Papers are written in one format but increasingly consumed in another.
For most of human history, scientific papers were written for human eyes alone — but that assumption has quietly dissolved. Large language models now read, summarize, and synthesize research at scales no individual scholar could match, yet the publishing formats that carry this knowledge were never designed with machines in mind. The gap between how science is written and how it is increasingly consumed is not merely a technical inconvenience; it shapes which knowledge becomes visible, which studies get integrated, and whose decades of work are honored or lost. A modest structural proposal — author-verified, machine-readable summaries alongside traditional abstracts — asks the scientific community to acknowledge that the reader has changed, and to build accordingly.
- LLMs now routinely process scientific literature at scale, yet they routinely confuse what a paper actually studied with what it merely mentioned — errors that quietly distort which evidence gets surfaced and which gets buried.
- Fields like neuroscience, which depend on synthesizing wildly different animal models, methods, and vocabularies, are especially vulnerable: the machines amplify existing biases toward conventionally formatted research, leaving unconventional work harder to find.
- A concrete experiment — the MetaBeeAI project on pesticide effects in bees — exposed exactly these failure modes, with LLMs stumbling over inconsistent terminology and mismatched method-to-finding pairings across a multi-disciplinary literature.
- The proposed fix is deliberate and bounded: a structured, LLM-drafted, author-corrected summary sitting beside every paper's abstract, capturing what was studied, how, and with what numerical results — verified by the people who know the work best.
- The stakes extend beyond accuracy: repeated LLM reprocessing of the same papers carries real computational and environmental costs, costs that standardized open summaries could largely eliminate while improving provenance and trust.
Scientific papers have always been reformatted as they move between journals, but the content stayed the same — the structure was designed for human eyes and printing workflows. That assumption no longer holds. Large language models now scan and synthesize scientific literature at scales no individual researcher could manage, yet publishing has not caught up to the reality that papers are written one way and increasingly consumed in another.
The problem compounds in subtle ways. When an LLM processes a scientific paper, it cannot reliably distinguish between a species that was actually studied and one merely mentioned in passing. When a paper describes multiple experiments, the machine struggles to correctly pair methods with findings. These errors propagate downstream, distorting which studies become visible to synthesis tools and amplifying biases toward conventionally structured research. Fields like neuroscience — which draws on molecular biology, anatomy, behavior, engineering, and insights from cephalopods, fruit flies, honeybees, and worms — are especially exposed. Decades of work across different vocabularies and experimental traditions become harder to integrate when the machines reading them are unreliable.
The MetaBeeAI project tested this directly, using LLMs to extract structured information from papers on how neurotoxic pesticides affect bees. The machines summarized key findings reasonably well but stumbled on terminology — different fields use different words for similar concepts — and on integrating results across many papers at once. The challenge of extracting facts in a format that can be compared across study types remained genuine.
The proposed response is not to resist machine reading but to design for it deliberately. Every paper should include an author-verified, machine-readable summary alongside the traditional abstract — not replacing it, but sitting beside it, freely available online. An LLM would draft this summary first, extracting what was studied, how, and the core numerical results. Authors would then review and correct it, ensuring accuracy and surfacing limitations. The verification step matters both for reliability and because it forces researchers to engage explicitly with how their work will be interpreted — a reflection valuable in its own right.
Precedents exist: structured reporting formats have improved reproducibility, genomics has demonstrated the power of shared metadata standards, and the STAR Methods initiative has shown what standardizing methodology can do. A machine-readable summary extends this logic to the paper itself. In the longer term, these summaries could evolve into explicit representations of scientific relationships — linking interventions to outcomes, circuits to behaviors, genes to phenotypes — making implicit prose connections legible to machines at scale.
The practical gains compound. Structured open summaries would reduce the computational waste of LLMs repeatedly reprocessing the same papers, improve provenance, and make literature synthesis inspectable rather than opaque. The choice is not between writing for humans or writing for machines. Scientific papers should remain rich human documents. But alongside them, the field needs transparent, author-verified representations that allow knowledge to integrate more reliably — because the reader has already changed, whether publishing has acknowledged it or not.
Scientific papers arrive at your desk formatted one way, then get reformatted again when they move to the next journal, and again after that. The content stays the same. The structure keeps changing. This made sense when humans were the only readers—publishers designed formats for human eyes and their own printing workflows. But something has shifted. Large language models now routinely scan, summarize, and synthesize scientific literature at scales no individual researcher could manage alone. Papers are being written in one format but increasingly consumed in another, yet the publishing world has not caught up to that reality.
The problem runs deeper than formatting quirks. When an LLM tries to extract structured information from a scientific paper, it encounters obstacles that seem small on the surface but compound quickly. A paper might mention a species in passing—something it compared against but did not actually test—and the machine cannot reliably tell the difference between what was studied and what was merely referenced. When a paper describes multiple experiments, the LLM struggles to correctly pair which methods produced which findings. These are subtle errors, but they propagate downstream. They distort which studies become visible to the tools that synthesize evidence and build models. Studies that fall outside standard paradigms become harder to find and integrate, and machines can amplify this bias by relying on consistent patterns in language that not all research follows.
