📊 How Far Has AI Penetrated Research? A Field-by-Field Survey

AI / Academia / Research / Trend Survey

Survey date: July 27, 2026

Table of Contents

• Conclusion
• 1. Numbers That Frame the Big Picture
• 2. Field-by-Field Penetration (Measured Directly on arXiv Full Text)
• 3. Field by Field: What Is AI Being Used For?
  • Computer Science / Machine Learning
  • Mathematics
  • Life Science / Drug Discovery
  • Chemistry / Materials Science
  • Physics / Astrophysics
  • Economics / Social Science
  • Medicine / Clinical Practice
  • Law
  • The Situation in Japan
• 4. How to Read This Situation
• References

Conclusion

“Writing papers with AI” is no longer the exception — in some fields, it has become the majority practice. That said, the degree of penetration is wildly different from field to field: in the direct measurement of arXiv full text, computer science sits at about 65%, versus about 0.7% for mathematics — nearly a hundredfold gap.

More importantly: the stage where “AI helps with the writing” and the stage where “AI drives the research itself” are two different things. The former has spread to nearly every field, while the latter is producing substantive results in only about three areas — formal proof in mathematics, drug discovery, and materials.

1. Numbers That Frame the Big Picture

That last row best captures where things stand right now. He and Bu’s study, published in PNAS, analyzed 5.2 million papers and concluded that “journal AI policies had almost no effect either on curbing AI use or on improving transparency.” There was no significant difference in the growth rate of AI use between journals that had a policy and those that didn’t.

Use is spreading; disclosure is not — that’s the default state of 2026.

2. Field-by-Field Penetration (Measured Directly on arXiv Full Text)

The measurement published by the detection tool Unslop in July 2026 is, at present, the most fine-grained data available. It scored the full text (not just the abstract, since the signal weakens there) of 12,750 arXiv papers spanning January 2023 to July 2026, calibrated against a 0.4% false-positive rate on 2021–22 papers from before ChatGPT existed.

Results for the most recent 12 months:

The pre-ChatGPT baseline was near zero across every field, so this represents the change over three and a half years.

A note on how to read this table

Limitations the researchers themselves point out:

1. Mathematics has a structural blind spot: it’s heavy on mathematical notation, which detectors have trouble reading as “AI-like prose.”
2. Detector sensitivity varies by AI model, so “the true share is at least this high.”
3. What gets flagged is “mechanical writing” — this also catches human manuscripts heavily edited by AI.

In particular, it would be a mistake to read mathematics’ 0.7% as “mathematicians aren’t using AI.” Mathematics is, in a different form, one of the fields where AI has cut in deepest (more on this below).

As background on the field differences, the PNAS study also reports that the increase is largest at non-English-speaking, physical-science, and high-open-access-ratio journals. Writing support for researchers who aren’t native English speakers is a major driver of the spread.

3. Field by Field: What Is AI Being Used For?

Computer Science / Machine Learning — Saturation and Institutional Breakdown Happening at Once

This is the field with the highest penetration, and at the same time, the field where the downsides started breaking the institution first.

An explosion in submissions

• NeurIPS: 9,467 papers in 2020 → 21,575 in 2025 (+220%)
• AAAI 2026: roughly 29,000 papers (over 75,000 total authors)
• ICLR 2026: 19,490 papers

The decline in quality shows up in the numbers

• NeurIPS 2026’s position-paper track: 273 of 969 papers (28.2%) scored highest on the detection tool Pangram. Even a more fine-grained pass judged 12.7% as substantially AI-generated. 178 papers (18.4%) were desk-rejected with no opportunity to appeal, and 123 were conditionally rejected pending “presentation of an online version with edit history.” The track had limited AI use to “peripheral edits such as copyediting.”
• The same is true on the review side: of about 75,800 reviews submitted to ICLR 2026, 21% were entirely AI-generated, and more than half showed some form of AI involvement (Pangram Labs analysis). Nature also reported that “more than half of researchers use AI for peer review against guidance.”
• NeurIPS 2025: at least 53 accepted papers (about 1%) passed peer review while containing fabricated citations.

The institutional response

In November 2025, arXiv shifted policy for review articles and position papers in the CS category, requiring “proof of having passed peer review and been accepted at a conference or journal” as a submission requirement. It also introduced a one-year submission ban for authors of clearly AI-generated material that hasn’t been checked by a human. First-time submitters now also need an endorsement.

The academic-politics angle

A position paper has appeared arguing that “denominator gaming” by fully automated scientific agents is a real risk: artificially inflating submission counts (the denominator) rather than paper quality (the numerator) can drive down the acceptance rate and manufacture the appearance of selectivity and prestige.

Mathematics — Last in the Text Metric, But Cutting-Edge in the Proofs Themselves

It sits at 0.7% on paper — dead last in the table — but mathematics is the first field where AI produced the substance of research in place of a human.

