slow, additive work of the sciences possible in a way that manuscript culture had not been. Third, print created a public record that a dispersed community could argue over at the same time. The scientific journal, the priority dispute, and the coordinated research program are all downstream of cheap, standardized reproduction.

Eisenstein’s own account of the consequences was candid about cost. Print did not produce only enlightenment. It produced the Reformation, which fractured Christian Europe, and it enabled the wide circulation of polemic, propaganda, and competing cosmologies that fueled a century of religious war. The same technology that made cumulative science possible made sectarian conflict cheaper too. Anyone arguing that abundance of knowledge tools is straightforwardly good has to account for the fact that the last great wave of it cut both ways.

That ambiguity is not a reason to discount the analogy. It is the analogy’s most useful feature. Print multiplied the number of people who could produce serious work and the number who could produce confident nonsense, and the two effects arrived in the same package.

When amateurs built the science

The tendency to imagine a past in which science was done by credentialed professionals and amateurs were dabblers gets the history backwards. In much of the nineteenth century, the “amateur” was not the person excluded from science. The amateur was frequently the person doing it.

A British Association for the Advancement of Science was founded in 1831, and its founders are often described as professionalizers. The historian Heather Ellis has argued that the term “professionalization” flattens a more complicated reality: alongside a push for specialist expertise, the association and its peers valorized the “gentleman-amateur” ideal well into the twentieth century, and much of the actual observational work of Victorian natural history was done by people with other jobs — clergymen, doctors, schoolmasters, and women barred from formal scientific employment (Ellis, “Knowledge, character and professionalisation in nineteenth-century British science,” History of Education, 2014).

The same pattern held in America. The word “scientist” was coined in 1834, and Paul Lucier’s study of nineteenth-century American men of science found that the category barely existed as a profession for most of the century; those who did the work were often self-taught and uncredentialed, and the effort to establish a pure, university-based, salaried scientific profession was a late-century reform movement rather than a description of the existing world (Lucier, “The Professional and the Scientist in Nineteenth-Century America,” Isis, 2009).

The professionalization that followed genuinely did contract the amateur’s role, particularly in sciences that came to depend on expensive laboratory apparatus. But it did not eliminate amateurs; it redefined what counted as a contribution. The observational sciences — ornithology, botany, astronomy, meteorology — kept a place for non-professionals because the work was distributed in space and time in ways a staffed institution could not cover. This is the pattern that Galaxy Zoo and its descendants would later inherit and accelerate.

Citizen science as a working model

What is striking about the modern citizen-science record is how much of it has already been done, and how much of it contradicts the intuition that non-specialists cannot contribute to real research.

Galaxy Zoo is the founding example. The design choice that made it work was the decomposition of a hard problem — classifying galaxy morphology across a huge survey — into a micro-task a person could complete in seconds. Because each galaxy was classified many times independently, the aggregation could quantify the error that any single classifier introduced, and the resulting catalog was robust enough that it became a standard resource in the field. Nearly 900,000 galaxies were eventually released as a public dataset, and later expansions asked volunteers for more detailed measurements of galactic structure.

Foldit took the approach into molecular biology. Launched by researchers at the University of Washington in 2008, it turned protein-structure prediction into a puzzle game and gave players the same physics-based scoring the laboratory’s own software used. In a 2010 paper in Nature, the team reported that more than 57,000 players had taken part and that top-ranked players had outperformed the lab’s automated method on a set of blind structure-refinement problems where the solution required substantial reworking of a protein’s backbone (Cooper et al., “Predicting protein structures with a multiplayer online game,” Nature, 2010). The detail worth noticing is why the humans won. The automated method was more consistent, but it got stuck probing a local optimum; the players had the patience to dismantle a structure they had already partly solved, accepting a worse score in the short term because they could see a better one further away.

The pattern across both projects is consistent. Non-specialists contribute most reliably when the task has been decomposed into evaluable units, when the tool does the bookkeeping and the scoring, and when the contributor brings something the automated system lacks — visual intuition in one case, strategic persistence in the other. That is a real division of labor and not a flattering way of saying the public is being kept busy. In both cases the professional literature treated the volunteers’ output as research data.

Where the machines change the arithmetic

The citizen-science model depended on a human supplying the thing machines could not: judgment, perception, or dogged search through a space that automated methods handled poorly. Automation has since moved onto that ground.

In 2024, a DeepMind system called AlphaGeometry solved 25 of 30 olympiad-level geometry problems on a standard test set, a performance that approached the average outcome of an International Mathematical Olympiad gold medalist (Trinh et al., Nature, 2024). A separate system, GNoME, screened millions of candidate crystal structures and proposed hundreds of thousands of chemically stable possibilities for synthesis (Merchant et al., Nature, 2023).

Two things follow, and they pull in opposite directions.

The optimistic reading is that these systems hand the hard part of exploratory work to anyone. If proving an olympiad problem or screening a compound library once required a trained specialist, and now requires a well-phrased question, then the number of people who can launch a real investigation grows enormously. The same observation applies to scholarship and art: the technical floor for producing a serious essay, a data analysis, a musical arrangement, or a translation has dropped, and it has dropped most for people who never had access to the specialized training.

The sober reading is that automating the contribution threatens the citizen-science model specifically. Galaxy Zoo worked because the human click carried information the machine did not have. If a model can classify the galaxies, the volunteers are no longer supplying a unique input, and the relationship shifts from collaboration to supervision. The blooming and the obsoleting arrive on the same curve.

