New Article: The Burden of Discernment
Scientific institutions are struggling to survive under a mountain of publications and poor standards of peer review. Artificial Intelligence could usher in a new era of knowledge, or be its downfall.
This article by Snezana Gvozdenovic was published on Palladium Magazine on May 23, 2026.
In 1633, the year after Galileo Galilei published his Dialogue Concerning the Two Chief World Systems, a book that made a comparative argument strongly favoring heliocentrism, his work was banned by the Roman Catholic Church, and he was sentenced to lifelong house arrest. On October 31, 1992, after a 13-year investigation into the condemnation of the Italian astronomer, Pope John Paul II officially closed the inquiry and formally acknowledged the Church’s error in the affair. The Supreme Pontiff also stated that science and the Christian faith are not in opposition, and that contemporary culture requires constant effort to synthesize knowledge: “a true culture cannot be conceived of without humanism and wisdom.”
As artificial intelligence reshapes our relationship with knowledge, this synthesis is perhaps more relevant than ever. It is also more nuanced than what is highlighted about the case in popular tellings of the history of science. It then isn’t a coincidence that Galileo’s case is also invoked to bolster rejected or inadequate claims—a logical fallacy called the “Galileo gambit” that conflates scientific criticism with undue persecution, and one that has seen use by alternative medicine enthusiasts, vaccine and climate-change skeptics, conspiracy theorists, and pseudoscience proponents. It isn’t only good ideas that attract passionate advocates.
Whether as individual thinkers or as a civilization, we cannot simply drop the burden of discernment, the often difficult task of distinguishing sound scientific challenges from pseudoscience, and ultimately of distinguishing what is correct from what is not. The case reveals a fundamental tension in how society handles challenges to consensus that remains today: while scientists face legitimate concerns about their ideas being stifled or dismissed prematurely, this same concern is exploited to shield poorly evidenced claims from appropriate scrutiny.
However, scientists and researchers who take on entrenched views in academia and in the public eye today face daunting moral and professional challenges. These challenges are more subtle than papal decrees or house arrests, but academic gatekeeping, algorithmic visibility controls, funding pressures, and a strained, long-overextended peer-review system can have a similar detrimental effect.
This problem of institutional discernment has perhaps been the fundamental problem of knowledge and humanity’s relationship with it all along. Can science do better, or are we bound to continue seeing the same problems in new technologically and socially updated guises?
The Institutions of Knowledge
Institutions dedicated to accumulating and sharing knowledge evolve in lockstep with technological progress, with each major technological advancement reshaping our information preservation methods and how we synthesize knowledge into understanding. Pre-literate societies relied on oral traditions with structured memorization and performance to transmit knowledge across generations. The advent of writing led to libraries, transforming knowledge from ephemeral recitation into collectible, searchable, and cumulative works. Later, medieval monastic libraries played an important role in preserving valuable texts through societal collapse, while the emergence of universities in the 11th and 12th centuries helped institutionalize structured debate and formal teaching.
The printing press enabled widespread distribution outside of localized institutions, giving rise to the Republic of Letters—an international network of intellectuals such as the botanist and physician Carl Linnaeus, who developed and popularized a system for classifying life or Gottfried Wilhelm Leibniz, one of the most influential mathematicians who ever lived, and the author of over 15,000 letters to hundreds of scholars. In the Enlightenment era, knowledge became a recognized commodity, with rising specialization prompting the creation of scientific disciplines and laboratories. In contrast to prior bodies geared toward archiving and deliberation, laboratories rigorously expanded insights in focused areas, fueling the momentum of modern scientific progress.
With the advent of laboratories, the need to share and discuss discoveries became paramount. The emergence of scholarly publishers who recorded and distributed research and experimental outcomes, enabled the scientific community to scrutinize, debate, and build upon the existing knowledge base. This tradition began with the early printed journals, such as Philosophical Transactions of the Royal Society dating back to the 17th century. While printed journals set the groundwork for the distribution of findings and sharing of ideas, contemporary digital scholarly publishing platforms have greatly expanded the volume and accessibility of published work.
The digitalization of knowledge brought its own challenges with scientists and researchers facing an overwhelming amount of data and published material. NASA’s EOSDIS program is estimated to archive over 500 petabytes of data by 2030, while CERN has already reached a one-exabyte milestone as of 2025. At the same time, arXiv recorded a steady year-over-year increase of paper submissions since August 1991, rising a thousandfold over 35 years. There are signs our ability to integrate and make sense of this digital deluge is lagging behind. Yet for the vast increase of the papers submitted, each paper is becoming less disruptive, and less likely to “break with the past in ways that push science and technology in new directions.”
