In early 2026, roughly 1 in 277 biomedical papers cited a source that does not exist.

Speed is the main driver of this trend. AI for science has made producing research output faster and cheaper than at any point in its history, and the checks that confirm whether that output holds have not kept up.

Most labs feel the acceleration but underestimate it. The volume of papers, models, and results keeps climbing. The share of that work a researcher can safely build on is not climbing with it.

The labs that handle this well have changed what they treat as the hard part. Producing a draft or a result is cheap now. Knowing whether the claim behind it is sound is the work that remains. So they add a verification step before they build on a finding.

This gap is already reshaping research. The practical question is where AI is moving fastest, which fields are falling behind, where speed is creating risk, and what labs can do now to keep their science verifiable.

‍

The Pace of AI Adoption in Research Is Not Linear

For most of its history, AI for scientific research played a supporting role. Models helped with formatting, summarizing papers, or generating boilerplate code. The researcher stayed in control of every decision.

That arrangement is changing fast. The current generation of agentic systems can plan a study, run computational experiments, write a full manuscript, and submit it for review without a human in the loop. Sakana AI's AI Scientist produced a paper that was accepted at a peer-reviewed ICLR workshop in 2025, the first time a fully AI-generated paper cleared that bar. The cost per manuscript was roughly $15.

Other architectures push throughput even further. Multi-agent systems now divide work across specialized modules for ideation, experiment design, execution, and writing. Some produce complete papers in hours, not months.

These systems still make basic errors. They cannot inspect their own visual output, which leads to broken figures and overlapping tables. They compare results to baselines unfairly. In early testing, one system modified its own code to extend a timeout rather than fix the inefficiency that caused it.

The pattern is clear. The rate at which AI generates research output is compressing, while the rate at which anyone verifies that output is not. Each cycle gets faster. The checks that used to happen naturally, because a human was doing the work, no longer happen by default.

This matters because adoption is not spreading evenly. Some fields absorbed these tools years ago. Others have barely started. And the gap between them is widening.

‍

Where AI for Scientific Research Is Already Embedded

The fastest adoption is concentrated in fields where physical laws and molecular structures map cleanly into deep learning architectures. 

These disciplines had large, structured datasets before the current wave of models arrived, and AI slotted into existing workflows with relatively little friction.

‍

Structural Biology and Proteomics

AlphaFold changed the economics of protein structure prediction almost overnight. The system now covers more than 200 million proteins, and by 2024 roughly one in five new structures deposited in global databases cited it. 

Adoption has been strongest in cryo-electron microscopy, where the hardest structures benefit most. The tool is not replacing experimental work, but rather pushing researchers toward problems they previously could not afford to attempt.

‍

Materials Science and Crystallography

AI for scientific discovery has had its most dramatic demonstration in materials science. Google DeepMind's GNoME used graph neural networks to predict the stability of inorganic crystals at a scale no human team could match. 

The system predicted 2.2 million new crystal structures and expanded the number of stable materials known to humanity from 48,000 to 421,000. Among them are candidates for next-generation batteries and superconductors. Traditional discovery methods would have taken decades to cover the same ground.

‍

Weather and Astronomy

Numerical weather prediction used to depend on computationally expensive fluid dynamics simulations. Modern deep learning models have largely bypassed that pipeline. 

Operational frameworks now generate high-resolution global forecasts in minutes, running an order of magnitude faster than their predecessors. In astronomy, foundation models trained on hundreds of millions of celestial objects signal a similar shift from field-specific tools toward general-purpose models for the physical sciences.

‍

The Fields Still Catching Up

While data-rich disciplines accelerate, other research domains face barriers that have little to do with the technology itself.

‍

Clinical Research

The potential for AI in diagnostics and therapeutic planning is well documented, but actual integration remains limited. Black-box models create liability risk in settings where every decision needs a clear rationale. 

Electronic health record systems are incompatible across providers, which makes it difficult to assemble the large, integrated datasets models need. Patient data protections restrict the kind of pooling that open-data fields take for granted. And complex billing structures make hospital administrators cautious about adopting tools without established precedent, even when the clinical case is strong.

These barriers reinforce each other. Fragmented data slows model development, which limits the evidence base for adoption, which keeps incentives misaligned.

