2026.04 · Pebblous Data Communication Team
Reading time: ~20 min · 한국어
Executive Summary
The era of AI autonomously writing scientific papers is dawning. Sakana AI Scientist v2 demonstrated end-to-end automation from idea generation to paper writing, and JAIGP (Journal of AI-Generated Papers) launched in February 2026 as the first dedicated journal for AI-authored papers. While all Big 5 publishers ban AI authorship, the boundaries of the academic ecosystem are already shifting.
Yet the gap between reality and hype remains wide. While it is true that AI Scientist v2 passed peer review at ICLR (International Conference on Learning Representations) 2025, this was at a workshop track (~32.6% acceptance rate), and only 1 of 3 submitted papers was accepted before being voluntarily withdrawn per a prior agreement. The paper published in Nature 651 (2026-03-26) was a human-authored paper by Lu, C. et al. describing the AI Scientist system — not a paper written by AI. As of April 2026, no AI-solo authored paper has been published in a major peer-reviewed journal.
The proliferation of AI-generated papers carries a structural risk: contamination of academic databases. The share of papers containing hallucinated citations surged ninefold, from 0.3% in 2024 to 2.6% in 2025. In this environment, the value of data provenance tracking and quality diagnostics rises sharply — underscoring the increasingly critical role of DataClinic's L1/L2/L3 diagnostic framework in the age of AI science.
1. The Rise of AI Scientists — Two Experiments
Between 2024 and 2026, science research automation emerged as a key competitive front for Big Tech (OpenAI, DeepMind, Anthropic) and startups (Sakana AI) alike. Two experiments at the center of this wave are fundamentally reshaping the academic ecosystem. One is Sakana AI Scientist v2's answer to the question "Can AI write papers?" and the other is JAIGP's answer to "Where do you publish AI-written papers?"
These two projects take complementary yet fundamentally different approaches. AI Scientist v2 automates the entire research pipeline — idea generation, code writing, experiment execution, and paper drafting. JAIGP builds the publication channel for such papers. Meanwhile, OpenAI aims to build an autonomous AI researcher by 2028, DeepMind is pursuing AI co-scientist and AlphaEvolve, and Anthropic invested billions through Claude for Life Sciences and the $400M Coefficient Bio acquisition.
Big Tech's AI Science Automation Race
| Company | Key Project | Strategy |
|---|---|---|
| OpenAI | Deep Research | Targets autonomous AI researcher by 2028 |
| DeepMind | AI co-scientist, AlphaEvolve | UK automation lab + algorithm design agent |
| Anthropic | Claude for Life Sciences | Coefficient Bio $400M acquisition, life sciences focus |
| Sakana AI | AI Scientist v2 | Series B $135M ($2.65B valuation), end-to-end paper automation |
2. JAIGP: The Birth of an AI-Only Journal
JAIGP (Journal of AI-Generated Papers) is the first dedicated academic journal where AI serves as the author and humans act solely as prompters. Since launching in February 2026, it has published 43 papers as of April 2026, with diverse AI models including Claude, ChatGPT, Gemini, and Grok listed as authors.
JAIGP's founder, César Hidalgo, is currently a professor at the Toulouse School of Economics (formerly at MIT Media Lab, 2010–2019). He created this alternative channel at a time when the traditional publishing system entirely excludes AI authorship. Since March 2026, JAIGP has introduced a multi-stage peer review system: AI screening → ORCID endorsement → community comments → AI reviewers → human final evaluation.
Key point: JAIGP represents a milestone in the history of academic publishing, but without an Impact Factor, it has not gained recognition from traditional academia. As of April 2026, JAIGP is not an "accredited academic journal" in the traditional sense.
Nevertheless, JAIGP's experiment raises important questions. How should we evaluate the value of AI-generated research? Can the traditional academic system permanently exclude AI authorship? JAIGP's 43 papers — and their quality variance, from strong to weak — are accumulating empirical data on these questions.
3. Sakana AI Scientist v2: From Hypothesis to Paper
Sakana AI's AI Scientist v2 automates the entire scientific research process — from idea generation to code writing, experiment execution, and paper drafting. Three key improvements over v1 stand out: (1) removal of template dependency, (2) Agentic Tree Search for research path optimization, and (3) a VLM (Vision-Language Model) feedback loop for automated interpretation of experimental results.
On the cost front, AI Scientist v2 generates a paper for roughly $20–$25 (experiments $15–$20 + writing ~$5). Compared to human researchers' publication costs — formatting alone at a median of $477, open access Article Processing Charges (APCs) of $3,500–$4,000, plus labor — the difference exceeds 100x.
