Report: Detachment of scientific conclusions from human authorship or oversight, especially with generative models, raises questions about accountability, originality, and reproducibility.

AI in Scientific Research: Promise, Perils and Policy Priorities in the EU

by ChatGPT-4o

Artificial Intelligence (AI) is no longer just a supporting tool for science—it is fast becoming a co-pilot in the scientific process, transforming everything from hypothesis formulation and experiment design to data analysis, writing, and community-building. The Joint Research Centre (JRC) of the European Commission presents a comprehensive, multidisciplinary report mapping out how AI is reshaping the scientific landscape, offering both unprecedented opportunities and raising profound concerns about epistemic integrity, bias, and the future of knowledge itself.

1. Transformative Role of AI Across the Scientific Process

The report outlines how AI now plays an active role in all eight core stages of scientific research: from asking questions and conducting literature reviews to designing experiments, analysing results, publishing findings, and building research communities. For instance:

• Large Language Models (LLMs) are used to mine massive corpora of scientific literature to identify research gaps or generate novel hypotheses.

• AI-driven experiment design enables the automation of lab workflows—so-called “self-driving labs”—and simulation of conditions too complex for human design alone.

• AI-enhanced data analysis uncovers patterns in multimodal data (e.g., genomics, satellite images) that human researchers might never detect.

• Generative AI is increasingly used in writing and visualising scientific findings—helping with translation, summarisation, and accessibility.

Deep dives into protein structure prediction (e.g., AlphaFold), materials discovery, and computational archaeology serve as compelling case studies of AI’s potential to revolutionise specific disciplines.

2. Epistemic Drift and the Risk of Hallucinated Science

Perhaps the most striking concept in the report is the idea of “epistemic drift”—a term used to describe how AI may subtly or radically change what counts as scientific knowledge. This happens in two ways:

• Over-reliance on AI may lead to the narrowing of scientific questions explored, favouring well-documented areas where AI performs better.

• Detachment of scientific conclusions from human authorship or oversight, especially with generative models, raises questions about accountability, originality, and reproducibility.

Coupled with this is the very real danger of hallucinated outputs—fabricated but plausible-sounding information generated by LLMs, which, if not caught, could distort the scientific record.

3. Uneven AI Uptake and the Need for Skills, Governance, and Infrastructure

The report makes clear that AI adoption across disciplines remains uneven, reflecting differences in data availability, compute resources, and technical skills. To address this, it stresses the need for:

• Investment in High-Performance Computing (HPC), “AI Factories,” and open scientific data repositories to democratise access.

• Hybrid research teams with both domain knowledge and AI fluency to ensure meaningful human-AI collaboration.

• Open science practices—open data, models, infrastructure, and publications—to preserve scientific integrity and reproducibility in the age of AI.

4. Europe’s Relative Strength and Vulnerabilities in AI Research

From a geopolitical standpoint, the report offers a sobering look at Europe’s AI competitiveness:

• Europe leads in scientific output, especially academic research.

• But it lags far behind in patent applications (3%) and venture capital investments (7%), with China dominating AI patents (76%) and the US leading in VC funding (53%).

• Despite this, EU-funded programs like Horizon Europe significantly strengthen the AI ecosystem, especially in smaller member states like Greece and Slovenia.

This imbalance means the EU is strong in fundamental science but vulnerable in terms of tech transfer, startup growth, and foreign ownership of AI firms (49% of foreign-owned EU AI firms are US-controlled).

5. Key Recommendations

For EU Policymakers:

• Establish robust AI governance frameworks for science, complementing the EU AI Act with specific guidance for research applications.

• Invest in public AI infrastructure—including open models, open datasets, and compute capacity tailored for research.

• Support interdisciplinary training and career pathways that blend AI skills with deep scientific expertise.

For Research Institutions and Funders:

• Promote AI literacy and critical thinking to counteract over-dependence on “black box” models.

• Build cross-sector collaborations, especially between academia, industry, and public research facilities.

• Require transparent documentation of AI-assisted research methods to ensure traceability and reproducibility.

For Scientific Publishers and Peer Review Bodies:

• Update authorship and plagiarism policies to account for generative AI contributions.

• Encourage disclosure of AI tools used in study design, writing, or data analysis.

• Develop AI-enhanced peer-review support systems that can assist—but not replace—human judgment.

For Industry and AI Developers:

• Engage with the scientific community to build domain-specific AI tools that respect open science principles.

• Prioritise explainability and auditability in tools deployed in research contexts.

• Avoid extractive practices that exploit open-access publications without reinvestment or accountability.

Conclusion: Embracing AI Without Losing the Soul of Science

This report does not offer a binary view of AI in science as good or bad—it recognises it as both a tremendous opportunity and a source of existential risk to the integrity of the scientific enterprise. The challenge ahead for Europe is to leverage its strengths in academic excellence and ethical governance, while closing the gap in innovation funding and startup growth.

The EU’s vision must not just be about using AI to accelerate science, but about ensuring that science—rigorous, inclusive, and transparent—shapes the future of AI.