Social Sciences and the Humanities

Many departments restrict AI use outright rather than teaching its responsible application. When students and researchers use these tools anyway, without guidance, the result is a shadow AI culture where outputs go unchecked, and risks go unaddressed.

The problem runs deeper than individual department policy. AI research funding and compute resources are concentrated in well-resourced private labs, and the models built there are optimized for technical workflows. They do not align well with the contextual reasoning, critical analysis, and open-ended inquiry that define these disciplines.

Why the Gap Keeps Widening

The lag matters beyond individual fields. As general-purpose models are shaped by the disciplines that adopt them earliest, the feedback loop further favors those fields. The longer a domain stays on the sideline, the harder it becomes to close the distance.

‍

Speed Without Validation Is a New Risk

The same tools that compress research timelines also compress the time between an error being created and its entry into the permanent record.

The most visible example is fabricated citations. Large language models generate text by predicting the next plausible word, not by querying verified databases. When used to draft manuscripts, they produce references that look real, including correct formatting, plausible publication dates, and real author names. But the underlying studies do not exist.

A Columbia University audit of nearly 2.5 million biomedical papers found over 4,000 fabricated citations across roughly 2,800 published articles. The rate grew more than 12-fold between 2023 and early 2026, with the sharpest increase beginning in mid-2024 as AI writing tools became widespread. In clinical fields, where treatment decisions depend on systematic reviews, a single fabricated trial reference can distort an entire guideline.

The problem is not limited to journal articles. In 2025, Springer Nature retracted a machine learning textbook after investigators found that roughly two-thirds of its sampled citations either did not exist or contained substantial errors. The book had passed through multiple layers of editorial review.

Fabricated references are the most measurable symptom, but the deeper issue is reasoning quality. On PaperArena, a benchmark testing end-to-end scientific tasks, the best AI agent scored 38.8% compared to 83.5% for PhD-level researchers. Models draft fluently, which makes their output easy to trust. But fluency is not accuracy. 

A confident, well-written paragraph can contain a flawed comparison, an invalid statistical claim, or a conclusion that stretches well past the data. The gap between how good the writing looks and how sound the reasoning is has become a risk factor in its own right.

‍

What Labs Should Be Doing Now

The response to faster, less reliable output is not to slow down. Instead, it’s adding verification as a standard step in the workflow. Several practical measures can close the gap between what AI produces and what a lab can safely build on, including:

• Adopt field-agnostic reporting standards: Any lab using machine learning models should follow a structured checklist. The REFORMS framework, developed by an interdisciplinary consensus of researchers at Princeton, covers 32 reporting items across eight modules: study goals, computational reproducibility, data quality, preprocessing, modeling, data leakage prevention, metrics and uncertainty, and generalizability. A checklist does not guarantee rigor, but it makes gaps visible.
• Standardize reproducible environments: Package code, dependencies, and computing specifications into containers so that results can be re-run by anyone. Without this step, reproducibility depends on undocumented assumptions about the environment the work was done in.
• Isolate validation from training data: Audit preprocessing pipelines to confirm that operations like normalization and oversampling happen after the train-test split. Leakage before the split inflates results and produces findings that do not generalize.
• Verify reasoning, not just prose: Move from using general-purpose language models as drafting tools to systems designed to test whether the claims in a manuscript actually hold. This is the step most labs still skip.

QED's Scientific Evaluation Engine is built for that last step by breaking a manuscript into individual claims, both explicit and implicit, and testing each against the published literature. It surfaces unsupported conclusions, weak inferences, and structural gaps, then returns structured findings AI for scientists can act on before submission. QED does not train on uploaded materials, so your work stays private.

‍

Build Validation Into Your Research Workflow

The gap between what AI can produce and what the scientific record can absorb is growing. 

Reporting checklists and reproducible environments help, but they do not test whether the claims in a manuscript are supported by the evidence behind them. That requires a different kind of tool.

QED's Scientific Evaluation Engine does that work. It identifies the claims in your research, checks each against the literature, and tells you where the argument is strong and where it needs attention. The entire process is author-controlled review that keeps your manuscript private, so your work stays yours from first draft to publication. 

Upload a manuscript or grant application and find out which claims hold before a reviewer does.

‍

FAQs