+- MKLab (https://mklab.gr)
+-- Forum: [INDEX] (https://mklab.gr/forumdisplay.php?fid=1)
+--- Forum: ARTFICIAL INTELLIGENCE (AI) (https://mklab.gr/forumdisplay.php?fid=5)
+---- Forum: ARTICLES (https://mklab.gr/forumdisplay.php?fid=33)
+---- Thread: AI in Science: Early Insights (/showthread.php?tid=2025)
AI in Science: Early Insights - mklabgr - 09-16-2026
AI in Science — 8 Key Insights
The paper “AI in Science: Early Insights” examines how AI is already changing scientific research, using 15 million Gemini interactions, more than 2,600 specialized scientific AI models, and a survey of 637 scientists. Its central message is that AI is producing meaningful productivity gains, but the main constraint is increasingly shifting from thinking and computing to testing, validating, and physically carrying out scientific work.
1. Scientists are among the heaviest users of AI
AI adoption in science is already widespread. Scientific occupations are substantially overrepresented in AI usage, and nearly half of surveyed scientists use AI daily. In the Gemini data, life/physical/social science occupations used AI at about 2.7× their share of U.S. employment. Adoption extends across almost all scientific disciplines.
2. General LLMs and specialized scientific AI are complements
The paper does not find that general-purpose LLMs are simply replacing specialized scientific models. Instead, they perform different jobs:
- LLMs: coding, literature review, statistical analysis, writing, troubleshooting and general research assistance.
- Specialized models: protein prediction, disease prediction, molecular engineering, simulations, classification and domain-specific generation.
3. AI is saving scientists substantial time
About three-quarters of researchers report time savings, averaging almost 7 hours per week. Most of that saved time is redirected toward additional research rather than simply reducing working time. Around 84% report increased professional/laboratory output over the previous three years, while about 90% expect output to increase over the following three years.
4. AI is broadening scientific thinking
The benefits aren't limited to speed. About 68% of scientists say AI gives them greater access to insights from other disciplines, while roughly 65% report broader research agendas. This suggests that LLMs can act as a kind of intellectual translator, helping researchers access knowledge outside their specialization.
5. AI is creating a new bottleneck: experiments and validation
This is perhaps the paper's most important insight. AI can accelerate computational analysis and generate hypotheses much faster, but physical experiments, clinical validation and data collection cannot necessarily accelerate at the same rate.
About 44% of scientists report that their main bottleneck has shifted downstream, while only about 14% report an upstream shift. Consequently, 41% report a growing backlog of untested hypotheses.
6. Verification consumes part of the productivity dividend
AI-generated scientific output cannot simply be accepted without checking it. Among scientists who save time with AI, 89% spend more than 10% of that saved time verifying AI outputs, while 46% spend more than 25% of it on auditing, debugging or fact-checking.
So the real productivity equation is closer to:
AI generation → verification → experiment → validation
rather than simply:
AI generation → scientific result.
7. AI may encourage safer, incremental science
There is a potentially important downside. 49% of surveyed scientists say AI pushes them toward safer, incremental questions, compared with 28% who say it enables riskier questions. AI works particularly well where there are established benchmarks, clean data and known evaluation criteria.
The authors therefore raise the possibility of a “Streetlight Effect”: AI could make already-easy, incremental research much cheaper while leaving difficult, uncertain research relatively expensive.
8. The future may be an AI-orchestrated scientific workflow
The paper points toward a future in which a general AI system acts as an orchestrator: it could formulate research plans, call specialized models such as protein- or materials-prediction systems, interpret their results, and potentially coordinate experiments.
But achieving this requires investment in the bottlenecks—physical experimentation, validation, verification tools, data and benchmarks. The paper also warns that specialized AI development is concentrated in a relatively small group of wealthy economies, potentially widening existing research-capacity differences.
Bottom line.
The paper's picture is more nuanced than “AI will automate scientific discovery.” AI is already making scientists faster and appears to be increasing research output and interdisciplinary access. But accelerating the computational parts of science can simply move the queue to the physical parts: experiments, validation and verification. The decisive question is therefore not only how powerful AI becomes, but whether scientific infrastructure can scale fast enough to exploit what AI produces.
ARTICLE [PDF]