AI Scientist v1 vs v2 Comparison
| Feature | v1 (2024) | v2 (2025) |
|---|---|---|
| Research design | Template-based | Template-free, open-ended |
| Search method | Linear pipeline | Agentic Tree Search (multi-path exploration) |
| Result interpretation | Text-based | VLM feedback loop (visual analysis) |
| Peer review outcome | Not submitted | ICLR 2025 workshop: 1 of 3 accepted |
| Cost | ~$15/paper | $20–$25/paper |
But behind the cost revolution lies significant quality variance. Of the 3 papers submitted to ICLR 2025, the accepted paper scored an average of 6.33/10 (individual scores: 6, 7, 6), while the 2 rejected papers scored only 3–4/10. Even the accepted paper contained quality issues, including LSTM concept attribution errors and definitional inaccuracies.
4. Fact Check — Between Reality and Hype
The achievements of AI science automation are real, but they are systematically overstated in media coverage. Below is a fact check of the most common claims.
"AI published a paper in Nature"
The paper published in Nature 651 (2026-03-26) was written by human researchers (Chris Lu et al. at Sakana AI) describing the design and results of the AI Scientist system. It is a paper about an AI system, not a paper written by AI. As of April 2026, no AI-generated paper has been published in a major peer-reviewed journal.
📌 Where did this misconception come from?
Two sources are primarily responsible.
- ▸ Sakana AI's official blog title — "The AI Scientist: Towards Fully Automated AI Research, Now Published in Nature" (2026.03). The phrase "Now Published in Nature" reads as if AI's own research was published, not a human-authored description of the system.
- ▸ phys.org — "AI writes a research paper that passes peer review" (2026.03). This framed the Nature 651 paper as if AI had directly authored it.
TechCrunch was a notable exception, titling their piece "it's a bit more nuanced than that."
Source: Nature DOI:10.1038/s41586-026-10265-5, Sakana AI Blog
"AI Scientist passed ICLR peer review"
Partially true, but context matters. One of 3 papers was accepted at a workshop track of ICLR 2025 (ICBINB — "I Can't Believe It's Not Better"). The workshop's acceptance rate was ~32.6% (14 of 43 papers, per PMLR Volume 296), which differs from the ICLR main track (32.08%, 11,565 papers) in both scale and competitiveness. Furthermore, the accepted paper was voluntarily withdrawn per a prior agreement, meaning it was never actually published.
Source: PMLR Volume 296, arXiv:2504.08066, SakanaAI GitHub
"An AI-solo paper has been published in a major journal"
As of April 2026, there is no case of an AI-solo authored paper being published in a major peer-reviewed journal. The 43 papers in JAIGP have not gained traditional academic recognition, and the AI Scientist v2 paper accepted at the ICLR workshop was voluntarily withdrawn.
"JAIGP is an accredited academic journal"
JAIGP is an experimental journal launched in February 2026 that holds significance as a milestone in academic publishing history. However, it has no Impact Factor and has not gained official recognition from traditional indexing systems (Scopus, Web of Science, etc.). That said, it continues efforts to build academic rigor, having published 43 papers and introduced a multi-stage peer review process.
Source: jaigp.org, César Hidalgo's blog
"It costs only $20–$25 to generate a paper with AI"
True, based on AI Scientist v2. Experiment costs of $15–$20 plus ~$5 for writing total $20–$25 per paper. Compared to human researchers' publication costs (APCs of $3,500–$4,000, plus labor), the difference exceeds 100x. However, 2 of 3 papers were rejected (scores 3–4/10), so the cost-quality tradeoff must be considered.
Source: SakanaAI GitHub README, Enago/Editage
5. The AI Policy Landscape in Academic Publishing
All Big 5 publishers ban AI authorship, though their specific policy implementations vary. COPE (Committee on Publication Ethics) states that "AI cannot be held accountable and therefore cannot be an author." ICMJE (International Committee of Medical Journal Editors) determined that AI cannot meet its four authorship criteria (substantial contribution, critical revision, final approval, and agreement to be accountable). In 2026, ICMJE added a dedicated AI section (Section V).
| Publisher | AI authorship ban | AI disclosure | AI image policy | AI in peer review | Notable |
|---|---|---|---|---|---|
| Springer Nature | Yes | Required | Banned | Banned | Nature editorial comment (2026.03) |
| Elsevier | Yes | Required | Limited | Banned | Separate AI usage section required |
| Wiley | Yes | Required | Banned | Banned | Hindawi 11,000+ retraction crisis |
| Taylor & Francis | Yes | Required | Limited | Banned | Enhanced CRediT author declaration |
| SAGE | Yes | Required | Limited | Banned | AI literacy education program |
However, strong defenses do not mean there are no points of entry. Project Rachel (arXiv:2511.14819) demonstrated that a fictional AI academic identity could accumulate citations and even receive peer review invitations. The Nature editorial board urged in a March 25, 2026 commentary that "institutions, funders, and publishers must respond." This marks an institutional turning point — an acknowledgment that current policies cannot fully block AI's infiltration into academia.
