Are AI Scientists Set to Reimagine Science? 5 Realities Researchers Must Know

In March 2026, Sakana AI's "The AI Scientist" became the first autonomous system to have its work published in Nature—a milestone that would have seemed implausible even two years earlier. The system generates hypotheses, designs experiments, runs them, writes complete manuscripts, and even reviews its own output, all for roughly $15 per paper. With over 50% of researchers already using AI during peer review, the question is no longer whether AI will reshape research, but how fast—and whether existing safeguards can keep up. This article unpacks what autonomous AI research systems can actually do today, where they fall short, and what researchers need to do differently as a result.

From Lab Assistant to Lab Partner

For decades, AI's role in research was narrowly defined: automating statistical analysis, flagging plagiarism, or polishing grammar. Researchers controlled the intellectual process end to end. That model began shifting when DeepMind's AlphaFold 2 demonstrated that AI could solve problems—predicting 3D protein structures from amino acid sequences—that had resisted human efforts for decades, generating scientific hypotheses at a scale no human team could match.

What systems like The AI Scientist represent, however, is a further leap. Rather than excelling at a single task, these systems attempt to replicate the entire research lifecycle: surveying literature, generating ideas, planning and iterating on experiments, writing manuscripts, and conducting automated peer review. An independent evaluation by Beel et al. (2025) confirmed that the system produces complete research manuscripts at a cost of $6–$15 with approximately 3.5 hours of human involvement—far faster and cheaper than traditional research timelines.

This is not a marginal productivity improvement. It is a structural shift in how knowledge can be produced.

The $15 Paper That Passed Peer Review

The AI Scientist didn't just generate papers in isolation. In a controlled experiment with IRB approval and cooperation from the ICLR 2025 conference organizers, three fully AI-generated manuscripts were submitted to the "I Can't Believe It's Not Better" workshop. One manuscript scored 6.33 out of 10 and exceeded the average acceptance threshold.

The paper was withdrawn before publication, as pre-agreed with organizers. But the result demonstrated something significant: an autonomous system had produced work that human reviewers, evaluating blindly, considered publishable.

That said, context matters. None of the three papers met the standard for the main ICLR conference (acceptance rate: 32. The system also required human filtering to select the most promising outputs before submission. These are meaningful caveats, but they do not diminish the trajectory this result signals.

What Editors are Already Seeing

These developments are not lost on the editorial community. At a recent Enago roundtable in London, editors working across disciplines underscored a consistent concern: AI produces outputs that are superficially polished but lack the intellectual depth that defines real research. As one editor from Imperial College London put it, AI takes existing information and recombines it—but it cannot penetrate the author's mind or generate genuinely new ideas. Editors described their role as far more than language correction: they serve as the last quality checkpoint for logical consistency and subject-matter accuracy, functions that no autonomous system currently replicates.

This ground-level editorial perspective aligns closely with what the research literature now confirms.

5 Realities Researchers Must Know

As AI systems move from assisting discrete tasks to influencing entire research workflows, researchers need a practical framework for engaging with them. The following five realities reflect where the technology stands today—and what that demands of the people who use it.

1. AI-generated ideas are novel but fragile in execution

A landmark study by Si, Yang, and Hashimoto (ICLR 2025) compared research ideas generated by an LLM agent against those written by over 100 expert NLP researchers. In blind reviews, LLM-generated ideas were rated significantly more novel (p < 0.05) than human-written ones. However, they scored lower on feasibility. More critically, a follow-up execution study found that when these ideas were actually implemented, LLM-generated ideas deteriorated significantly more than human ideas across novelty, excitement, effectiveness, and overall quality.

The takeaway is clear: AI can generate promising starting points, but those ideas require rigorous human evaluation before committing resources to execution. Treat AI outputs as hypotheses, not conclusions.

2. Manuscript quality remains far below expert standards

The same independent evaluation that confirmed the AI Scientist's cost efficiency also assessed the quality of its output. Beel et al. (2025) described the manuscripts as comparable to the work of "an unmotivated undergraduate student rushing to meet a deadline." The system struggles with justifying design decisions, demonstrating genuine domain expertise, and producing high-impact hypotheses.

Researchers using AI-generated drafts as starting points must invest substantial effort in verification, restructuring, and critical evaluation. Skipping this step risks propagating superficially competent but substantively weak research into the literature.

3. Disclosure obligations now extend beyond writing

Major publishers—including Nature Portfolio, Elsevier, and Wiley—now require disclosure of AI use in manuscript preparation. But as AI's role expands into idea generation, experimental design, and data analysis, disclosure limited to "AI-assisted writing" no longer captures the full picture.

COPE's updated guidelines are unequivocal: AI cannot be listed as an author, and human researchers bear full responsibility for all content. Researchers who use AI at any stage of the research process should document that use explicitly and be prepared to defend every claim, citation, and methodological choice as their own.

4. Verification is non-negotiable—and more difficult than it appears

Autonomous research systems inherit the well-documented limitations of large language models: hallucinated references, fabricated statistics, internally inconsistent reasoning, and overconfident claims unsupported by evidence. These problems are harder to catch in the context of a polished, well-structured manuscript because the surface quality masks deeper flaws.

Researchers must cross-check every numerical result, verify every cited source, and independently validate experimental outputs. This is especially critical in high-stakes domains such as clinical research or pharmacology, where a single fabricated reference can have downstream consequences.

5. The researcher's role is evolving, not disappearing

AI systems are unlikely to replace researchers. But they are already changing what "doing research" means. As autonomous systems handle more of the generative and iterative work—literature scanning, hypothesis exploration, draft writing—researchers increasingly function as validators, supervisors, and ethical gatekeepers.

This shift requires new competencies: the ability to critically evaluate AI-generated outputs, design verification protocols, and make judgment calls about when AI assistance is appropriate and when it introduces unacceptable risk. These are not skills most researchers were trained for, and the gap between AI capability and researcher readiness is widening.

What comes next

The emergence of autonomous research systems does not mean the end of human-led science. It means the standards for human oversight must rise to match the scale and speed at which AI can now produce research outputs. Researchers who learn to evaluate, verify, and govern AI-generated work effectively will be better positioned to use these systems as genuine force multipliers—accelerating discovery without compromising integrity.

For researchers already integrating AI into their workflows, working with subject-specialist editors who can evaluate scientific reasoning, logical consistency, and methodological rigor offers an immediate layer of protection—particularly when AI-generated content has influenced any stage of the manuscript.

At the institutional level, Enago's Responsible AI Movement has already helped over 3,000 researchers build practical skills for navigating AI in the research lifecycle—through workshops, webinars, and training sessions developed in partnership with publishers and universities. As autonomous systems move from experimental milestones to everyday tools, this kind of structured preparation is no longer optional.

AI scientists are here. The researchers and institutions that invest in understanding them now will be the ones best positioned to use them responsibly.