The Atrophy Risk: What AI Could Cost Science Advice

DOI: https://doi.org/10.25453/plabs.32781210

Published on June 24th, 2026

When we published the Science for Policy Handbook at the Joint Research Centre in 2020, artificial intelligence appeared in the chapter on big data and the modernisation of the policy cycle. It was a future opportunity. Five years later, AI is operational in the evidence-to-decision chain across Europe and beyond, from weather forecasting systems feeding flood warnings to machine-learning classifiers accelerating systematic reviews at Cochrane. The question for the science advice community is no longer whether AI will arrive, but whether we are paying attention to what it may be quietly taking away.

The Handbook described the evolution from Science for Policy 1.0 (the deficit model, where scientists supply knowledge to fill policymakers’ gaps) to Science for Policy 2.0, built on co-creation. AI is now creating a third configuration, which I have been calling Science for Policy 3.0. But its character is not yet settled. It could genuinely enhance the science-policy interface. Or it could hollow it out.

Where AI is working well, the strongest implementations share a common design: human oversight, rigorous validation, and explicit acknowledgement that AI outputs are decision support, not decision-making. The European Centre for Medium-Range Weather Forecasts’ (ECMWF) AI forecasting system, operational since February 2025 and now feeding Copernicus flood warnings, runs side by side with traditional models and is continuously verified against observations. The World Health Organization’s (WHO) Epidemic Intelligence from Open Sources platform applies machine learning to hundreds of thousands of daily items, but every source undergoes expert-led validation before integration. Cochrane’s randomised controlled trial classifier is deliberately calibrated to prioritise recall, a governance choice to minimise the risk of missing evidence.

These cases work because AI operates in the high-volume, low-judgment segments of the pipeline: screening, triage, extraction, forecast generation. The feedback loops are tight and the governance infrastructure matches the ambition.

The danger lies elsewhere, in the judgment-intensive segments where evidence meets politics, where values compete, where context determines meaning.

Consider the United Kingdom. The government’s Incubator for AI developed tools for civil servants, including Redbox for document summarisation, Parlex for parliamentary analysis, and Lex for legislation search, serving over 5,300 officials. A trial of 20,000 civil servants showed time savings of nearly two weeks per person per year. But the Institute for Government identified the tension: these tools could arm ministers with superficial talking points “to which they become overly attached,” making it harder for officials to give nuanced advice. The Redbox team found that 70 per cent of interactions were users simply chatting with a large language model for routine tasks, not the complex analysis it was designed for. Researchers at Durham University argued that the prompt itself constitutes a new form of political technology, restructuring how government knowledge is produced.

In the United States, the Department of Government Efficiency went further, using AI to determine whether federal employees’ work was “mission-critical.” An AI system hallucinated contract values at the Veterans Administration, nuclear security staff were terminated before officials realised the positions were essential, and nearly 200,000 federal employees left government.

I call this the atrophy risk. In the augmentation phase, AI enters as a productivity tool and satisfaction is genuine: Redbox users reported that 80 per cent believed it improved the quality of their work and 89 per cent said it saved them time. In the substitution phase, efficiency gains create pressure to reduce headcount, and the humans who knew how to do the work are let go. In the dependency phase, institutional capacity to function without AI degrades; tacit knowledge of which minister responds to which framing, which evidence is contextually relevant, which stakeholders will block implementation, walks out the door. In the vulnerability phase, when AI produces errors, there are too few people left to catch them. The system has lost its error-correction capacity.

This pattern is already visible in aviation and financial services. What makes the science-policy case distinctive is the nature of what atrophies. The skills that the Handbook identified as essential, understanding policy context, building trust, navigating interests, translating between scientific and political reasoning, are deeply experiential. They develop through friction, through sitting with a sceptical policymaker and learning to read the situation. If an AI drafts your policy brief, simulates your stakeholder mapping, and pre-generates your policy options, you may never develop those muscles at all.

For Europe, this raises particular questions. Institutions like the Joint Research Centre, the European Food Safety Authority, and the European Environment Agency were built over decades on the principle that independent scientific assessment should inform public policy. They represent accumulated human capital that is expensive to build and easy to lose. If budget pressures and the illusion of AI equivalence lead to workforce reductions, we may dismantle capabilities that cannot be recreated once the limitations of current AI become apparent.

