“"Artificial scientists" have sprung up like mushrooms in recent years, yet the dream of an automated researcher capable of executing every step—from ideation to the final publication—on its own has not materialized. Artificial intelligence can handle individual tasks such as literature reviews or data analysis, but hypothesis generation and interpretation remain the domain of humans. Even the advent of language models has not fundamentally changed this. While Large Language Models (LLMs) can produce impressive results, they often stumble over the simplest tasks because they lack an inherent understanding of the research process.
At a well-attended international machine learning conference, a fully AI-generated paper has now successfully passed the peer-review process. With the organizers' knowledge, three machine-generated papers from the Japanese company "AI Scientist" were submitted; one of them passed the review process within a workshop track.
Human reviewers were unable to detect the papers' machine origins.
The paper describes remaining obstacles for AI in the field of machine learning—almost as if warning against inflated expectations regarding its own capabilities. Its developers promote the promise of a fully automated research process. From ideation, programming, and experimental testing to publication—including self-evaluation—the "AI Scientist" is designed to handle everything independently, just as it did with the submitted research paper.
Reviewers rated the machine-generated output as having moderate but sufficient quality. It did not clear the hurdle for the conference's main proceedings. While clearly not an outstanding paper, it was nonetheless of a quality that could potentially be published in a scientific journal. Will the scientific publishing market soon be flooded with AI-generated papers that distort academic competition and render publication useless as an indicator of reputation? And what role would be left for scientists? A paper published in *Nature* in the wake of this successful peer review offers a measured perspective on these questions (Chris Lu et al.: Towards end-to-end automation of AI research. Nature, Volume 651, March 2026). Ultimately, only one out of three submitted papers passed peer review. A flawless AI-generated research output is evidently the exception rather than the rule. In the authors' experience, AI-generated papers continue to suffer from numerous weaknesses: they posit naive hypotheses, make errors in experimental execution, arbitrarily duplicate figures and images, or blithely hallucinate data. Overall, they lack methodological rigor.
Furthermore, the artificial researcher remains limited for now to specific fields—such as machine learning—where the research process takes place entirely on a computer. However, concepts for automated chemical laboratories, where experiments are conducted under AI guidance, already exist. *Nature* recently reported remarkable AI-driven advances in drug development.
Based on comprehensive literature analysis, an AI proposed unusual compounds for hard-to-treat diseases that proved to be astonishingly effective.
This could vastly accelerate research and development, especially as AI makes huge strides in eliminating sources of error. Nevertheless, there is no guarantee that the tendency to hallucinate can be entirely eradicated; one of the proposed drug candidates would have had fatal consequences. It also remains unclear whether AI can truly arrive at original insights or if it will remain confined to pattern recognition within a strictly defined scope—though even that would yield immense progress.
If AI-generated papers were to become mass-produced commodities, it would be difficult to sift original research contributions from the mountain of "instant" output. There is already no shortage of dubious business models profiting from this trend. It is possible to generate papers that appear flawless yet rely on false data—a tempting prospect for scientists seeking a quick boost to their job prospects. Fittingly, the European research council has just reiterated that while AI may assist in the preparation of publications and expert reports, it must not take over either task entirely. Whether this can be prevented in practice remains an open question. AI researchers have a vested interest in viewing the collaboration with artificial intelligence as a relationship of assistance. The logical consequence of a fully automated research process would be to hand over publications, awards, and patents to machines—and eventually, professorships as well. The displaced scientists would then have plenty of time to pursue other activities. But which ones?” [1]
1. Die Vision einer voll automatisierten Forschungswelt: Eine vollständig KI-generierte Publikation besteht das Peer-Review-Verfahren: Welche Folgen hat das für den wissenschaftlichen Publikationsmarkt? Frankfurter Allgemeine Zeitung; Frankfurt. 08 Apr 2026: N4. THOMAS THIEL
Komentarų nėra:
Rašyti komentarą