6. Data Quality and Scientific Trust
The spread of AI-generated papers creates a chain of risks: hallucinated citations, data contamination, and model collapse. Understanding these dangers requires facing the numbers at each stage.
The Surge in Hallucinated Citations
The share of papers containing hallucinated citations — references to papers that do not actually exist — surged from 0.3% in 2024 to 2.6% in 2025, a ninefold increase. In computer science, 2–6% of all papers contain unverifiable references. AI-assisted papers now account for 13.5% of all publications, rising to 22.5% in CS.
The Recursive Loop of Model Collapse
Shumailov et al. (Nature 2024) demonstrated that model collapse occurs when the proportion of synthetic data exceeds 30–50%. When AI-generated papers enter academic databases and then feed back into AI training data, a recursive contamination loop forms: quality degradation → worse AI papers → further quality degradation.
Data Contamination Loop
AI Detection Tools: Current State and Limitations
Currently, GPTZero (99.3%), Originality.ai (99%), and Turnitin (92–100%) show high accuracy. However, even light editing causes accuracy to drop by 20–30%. More critically, approaches like AI Scientist v2 — which actually executes code and generates experimental data — are difficult to catch with simple text detection. An arms race between detection and generation technology is now underway.
The retraction landscape is equally alarming. In 2023, a record-breaking 10,000+ retractions occurred, with 335+ directly AI-related (2023–2024, per Frontiers/PMC). Wiley/Hindawi alone accounted for 11,000+ retractions.
7. Opportunities and Risks
AI science automation offers opportunities for cost reduction and speed innovation in research, but adoption without quality verification carries academic and legal risks.
Immediately Actionable Use Cases
- Literature review automation: bulk paper screening and summarization
- Hypothesis generation support: exploring new research directions from existing patterns
- Experiment code writing: automated statistical analysis and data preprocessing code
- Draft writing assistance: accelerating paper structuring and first drafts
Essential Safeguards
- Hallucinated citation verification: confirm that every AI-generated reference actually exists
- Publisher policy compliance: all Big 5 require AI usage disclosure
- Reproducibility assurance: independent verification of AI-generated experimental results
- Legal liability clarity: no legal framework exists for AI authorship
Where Synthetic Data Meets AI Science
AI Scientist v2 generated synthetic arithmetic expression datasets to test compositional generalization — directly linking to the field of synthetic data generation and augmentation. For AI to autonomously design and run experiments, input data quality must be guaranteed — this is the core premise of AI-Ready Data.
"In the age of AI-written science, trustworthy data becomes even more important." The AI Scientist v2 pipeline — "hypothesis → code → experiment → paper" — is structurally isomorphic to DataClinic's "diagnose → prescribe → optimize" pipeline. The greater the quality variance of AI-generated papers (1 workshop acceptance vs. 2 rejections), the higher the market value of systematic data diagnostic tools. The DataClinic (diagnosis) → DataGreenhouse (synthesis/augmentation) → AI-Ready Data (quality assurance) chain can serve as the critical data quality infrastructure for the age of AI science.
FAQ
Did AI really publish a paper in Nature?
No. The paper published in Nature 651 (2026-03-26) was authored by human researchers (Chris Lu et al. at Sakana AI) describing the AI Scientist system's design and results. It is a paper about an AI system, not a paper written by AI. As of April 2026, no AI-solo authored paper has been published in a major peer-reviewed journal.
What does it mean that AI Scientist v2 passed ICLR peer review?
One of 3 papers was accepted at the ICLR 2025 workshop track (ICBINB — "I Can't Believe It's Not Better"). The workshop's acceptance rate was ~32.6% (14 of 43 papers), which differs from the ICLR main track (32.08%, 11,565 papers) in scale and competitiveness. The accepted paper was voluntarily withdrawn per a prior agreement and was never published.
Is JAIGP an accredited academic journal?
As of April 2026, JAIGP has not gained official recognition from traditional academia (such as an Impact Factor). Since launching in February 2026, it has published 43 papers and introduced a multi-stage peer review system starting March 2026, but it has not yet established academic authority. Its significance lies in being a historic experiment as a dedicated publication channel for AI-generated papers.