The Cochrane, Campbell Collaboration, JBI (formerly Joanna Briggs Institute), and Collaboration for Environmental Evidence joint position statement of 2025, insisting on human oversight and transparent reporting of AI use in evidence synthesis, points in the right direction. But even in their ecosystem, a review of over 2,000 evidence syntheses found AI adoption outpacing quality control. And nobody is systematically measuring what human capabilities are being lost.

Can we develop institutional “capability audits,” analogous to pilots maintaining manual flying currency, to ensure advisory staff retain core judgment skills? Should we distinguish formally between pipeline segments where AI substitution is safe and those where it threatens the interpretive capacity on which good advice depends? And how do we ensure that the next generation of science advisors develops the experiential judgment that no training dataset can provide?

Practical answers are beginning to take shape.

Capability audits should start where advisory failure is most costly: public health, climate, and urban systems. Bodies such as the European Centre for Disease Prevention and Control, the European Environment Agency, and the technical units advising on infrastructure decisions are well placed to lead. The instrument is simple: periodic exercises in which advisors produce a briefing, options paper, or risk assessment without AI assistance, assessed by senior peers against the same standards as their AI-assisted work. Sectoral regulators and chief scientific advisers' offices have the standing to make such reviews routine.

Concrete cases clarify the line between safe and unsafe substitution. ECMWF's AI forecasting is safe: the output space is narrow, ground truth arrives daily in observations, and verification infrastructure already exists. Drafting the executive summary or options section of a ministerial brief is not. There is no observational ground truth, the framing shapes what becomes politically thinkable, and the choice of words turns on knowing how a particular minister reads a particular argument.

Building experiential judgment in the next generation means protecting the tasks AI performs most efficiently. If junior advisors never produce a stakeholder map, draft a contested options paper, or sit through the feedback that turns a poor brief into a usable one, they will not develop the capacity those activities build. The methods are familiar: apprenticeship pairings, rotation across ministries and scientific institutions, post-decision retrospectives in which trainees must articulate why a final decision diverged from the formal evidence. What is new is that they have become indispensable.

Underlying all three is the governance of the evidence infrastructure itself. Institutions deploying AI in advisory work should maintain registers of which models are in use, on what data they were trained, what assumptions they encode, and where dependencies on external vendors create strategic exposure. Every AI-assisted brief should leave a retrievable audit trail, so errors can be traced and patterns of failure recognised.

We are in a critical window. The decisions being made now about advisory workforce structures, institutional investment, and governance requirements for AI in policy processes will determine whether Science for Policy 3.0 is a genuine enhancement or a hollowed-out imitation. The science-policy community has built something valuable over the past decades. It would be a bitter irony if, in the pursuit of efficiency, we consumed the very capabilities on which effective governance depends.

References

1. V. Šucha and M. Sienkiewicz (eds.), Science for Policy Handbook, Elsevier, 2020. https://doi.org/10.1016/C2018-0-03963-8

2. V. Šucha, “Before AGI Arrives: Why 2025–2030 Determines the Future of Democratic AI Governance,” EUI STG Policy Brief, 2025. https://data.europa.eu/doi/10.2870/7885333

3. V. Šucha, “Preserving Human Agency in the Age of AI,” EUI STG Policy Brief 2025/12. https://hdl.handle.net/1814/93015

4. L. Amoore et al., “Politics of the prompt: Government in the age of generative AI,” Economy and Society, 2025. https://doi.org/10.1080/03085147.2025.2560177

5. Institute for Government, Policy Making in the Era of Artificial Intelligence, 2025. https://www.instituteforgovernment.org.uk/publication/policy-making-era-artificial-intelligence

6. E. Flemyng et al., “Position statement on AI use in evidence synthesis,” Environmental Evidence, 14(20), 2025. https://doi.org/10.1186/s13750-025-00374-5

7. UK Incubator for AI, Redbox operational documentation, 2025. https://ai.gov.uk/our-work/government/

8. “AI tools as science policy advisers?” Nature, 27 September 2023. https://www.nature.com/articles/d41586-023-02999-3

Copyright: © 2026 [author(s)]. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in Frontiers Policy Labs is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.