Consider the landscape of modern neuroscience. The field sprawls across molecular biology, anatomy, neurophysiology, behavior, computer science, engineering. No single researcher can keep pace with all of it, yet the work demands a holistic understanding. Neuroscience has always advanced by drawing insights from diverse animal systems—neural activity in cephalopods, sleep in fruit flies, decision-making in honeybees, plasticity in crayfish, circuit function in worms. These studies span different levels of biological organization, use different methods, speak different vocabularies, follow different experimental traditions. Their datasets may not be directly comparable, but together they paint a broader picture of what nervous systems can do and how they adapt. If the information locked inside decades of scientific writing could be better integrated, it would honor not only the work of the scientific community but also the animal lives and public funding that made the research possible.
One research effort, the MetaBeeAI project, explored using LLMs to extract structured information from scientific papers—specifically, papers describing how neurotoxic pesticides affect bees. The work spanned molecular toxicology, neurobiology, behavior, and ecology. The machines proved good at summarizing key findings but stumbled in ways that matter for scientific synthesis. Terminology posed another obstacle. Different fields use different words for similar concepts. Even basic categories like "control" or "treatment" are not always described consistently across studies. LLMs can handle some of this variation, but performance drops when trying to integrate across many papers. Extracting facts in a standardized format that can be compared across different types of studies remains a genuine challenge.
The solution is not to resist this shift or to rely entirely on machines to interpret papers after the fact. Instead, scientific publishing should adapt to the reality that both humans and machines now read the literature. The proposal is straightforward: every paper should include an author-verified, machine-readable summary alongside the traditional abstract. This summary would not replace the abstract but sit beside it, freely available online. It would capture, in structured form, what was studied, how it was tested, and the core numerical results, with direct links to datasets where available. An LLM would generate a first draft, extracting key elements into a structured format. The authors would then review and correct the output, ensuring it accurately reflects the study and its limitations. This verification step matters not only for reliability but because the process itself forces researchers to engage explicitly with how their work is represented and interpreted—a reflection that would be valuable in its own right.
Precedents exist for this kind of approach. Structured reporting formats have improved reproducibility by standardizing how methods are described. In genomics, shared standards and metadata frameworks have enabled large-scale data integration. The STAR Methods initiative in the biosciences has demonstrated how standardizing methodology and protocols can enhance experimental reproducibility. A machine-readable summary would extend this idea to the level of the paper itself. In the longer term, these summaries could evolve beyond simple metadata into structured representations of scientific relationships themselves: explicitly linking interventions to outcomes, circuits to behaviors, genes to phenotypes. Scientific papers already contain these relationships implicitly in prose, but prose is ambiguous and difficult for machines to interpret consistently. Making these relationships explicit and machine-readable could fundamentally change how knowledge integrates across neuroscience and across fields that use different species, methods, and terminology.
The practical benefits compound. Right now, LLMs repeatedly process the same papers—accessing, reading, summarizing them again and again, with real computational and environmental costs. Structured, open summaries would render much of this unnecessary. AI systems could draw directly from these summaries rather than reinterpreting the full text each time. More importantly, they would have access to a consistent, author-verified representation of each study, improving provenance and trust. Researchers could inspect exactly what information was extracted and how it was structured, rather than treating literature synthesis as an opaque process. Science would become more accessible and efficient.
The choice is not between writing for humans or writing for machines. Scientific papers should remain rich human documents designed for interpretation, debate, and learning. But alongside them, we should build structured, transparent, and author-verified representations that allow machines to integrate knowledge more reliably at scale. Right now, the system is drifting toward a world in which science is increasingly interpreted by machines, with all the distortions that entails. If we want science to be more open, reusable, and easier to integrate, we need to take that new reality seriously and build publishing infrastructure that serves both readers and the tools that now read alongside them.
Citações Notáveis
We are entering a world where papers are written in one format but increasingly consumed in another, yet scientific publishing has not adapted to that reality.— The author
If we could better integrate the information locked inside decades of scientific writing, we would honor not only the work of the scientific community, but also the animal lives and public funding that made this research possible.— The author
A Conversa do Hearth Outra perspectiva sobre a história
Why does it matter if a machine confuses what a paper studied versus what it mentioned in passing?
Because that confusion scales. When you're synthesizing one paper, the error is small. But when an LLM is pulling patterns across hundreds or thousands of papers, those small mistakes compound into systematic bias. Studies that don't fit standard patterns become invisible.
You mention neuroscience specifically. Why is that field particularly vulnerable to this problem?
Because neuroscience is fundamentally interdisciplinary. A fruit fly researcher and a cephalopod researcher are studying the same principles but using completely different vocabulary, methods, even animal models. A machine trained on consistent patterns will struggle to see what they have in common.
The author-verified summary sounds like extra work for researchers. Why would they do it?
It's actually not that different from what researchers already do—they write abstracts, they review papers, they think about how their work fits into the field. The difference is that this process would be explicit and structured. And the payoff is real: better integration of their work into the broader literature, less computational waste, more accessible science.
Could these summaries eventually replace the full paper?
No, and that's important. The full paper is where the debate lives, where you see the uncertainty, the limitations, the thinking process. The summary is a bridge—it lets machines find and integrate what matters, but humans still need the rich, messy original.
What happens if an author gets the summary wrong?
That's why verification matters. The author has to check it. They know their work better than anyone. And because it's structured and transparent, other researchers can inspect it directly. If something is wrong, it's visible.
This feels like it could slow down publishing even more.
It could, or it could speed things up. Right now, every LLM that touches a paper has to reinterpret it from scratch. Structured summaries would eliminate that redundant work. And the verification step, while it takes time, forces clarity that often improves the paper itself.