How much money does AI save when writing papers?
With AI Scientist v2, generating one paper costs about $20–$25 (experiments $15–$20 + writing $5). Human researchers face formatting costs (median $477), open access APCs ($3,500–$4,000), plus labor — often totaling tens of thousands of dollars. However, 2 of 3 AI-generated papers were rejected, so the cost-quality tradeoff should be carefully weighed.
What is the stance of academic publishers on AI authorship?
All Big 5 publishers (Springer Nature, Elsevier, Wiley, Taylor & Francis, SAGE) refuse to recognize AI as a paper author. COPE (Committee on Publication Ethics) states that AI cannot be held accountable and therefore cannot be an author, and ICMJE (International Committee of Medical Journal Editors) determined that AI cannot meet the four authorship criteria. However, they do not ban AI usage itself — they require transparent disclosure.
Can AI-written papers be detected?
Currently, GPTZero (99.3%), Originality.ai (99%), and Turnitin (92–100%) show high accuracy. However, even light editing drops accuracy by 20–30%, and approaches like AI Scientist v2 — which executes real code and generates experimental data — are difficult to catch with simple text detection. An arms race between detection and generation technology is underway.
Can AI-generated papers contaminate academic data?
This is already happening. AI-assisted papers account for 13.5% of all publications (22.5% in CS), and papers with hallucinated citations reached 2.6% in 2025 — a ninefold increase from the prior year. Shumailov et al. (Nature 2024) demonstrated that model collapse occurs when synthetic data exceeds 30–50%. The recursive loop of AI-generated papers entering academic databases, being used as AI training data, and producing lower-quality output is accelerating.
Can corporate R&D teams use AI Scientist-type tools?
Literature review, hypothesis generation, experiment code writing, and draft writing are immediately actionable use cases. However, verifying hallucinated citations, complying with publisher policies, and ensuring reproducibility are essential. The core question is "How do you verify what AI generates?" — which is why building data quality diagnostic infrastructure ahead of time is critical.
References
1. Lu, C. et al. "Towards end-to-end automation of AI research." Nature 651(8107), 914–919 (2026). DOI: 10.1038/s41586-026-10265-5
2. Yamada, Y. et al. "The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search." arXiv:2504.08066 (2025)
3. Monperrus, M. et al. "Project Rachel: Can an AI Become a Scholarly Author?" arXiv:2511.14819 (2025)
4. Shumailov, I. et al. "AI models collapse when trained on recursively generated data." Nature (2024). DOI: 10.1038/s41586-024-07566-y
5. Tie, G. et al. "A Survey of AI Scientists." arXiv:2510.23045 (2025)
6. "Can Generative AI Survive Data Contamination?" arXiv:2602.16065 (2026)
7. COPE Position Statement: Authorship and AI Tools — publicationethics.org
8. ICMJE Recommendations (2026 update) — icmje.org
9. Sakana AI, "The AI Scientist Generates its First Peer-Reviewed Scientific Publication" (2025-03-12) — sakana.ai
10. Sakana AI, "The AI Scientist: Now Published in Nature" (2026-03) — sakana.ai
11. Nature Editorial, "AI scientists are changing research — institutions, funders and publishers must respond." Nature 651, 853–854 (2026) — nature.com
12. Nature News, "How to build an AI scientist." — nature.com
13. JAIGP (Journal of AI-Generated Papers) — jaigp.org
14. Hidalgo, C.A. "An AI Tsunami is about to Hit Science" (2026-03-06) — cesarhidalgo.com
15. TechCrunch, "Sakana claims its AI paper passed peer review — but it's a bit more nuanced than that" (2025-03-12)
16. TechCrunch, "Academics accuse AI startups of co-opting peer review for publicity" (2025-03-19)
17. TechCrunch, "Sakana AI raises $135M Series B at a $2.65B valuation" (2025-11-17)
18. OpenAI, "Introducing Deep Research" (2025-02) — openai.com
19. Google Research, "Accelerating scientific breakthroughs with an AI co-scientist" (2025-02) — research.google
20. DeepMind, "AlphaEvolve: A Gemini-powered coding agent" (2025-05) — deepmind.google
21. Anthropic, "Claude for Life Sciences" (2025-10) — anthropic.com
22. GPTZero Benchmark — gptzero.me
23. Retraction Watch — retractionwatch.com
24. PMLR Volume 296 (ICBINB Workshop Proceedings)
25. SakanaAI GitHub — github.com/SakanaAI
26. AI Scientist ICLR 2025 Workshop Experiment — github.com/SakanaAI