In today’s rapidly evolving technological landscape, the boundaries of science are being redrawn—not only in what we can discover, but in how we discover it. At one of the sessions at the Collective Learning Workshop held from June 30 to July 2, 2025, at the Complexity Science Hub Vienna, researchers from across the globe gathered to explore a timely question: How are AI tools—particularly large language models (LLMs)—reshaping the norms and practices of science itself?
With attendees from institutions including MIT, Yale, Carnegie Mellon, the Santa Fe Institute, and the London School of Economics, the workshop aimed to surface collective insights, map ethical risks, and imagine new models of collaboration between humans and machines. What emerged was a vibrant blend of optimism, caution, and intellectual provocation.
Perhaps the most consistent theme to arise was that AI is transforming the role of the scientist. “The role of scientists will shift towards setting problems as AI handles the experiments,” noted one participant. In other words, the act of "doing science" may increasingly mean asking the right questions, while AI handles literature reviews, experimental design, and initial analyses insights.
LLMs were seen as particularly powerful in enhancing research independence—allowing individuals or small teams to do the work of many, and accelerating the journey from idea to insight. “LLMs can facilitate interdisciplinary communication and coding,” said another participant, pointing to their usefulness in reducing the friction of jargon across fields insights.
But this newfound speed introduces a tension: Does velocity risk flattening nuance?
Science is not just about method and data—it's also about voice, creativity, and intellectual style. Several participants expressed concern that as AI-generated text becomes ubiquitous, it may standardize scientific expression to a point where originality is hard to recognize.
“Standing out with unique style will be more important than ever despite AI use,” remarked one attendee. Another posed a developmental question: “How will budding scientists develop writing and thinking skills if models do the heavy lifting?”
There’s a deeper worry here: that scientific training, long based on mentorship and slow cognitive maturation, might be eroded by shortcuts, leaving behind a generation proficient in prompting but lacking in foundational reasoning insights.
Beyond aesthetics and workflow lies a stark ethical terrain. Participants voiced concerns over LLMs hallucinating results, generating misleading citations, or producing plausible-sounding but false conclusions. In scientific settings, such mistakes aren't trivial—they could endanger lives or undermine trust.
“There should be stringent testing and formal proofs before deploying safety-critical algorithms,” urged one voice. Others suggested importing standards from nuclear safety or aviation, where systemic risk is treated with commensurate regulatory rigor insights.
Moreover, new grey zones are emerging: What happens when LLMs become quasi-participants in experiments, subtly influencing outcomes? How do we prevent models from being manipulated to produce biased or even dangerous results?
The consensus was clear: We need strong, adaptable frameworks for both the responsible use and governance of AI in science.
While no one at the workshop argued that AI should—or could—replace scientists wholesale, many agreed that the future lies in co-evolution, not competition. One shared insight noted that human-AI collaboration frameworks will be central to scientific practice going forward insights.
At the workshop, participants experimented with AI-guided discussions, where group summaries were seeded by LLM-generated syntheses of individual reflections. These sessions sparked lively debates—not just about the topic at hand, but about how AI shapes what is foregrounded, what gets left out, and how collective reasoning unfolds.
Some groups proposed developing new protocols for transparency and interpretability in human-AI co-authorship. Others began sketching outlines for training curricula that incorporate AI literacy alongside traditional scientific skills.
Rather than treating AI as a one-time disruption, participants saw it as a call to reimagine the scientific enterprise itself—from publication norms and peer review to funding models and educational pathways.
Concrete ideas included:
These are not hypothetical futures—they are infrastructure-level questions that demand attention now.
While views diverged on pace and priority, one sentiment echoed across the workshop: AI forces us to confront what we value most about science. Is it speed, scale, precision? Or is it creativity, skepticism, community?
“AI may help us ask better scientific questions, but it also demands better answers—about responsibility, style, and what kind of science we want to do,” said one participant. It’s a fitting summary—and a challenge.