• In May 2026, Google DeepMind announced AlphaProof Nexus, an agentic framework combining Gemini 3.1 Pro with the Lean compiler
• Of 353 Erdős problems it attempted, it autonomously solved 9 — 2 of which had been unsolved for 56 years
• On OEIS conjectures, it proved 44 out of 492
• Harmonic’s Aristotle reached gold-medal-equivalent performance at IMO 2025 and now offers a public API

In mathematics, AI is entering not as “a tool for writing the prose of a paper” but as “a tool for searching out proofs verifiable in Lean” — an axis orthogonal to what the detectors are measuring. Because formal verification exists here, the “plausible-sounding lie” problem happening in CS is structurally much harder to reproduce.

Life Science / Drug Discovery — Industry Investment Is Reshaping How Research Is Done

Quantitative biology’s text-side number is 56.3%, the second highest, but the real action in this field is autonomous discovery systems.

Kosmos (Edison Scientific)

• Given a dataset and a natural-language goal, it runs up to 12 hours of cycles of data analysis, literature search, and hypothesis generation, and synthesizes a cited scientific report
• Across 7 case studies in metabolomics, materials science, neuroscience, and statistical genetics, it reproduced unpublished results and proposed novel mechanisms
• However, human-scientist verification found the descriptions only 80% accurate, and dataset selection and validity checking remain the premise that humans handle

Serious industry investment

• In early 2026, GSK paid NOETIK a $50 million upfront fee
• Eli Lilly is paying Chai Discovery tens of millions of dollars a year in access fees
• Lilly and Thermo Fisher have each announced collaborations with NVIDIA on autonomous labs

The paper-mill problem

• An estimated 1.5–2% of papers published in 2022 closely resemble paper-mill output
• A survey of 200,000 systematic reviews found 0.15% cite retracted paper-mill papers (oncology is most affected)

Chemistry / Materials Science — Toward the Second Generation of Self-Driving Labs (SDL)

The shared view across 2026’s review papers is that the field is moving from “SDL 1.0,” which showed the feasibility of closed-loop discovery, toward SDL 2.0, equipped with interoperability and breadth of application.

• Nanomaterials/catalysts: robotic coin-cell assembly, autonomous nanoparticle synthesis, automated solid-oxide-cell workflows
• Organic chemistry/bioengineering: AI-driven robotic chemists, ultra-fast flow-synthesis platforms, autonomous nanoparticle manufacturing
• South Korea, backed by its semiconductor, battery, and chemical industries, is one of the hubs of SDL adoption

On benchmarks, frontier models beat the average human expert on ChemBench, a set of over 2,700 chemistry problems.

Physics / Astrophysics — Cautious, But Foundation Models Have Arrived

Text-side penetration is low, at 10–34%. In a 2024 American Astronomical Society survey of roughly 500 people, usage was 33% for writing assistance and 48% for programming, showing that the norm of “use it for code generation, but not for the manuscript” is relatively holding.

The limits of capability are also clearly measured

• ReplicationBench: frontier models succeed at reproducing astrophysics papers under 20% of the time
• PaperArena: the best agent scores 38.8% vs. 83.5% for PhD holders
• On real bioinformatics tasks: roughly 17% for frontier models

Meanwhile, its track record as a tool keeps building up

• AION-1, the first astronomy foundation model, trained on over 200 million celestial objects from five major surveys, has arrived
• In weather, FourCastNet 3 generates a 60-day forecast in under 4 minutes (8–60× faster than conventional methods). In 2025, an AI system ran a weather forecasting pipeline end-to-end for the first time.

Economics / Social Science — “Text” at 47%, “Agents” at 20%

Text-side penetration in economics/finance is high, at 47%, but a survey of 1,260 quantitative social scientists that Anthropic ran in early 2026 shows a different picture.

Adoption rate

• Have tried an AI chatbot: 81%
• Regularly use a coding agent (e.g., Claude Code): 20%
• 86% of agent users use Claude Code

Usage skews overwhelmingly technical

• 97% use it for generating data-analysis code
• #2 is editing prose, followed by advice on methods and literature review
• Of all AI users, only a third used it for drafting the manuscript itself

The adoption gap is stark

• Graduate students/postdocs use it at 2× the rate of tenured professors
• Researchers with male-sounding names use it more than 2× as much as those with female-sounding names
• Those affiliated with a top-25 university use it 40% more than others
• This gap is sharper for coding agents than for general AI use

The issue on the qualitative-research side

Automation of thematic analysis and coding via LLMs is progressing, but questions remain unresolved: the validity of contextual understanding, the lack of systematic comparison against human coders, and the privacy problem of feeding sensitive data into ChatGPT. The dominant framing is not “replacement” but “a complement that lets manual coding be allocated more strategically.”

Medicine / Clinical Practice — Adoption in Practice Is Leading Adoption in Research

• 81% of US physicians use AI in clinical practice (AMA survey; doubled from 38% in 2023)
• More than three-quarters of physicians believe “AI improves the ability to care for patients” (up from 65% in 2023)
• On the research side, tools like Elicit, Consensus, SciSpace, and Scopus AI are cutting the time needed for a systematic review from months down significantly
• AI is also starting to be used on the research-integrity side — automated checks for compliance with reporting guidelines are an attempt to shift “from catching misconduct after the fact to gatekeeping beforehand”

Law — The Field Where Hallucinated Citations Became Visible as Real Harm

This is more a matter of practice than academia, but the impact reaches legal scholarship too.