The noise that abundance brings

The history of print warns against reading abundance as progress. The relevant modern evidence says the same.

Cheetham and Seshadri, two working materials chemists, examined a sample of the GNoME database and concluded that its entries demonstrated “scant evidence for compounds that fulfill the trifecta of novelty, credibility, and utility” — that many were trivial variations on known compositions, that some were chemically implausible, and that the appropriate term for a predicted compound with no demonstrated function was not “material” (Cheetham & Seshadri, Chemistry of Materials, 2024). The exchange is instructive not because the system was worthless — the critics acknowledged the method held promise — but because the volume of output outran the field’s ability to evaluate it. When two million candidates arrive at once, the bottleneck moves from generation to triage, and triage is work that still needs people who know the domain.

The same dynamic appears in creative and scholarly production, where the triage problem is not a shortage of reviewers but a shortage of shared standards for what counts as a contribution. A person who assembles a rigorous local history from primary documents and a person who prompts a model into a fluent imitation of one produce artifacts that look alike at a glance and differ entirely in what they cost to verify. The signals that used to carry that information cheaply, a known press, a known author, a slow publication process, were themselves bureaucratic overhead that nobody is eager to restore.

Abundance of tools and abundance of quality are different quantities. Print supplied both, and the institutions that made print productive — the journal, the peer review, the encyclopedia, the public library — were built to manage the second. A second Renaissance, if it comes, will be less a consequence of better tools than of better filters.

What actually made the first one work

It is worth separating the parts of the first Renaissance and the scientific revolution that depended on tools from the parts that depended on institutions.

Print was necessary. Eisenstein’s case is that the cumulative, self-correcting character of early modern science depended on fixed, widely available texts. But print alone did not produce the results. What turned cheap reproduction into durable knowledge was a set of social arrangements that emerged alongside it: the correspondence network of natural philosophers, the founding of societies that published proceedings, the convention of citing prior work, and the norm that a claim had to survive the scrutiny of distant strangers. The Royal Society’s motto, tellingly, was nullius in verba — take no one’s word for it. The tool made the culture possible; the culture did the work.

The citizen-science successes fit this pattern. Galaxy Zoo and Foldit worked because a professional group took responsibility for verification, designed the task to be checkable, and published the results in the ordinary way. The volunteers supplied labor and insight; the institution absorbed the responsibility for whether the output was true.

That division of responsibility is the part of the current moment that is genuinely unsettled. If automation makes it easy for a person to produce work that looks like scholarship, art, or science, and if the verifying institutions are already strained, then the natural failure mode is not a scarcity of production but a collapse of the trust that makes production useful. A generative tool can help an amateur reach a real result. The same tool can help a thousand others reach confident, fluent, wrong results, and the cost of telling them apart falls on the same overstretched reviewers.

Three conditions for a genuine flowering

The question of whether cheap tools produce a new Renaissance can be given a disciplined answer if it is broken into conditions rather than predicted wholesale.

The first condition is a task structure that rewards distributed effort. The historical and modern amateur successes share a shape: work that is decomposable, that can be scored or checked by an independent standard, and that benefits from more contributors. Where those conditions hold, the widening of participation is real. Where the work is not decomposable into checkable pieces, adding participants adds variance rather than knowledge.

The second condition is verification capacity that scales with production. Print’s own history suggests this is the binding constraint. A journal that receives ten times as many submissions needs either a much better filter or a much larger reviewing community; absent both, the signal in the noise is what degrades. The plausible mitigation is not to slow the tools down but to build better triage — automated checks, provenance standards, and a healthy willingness to publish negative or null results. Whether the relevant institutions will adapt is a matter of choices being made now, not of technological inevitability.

The third condition is that participation leaves the contributor more capable. This is the condition the labor studies cannot yet settle, and it is the one that decides whether the new abundance compounds. A person who uses an AI assistant to complete a piece of research while understanding less of it each time is participating in a transfer of capability outward, away from themselves. A person who uses it to attempt work at the edge of their competence, with enough feedback to learn from the failures, is doing something closer to an apprenticeship. The tools are compatible with both; the difference is in how they are used and, crucially, in whether the design of the system exposes its reasoning or hides it.

The honest version of the analogy

The Renaissance analogy is popular because it promises that abundance of knowledge tools leads to a flowering of human achievement. The historical record supports a narrower and more useful version of that claim. The availability of a transformative knowledge technology — print, then cheap instruments, then the internet, now automated generation — reliably widens who can attempt serious work. It does not reliably raise the average quality of what gets produced, and it tends to increase the volume of plausible material faster than the institutions for judging it can keep up. The flowering, when it happened, was produced by the combination of a new tool and a new set of institutions that made the tool’s output trustworthy.

The evidence for the tool side is already in hand. Galaxy Zoo’s volunteers matched professional classifications. Foldit’s players beat the lab’s own algorithm on problems that required strategic patience. AlphaGeometry and its successors now perform at a level that would have been remarkable institutional science a decade ago. The evidence for the institution side is thinner, because institutions change slowly and the tools are arriving quickly.

So the answer to whether cheap knowledge tools could produce a second Renaissance is that the tools make it possible and do not make it happen. What decides the outcome is whether anyone builds the verifying, funding, and preserving structures that turned the first flood of printed pages into cumulative knowledge. Two million candidate compounds and forty million galaxy classifications are, formally, the same kind of artifact: a mountain of raw material that becomes science only when someone takes responsibility for saying which parts are true. That responsibility is still, as it has always been, a human arrangement. The tools have made the mountains larger. They have not made the mountains climb themselves.