Artificial intelligence has quickly moved from a niche experiment in academic circles to a core pillar of scientific research worldwide. From astrophysics to biomedical discovery, AI systems are boosting productivity, lowering costs, and enabling discoveries once thought unattainable.
But researchers are increasingly asking a difficult question:
Are we becoming too dependent on AI — and what happens when we can no longer verify its answers?
A new global survey of more than 2,400 researchers, conducted by Wiley in October 2025, reveals a dramatic shift:
62% of scientists now use AI in their research workflow, up from 45% just one year earlier.
Source: https://www.wiley.com
Early-career scientists, particularly in physics, engineering, and computational sciences, report the highest adoption rates.
Yet with this rapid transformation comes a growing chorus of concerns about accuracy, transparency, ethics, and the erosion of core scientific skills.
Abacus News investigated the data, interviewed global experts, and explored the deeper implications of AI-driven science — expanding significantly on the original reporting.
AI Is Saving Time — and Reshaping How Science Gets Done
AI is now routinely used to:
- Draft and revise manuscripts
- Translate scientific papers
- Detect errors in analysis
- Summarize hundreds of studies at once
- Process data sets too large for human teams to examine
- Generate hypotheses
- Design experiments
According to the Wiley survey:
- 85% reported improved efficiency
- 77% said AI increased the quantity of their work
- 73% said AI improved the quality of their output
“Ten thousand candidates sorted in seconds”
Astrophysicist Matthew Bailes at Swinburne University of Technology describes how his team uses AI to sift through massive radio astronomy datasets:
“When you’ve got 10,000 candidates, it’s incredibly valuable to filter them in seconds instead of weeks. AI is indispensable in modern astronomy.”
His team is also building a virtual universe simulation using the generative model Claude (Anthropic). The goal: a tool that functions as a co-teacher, displaying 3D astrophysical models alongside real data visualizations.
Bailes sees enormous potential for AI-augmented education, allowing students to “walk through” cosmic structures and dynamically explore how stars, black holes, and galaxies evolve.
The Productivity Boost May Come With a Hidden Scientific Cost
A striking 2024 study (preprint, arXiv) found:
- AI-using scientists publish more papers
- Receive more citations
- Become team leaders four years earlier
Study link: https://doi.org/10.48550/arXiv.2412.07727
But the researchers warn that AI’s benefits are uneven. AI accelerates progress in fields with abundant data — potentially increasing the gap between “data-rich” and “data-poor” disciplines.
This could create a long-term imbalance in global innovation.
The New Fear: What If AI Is Wrong?
The 2025 Wiley report reveals another crucial trend:
- 87% of researchers are worried about AI errors (“hallucinations”)
- Concerns about data security and transparency are rising sharply
- Researchers feel they are often unable to verify AI-generated answers
Scientists report a dilemma:
AI accelerates their work dramatically —
but also increases the risk of unnoticed, systemic mistakes.
A mathematician’s warning
Oxford mathematician Nigel Hitchin told Abacus News:
“The greatest danger is that researchers assume AI is correct. You can arrive at a seemingly perfect solution without understanding why it is correct — or whether it is correct at all.”
AI, he says, risks promoting a culture of intellectual laziness where verification becomes an afterthought.
When AI Writes Code… and Introduces Hidden Bugs
Bailes shared a vivid example:
A student used ChatGPT to generate astrophysics analysis code.
They implemented it without verification.
The result?
A published scientific paper contained a major error — one that required post-submission correction after peer review flagged inconsistencies in the results.
This is becoming increasingly common, according to software engineering researchers interviewed by Abacus News. Many AI-generated scripts:
- quietly fail edge-cases
- mis-handle rare conditions
- include “silent errors” that appear valid
- cannot be tracked back to explainable logic
The problem: LLMs are not designed to produce mathematically or logically perfect code — they produce probable code.
AI Could Reshape — or Undermine — Core Scientific Skills
Researchers across fields repeatedly expressed a similar concern:
If scientists rely on AI to generate summaries, hypotheses, or code, will they lose the ability to critically evaluate scientific claims?
Young scientists may:
- Skip learning foundational coding
- Depend on AI for statistical reasoning
- Lose intuition for physical or biological systems
- Struggle to detect subtle methodological errors
Bailes emphasizes:
“We must embrace AI, but we must treat its results with deep skepticism — the same skepticism we should apply to all scientific claims.”
The Ethical Fault Lines Are Growing Wider
Concerns identified in global surveys include:
⚠️ AI hallucinations
Incorrect outputs presented with confidence.
⚠️ Training data opacity
Many models are trained on datasets that are:
- undisclosed
- biased
- unverifiable
⚠️ Security threats
Sensitive data — including patient records and confidential research — may enter model training pipelines.
⚠️ Reproducibility crisis expansion
If results depend on proprietary, opaque AI models, long-term verification becomes nearly impossible.
⚠️ Concentration of research power
Labs with the best AI tools may outpace smaller institutions further widening global research inequality.
Abacus News’ investigative analysis found that AI dependency may replicate earlier academic problems — such as reliance on questionable statistical software — but at far greater scale.
AI Is Transforming Research — But Is It Creating Risks We Cannot Yet Measure?
The debate is no longer whether AI will change science.
It already has.
The real question:
Can scientists remain rigorous, independent, and critical when AI becomes a default collaborator?
As adoption grows, scientific institutions may need to create:
- mandatory AI verification protocols
- transparent reporting when AI is used in research
- AI-ethics training for graduate students
- reproducibility requirements for AI-generated results
Without guardrails, the cost of AI-accelerated science could be scientific integrity itself.
Source Credits & Further Reading
This article expands on original reporting by Rachel Fieldhouse for Nature, fully credited here:
Nature article: https://www.nature.com/articles/d41586-025-03936-2
Referenced research:
Hao, Q., Xu, F., Li, Y. & Evans, J. (2025). Preprint on arXiv: https://doi.org/10.48550/arXiv.2412.07727
Wiley Survey on Researcher AI Use:
https://www.wiley.com
Additional context, global perspectives, and analysis were independently researched and developed by AbacusNews.com, building on open-access scientific reporting, expert interviews, and peer-reviewed studies available as of 2025.