• Court cases involving AI-fabricated citations reached 1,598 by June 2026 (up from about 200 a year earlier)
• Confirmed in 1,313 court proceedings as of April 2026, with 496 involving a licensed attorney
• The largest sanction: Couvrette v. Wisnovsky (Oregon district court), about $109,700
• In Q1 2026 alone, US courts imposed over $145,000 in sanctions
• Sanction amounts over time: $5,000 (2023) →$55,597 (2025) — an 11x increase in 18 months
• In Nebraska, an attorney was fully disbarred over an AI-submitted filing — a first in the US
• Technical background: even legal-specialized RAG tools hallucinate 17–34% of the time (Stanford/Yale, Dahl et al., 2024)

The Situation in Japan

• A survey of 532 universities nationwide found that about 60% are using generative AI in education. There are advanced cases using it for grading and admissions, while 40% remain stuck at the consideration/preparation stage out of caution about misuse — a split into two camps (Nikkei)
• Society-level guideline development is underway. The Japan Society for Public Relations and Communication published its “Guidelines on the Use of AI Tools in Research Activities” on June 30, 2026. It covers planning, investigation, analysis, writing, and peer review, and explicitly states that AI tools are not recognized as an author or co-author, and the author bears responsibility for the content of the research.
• The University of Tokyo’s university-wide guideline acknowledges the potential of AI while listing risks: misinformation, rights infringement, plagiarism, misconduct, and the erosion of thinking skills

4. How to Read This Situation

(1) Don’t confuse “text that reads as AI-like” with “research that has become AI-driven”

Much of CS’s 65% figure is writing support and polish for non-native English speakers, and doesn’t mean the substance of the research is AI-derived. Conversely, mathematics’ 0.7% doesn’t contradict at all the fact that AlphaProof Nexus is solving Erdős problems.

To compare fields fairly, you need to separate out these four layers:

(2) The more verifiable a field is, the more real AI’s output becomes

Mathematics (the Lean compiler), drug discovery (wet-lab verification), and materials (autonomous experiments) each have a mechanism to mechanically or physically falsify AI’s output, so hallucinations are less likely to make it into the final result.

By contrast, in CS and social science, where peer review is the only verification mechanism, hallucinations settle straight into the literature (the 53 NeurIPS 2025 papers are a real example).

(3) The institutional response is clearly playing catch-up

70% of journals have an AI policy, yet the disclosure rate is 0.1%, and there’s no significant difference in usage rate between journals with and without a policy — meaning norm-based control isn’t working.

The only measures that had real effect were ones involving detection plus mechanical rejection, like arXiv’s policy shift for CS or NeurIPS’s desk rejections. But detectors have a false-positive problem, and even Unslop’s 0.4% calibration produces a non-negligible number of false accusations at large scale.

(4) An access gap is showing up in a new form

The disparity Anthropic’s survey revealed — postdocs at 2× professors, male-sounding names at over 2× female-sounding names, top-25 universities at 40% higher — shows that productivity tools aren’t being distributed evenly.

The endorsement requirement arXiv introduced is reasonable as a countermeasure against AI spam, but it has been flagged as having the side effect of shutting out early-career researchers without an affiliation.

References

Primary Data / Direct Measurements

• Over 30% of papers submitted on arXiv read as AI-written (primary field-by-field measurement data)
• Academic journals’ AI policies fail to curb the surge in AI-assisted academic writing | PNAS
• AI-Generated Papers in the NeurIPS 2026 Position Paper Track – NeurIPS Blog
• Coding agents in the social sciences \ Anthropic
• Science | The 2026 AI Index Report | Stanford HAI

Academic Institutions / Research Integrity

• Position: Academic Conferences are Potentially Facing Denominator Gaming Caused by Fully Automated Scientific Agents
• Attention Authors: Updated Practice for Review Articles and Position Papers in arXiv CS Category
• arXiv Changes Rules After Getting Spammed With AI-Generated ‘Research’ Papers - 404 Media
• More than half of researchers now use AI for peer review — often against guidance | Nature
• GPTZero finds 100 new hallucinations in NeurIPS 2025 accepted papers
• Nearly 30,000 Submissions Flood NeurIPS | 36kr

Field by Field

• Google DeepMind’s AlphaProof Nexus Solves Erdős Problems
• Kosmos: An AI Scientist for Autonomous Discovery
• 12 AI Co-Scientists of 2026: From Biology to Math | Turing Post
• Toward self-driving laboratory 2.0 for chemistry and materials discovery | Materials Horizons
• AMA: AI usage among doctors doubles as confidence in technology grows
• Adoption of generative AI at the Italian institute for astrophysics | Frontiers
• AI Hallucination Legal Cases: A Sanctions Tracker (2026)

Japan

• 大学、生成AI活用6割 532校調査 向き合い方二極化 | Nikkei (Japanese)
• Japan Society for Public Relations and Communication publishes guidelines on AI tool use in research activities (Japanese)