Artificial Intelligence technology which excels at distilling large bodies of text and structured data, is in a timely alignment with the need for a transformation in science, promising both the potential to accelerate the expansion of academic publishing and the problems that come with it even as it holds vast potential to help humanity advance science. For now, we are yet to see its real impact on science and knowledge as a whole.
Each subsequent technological breakthrough has necessitated novel frameworks to synthesize knowledge. All relied on human judgment in some form or another. Oral traditions required the judgment of elders to decide which stories mattered; libraries required scholars to determine which texts to preserve and copy, and laboratories required peers to judge which findings held. These transformations can be seen as differing approaches for how to structure distributing the burden of discernment.
The Volume of Papers Has Overwhelmed Academic Journals
Despite a narrow audience, academic publishing is one of the highest-margin businesses today, with profit margins as high as 40% mirroring tech giants like Google and Microsoft. For many years, this field primarily functioned as a service to the academic community and was managed by scientific societies relying on handouts to cover printing costs rather than large conglomerates. The transformation occurred in the post-World War II era, when in 1951, Robert Maxwell, the enigmatic Czechoslovak-born British entrepreneur and father of the infamous socialite and convicted sex trafficker Ghislaine Maxwell, took over a small scientific publisher that became Pergamon Press.
Maxwell recognized an opportunity to capitalize on what he called a “perpetual financing machine” of academic publishing with the unlimited demand for information about emerging scientific fields and free academic labor. The business model Maxwell pioneered is still largely in use today: it relies on scientists providing articles to journals based on publicly funded research and conducting peer review without compensation, while publishers editorialize and package this labor into scientific journals and sell subscriptions back to academic institutions at premium prices.
Today, scientific journals are primarily online, with several open-access business models emerging in recent years in addition to the traditional subscription-based model: green open access, by which publishers permit eligible authors to self-archive a free copy of their article (usually a preprint); gold open access, with article processing charges, where authors pay to publish; diamond or platinum open access funded by grants or institutions (free for both authors and readers); and institutional membership agreements. Separately, there are preprint servers like arXiv that distribute research freely but without the peer review process. These models are a result of decades of efforts by the scientific community to change the current system in favor of open-access publishing and open science, but they hardly solve the challenges that stem from the shift originating in the 1950s.
Maxwell’s original paywall model created access inequality and hypercompetitive conditions in academic publishing. It didn’t invent the ‘publish or perish’ culture rooted in the academic tenure system, but the commercialization of journals and increased competition raised the stakes of where one published. In the early 1960s, Eugene Garfield established the Journal Impact Factor (JIF) as an aid to help librarians select journals. JIF stands as a measure of a journal’s impact, calculated from how often a journal’s recent articles are cited. However, what was designed to be a tool to sharpen discernment has in practice come to be seen as a measure of prestige and influence. It is no surprise that this, and metrics like the “h-index,” which scores authors by citation count, are seen as perverse incentives: they can be gamed yet still drive funding and tenure decisions. The more we lean on them, the more our own judgment atrophies.
At the same time, the mounting number of publications following the commercialization of scholarly publishing and the rise of the internet is straining peer reviewers—independent field experts tasked with evaluating quality, validity, and originality of research. Editorial retractions and self-retractions are to be expected in science, but the fact that we shape our world with potentially flawed research outputs is troubling. An example of this is a 2024 Nature study that projected climate-change damages far higher than previous estimates—figures later found to be inflated by flawed data and methodology, leading the authors to retract in 2025. Prior to its retraction, the study echoed around the world–generating numerous citations and headlines, and even being “incorporated in risk management scenarios used by central banks.”
In addition, there are growing concerns about misconduct, plagiarism, and error. The rising number of papers withdrawn for said reasons points to a system that’s failing under pressure. Some individuals take matters into their own hands. The molecular biologist and blogger Sholto David spends a significant amount of his free time reviewing published scientific papers and pointing out issues which, he says, “run the gamut from avoidable human error to evidence of scientific fraud.” Elisabeth Bik, microbiologist also known as an expert on scientific integrity and the founder of Microbiome Digest, works on identifying potentially doctored scientific images in published papers, and has found thousands of such cases.
Clearly discernment capacity hasn’t scaled with the sheer quantity of output. Generative AI, seemingly eagerly adopted by some to produce research papers—at times with fabricated references—makes such institutional discernment capacity even more pressing. It is ultimately the researcher’s responsibility to go to the source of their references and validate them, but that doesn’t change the fact that keeping carelessness, or even fraud, in check has become more difficult. This has been partially counteracted by software such as Proofig AI and Imagetwin that have been used both by independent investigators and peer reviewers to identify potential flaws in research papers.
