Cohere Research: What AI exposure scores actually measure — and why that is not enough for policymakers

Four researchers from Cohere Research — Campbell Lund, Thomas Euyang, Zanele Munyikwa, and Marzieh Fadaee — published on June 10, 2026 a comprehensive critical analysis of one of the most influential tools for assessing AI’s impact on the labor market: so-called AI exposure scores (AES). The paper is accompanied by a companion preprint on arXiv.

What are AI exposure scores?

Exposure scores are the fundamental tool in the debate about AI’s impact on the labor market. The starting point is the “GPTs are GPTs” framework (Eloundou et al., 2023), which defines occupational exposure as the percentage of tasks within a given occupation that a large language model (LLM) can perform or assist with. The higher the percentage, the more “exposed” the occupation is considered — whether to automation or to AI-assisted work.

Such scores quickly became a reference point for policymakers, economists, and media outlets trying to predict which sectors AI will change and to what extent.

Limitations of static scores for shaping policy

The authors identify two fundamental gaps between research methodology and the actual needs of policymakers.

Structural gap: Static scores measure the performance of a specific AI system, applied to a specific occupational taxonomy, at a specific point in time. This means they are inherently limited along three dimensions: temporal (what is true today may be outdated tomorrow), geographic (occupational taxonomies differ between countries), and ontological (the very definition of “tasks” within an occupation is contested and unstable). Policymakers, however, need measures that are flexible along all three dimensions.

Coordination gap: The research community has developed more sophisticated methodologies that address precisely those limitations — but these methodologies remain trapped within academic circles. Meanwhile, policymakers continue citing outdated static scores because they are unaware of newer approaches. The result is a systematic disconnect between what is methodologically available and what is applied in practice.

Five new methodological families

The authors catalog five new methodological directions that the academic community is developing but that have not yet entered standard policy use:

1. Dynamic and benchmark-based measures — rather than a single static assessment, tracking how exposure changes over time as AI systems advance.
2. Ensemble methods — combining multiple different models and assessment approaches to reduce bias from any single system.
3. Task-framework extensions — better mapping of what workers actually do, rather than relying on standardized occupational classifications.
4. Worker-centered metrics — measuring AI’s impact from the perspective of the worker themselves, not just an abstract “task.”
5. Adoption data — using empirical data about how and how much AI is actually used in the workplace, rather than theoretical estimates.

Shared responsibility: researchers and policymakers

The study emphasizes that responsibility for bridging both gaps is shared.

For policymakers the authors recommend: broaden the evidence base by drawing on multiple types of measures, include workers as epistemic partners — people who possess directly relevant experience and knowledge of their own working conditions — and shift focus from prediction to preparedness. Instead of asking “which occupations will disappear?”, the more relevant question is: how do we now prepare the workforce for uncertain change?

For researchers the recommendations are complementary: build data infrastructure designed from the outset to answer policy needs, and use interdisciplinary methods that bridge the gap between economics, sociology, computer science, and public policy research.

Broader context

Cohere Research’s paper comes at a moment when debates about AI regulation and its impact on the labor market are being conducted across the EU and beyond. Static scores such as those from Eloundou et al. (2023) have become almost the standard analytical input in such debates — and this is precisely why such a methodological critique is timely.

Without a better methodological foundation, policies rest on measures that oversimplify reality: they do not account for the pace of change, differences between labor markets in different countries, or the perspective of the workers whose lives are at stake. This paper sets a concrete agenda for how to fix that situation.