A recent study analyzing more than one million scientific papers and preprints that have been published on arXiv, bioRxiv, and Nature portfolio journals between 2020 and 2024, estimates that up to 22% of the text in computer science papers shows signs of large language model modification. Tools to detect AI writing are perhaps as fraught as language models themselves, so this finding should be taken with a grain of salt. The use of models in writing academic papers is however so widely observed and known that perhaps the exact number doesn’t even matter. It is large enough.
It’s important to note that peer review, as a concept, regardless of who is conducting it, is inherently limited. It forces researchers to produce work that fits certain criteria and can be understood by reviewers, making science more incremental as opposed to probing. Preprint servers were designed to circumvent that in an effort to give science more room to breathe, but in practice most preprints are way-stations to publication: the majority are ultimately submitted to peer-reviewed journals, so they tend to be written with that gate already in mind. AI review tools promise to allow people to tap into all of humanity’s knowledge to challenge scientific claims and help illuminate gaps, but it’s doubtful these tools can meaningfully shift the old mindset.
There are real practical problems with folding AI into how we generate knowledge. But they sit beneath far larger questions raised in light of the rapidly evolving nature of this new technology.
The Accelerating Scientific Age
If AI can evolve to create hypotheses, develop and perform experiments, and write papers, what remains for humans in the realm of science? This point of anxiety has come to be voiced by more and more scientists and researchers. The answer is judgment. While we have undeniably come to productively use AI to reduce our workloads, for now human input and evaluation remain necessary at every step. And even as AI inevitably advances, we must continue not to outsource our capacity to discern which questions and answers matter, and which paths are worth pursuing. To do otherwise would amount to humanity abdicating its claims to truth.
In his essay Situational Awareness: The Decade Ahead, the former OpenAI technical staff member and investor Leopold Aschenbrenner suggests that superintelligence taking off this decade is plausible and could be the most volatile and dangerous period in human history. Once general artificial intelligence is achieved, Aschenbrenner argues that AI research will effectively automate itself through recursive self-improvement: think of AlphaGo learning from the best human samples of the game Go, then AlphaGo Zero playing against itself only to perform better than any human ever could. Perhaps that is indeed the beginning of a chain reaction of a kind of intelligence explosion. The prospect of, at the very minimum, AI and machine learning research being fully automated by AI itself is both appealing and unnerving, especially if it comes to drive expansion in all other scientific fields as well.
Perhaps we abdicated our claim to truth first and only began developing the technology to act on it after. In the second half of the twentieth century, science fiction authors, philosophers of science, and even physicists themselves, frustrated by seemingly slow scientific progress, had begun to wonder whether superhuman intelligence was in fact needed for the next scientific breakthroughs. This was long before AI started producing plausibly important scientific results such as the recent announcement that an OpenAI model has made a breakthrough on one of the challenges posed by the Hungarian mathematician Paul Erdős in 1946 and then remained unsolved for eighty years. As we are confronted with such pressing technological developments and ponder the state and structure of science and our relationship with knowledge, we must be open to considering that the solution might just require us to build an entirely new system for communicating and evaluating science.
A system that would realign academic incentives and behaviors, and apply tools that extend human intuition, taste, and agency to work in tandem with artificial intelligence. The goal, of course, is to eliminate the displacement risk scientists fear by allowing humans to continue to gain and master knowledge, without stifling its growth. We’ll almost certainly need artificial intelligence to achieve all this. This won’t be easy, and will require making some difficult choices.
Yes, science can do better. However, the systems we need aren’t necessarily the ones that are faster at processing more papers. If technology is our forbidden fruit, we’ve already eaten it; we urgently need to solve the problem of discernment at an unprecedented scale and pace. We should use this urgency as an opportunity to rebuild our collective capacity for judgment—extending humanity’s scientific discernment rather than replacing it, distributing the discernment burden as we’ve successfully done in past technological disruptions of knowledge production, rather than automating it away and abdicating human claims to truth. Whether we’ll successfully engineer such new frameworks or proceed haphazardly with an obsolete mindset that will likely amplify already entrenched inefficiencies remains to be seen.
Snezana Gvozdenovic is the founder of Universal Scientific Protocols, Inc., a San Francisco based company designing systems to address the challenges in scientific and research communication and learning. You can follow her at @snnneee.