AI Exposure Scores: what they measure, what they miss, and what comes next

Technical Summary: AI Exposure Scores: What they measure, what they miss, and what comes next

Problem Statement
The paper addresses a critical asymmetry in the "future of work" debate between researchers and policymakers. Since 2023, the "GPTs are GPTs" scores (Eloundou et al., 2023) have become the dominant empirical input for policy analysis regarding AI's labor market impact. These scores define "exposure" as the percentage of occupational tasks a Large Language Model (LLM) can assist with, calculated against the U.S. Department of Labor's O*NET taxonomy.

The core problem identified is that while the original authors explicitly noted the limitations of their static, task-based methodology, these caveats have not traveled with the scores as they have been adopted by downstream policy analyses. This has created a widening gap where policymakers rely on static, backward-looking signals to answer complex, forward-looking questions about who is harmed, how, and when. The paper argues that the current reliance on a single, fixed metric fails to account for temporal shifts in AI capability, geographic variations in labor markets, and the ontological limitations of decomposing work into discrete tasks.

Methodology and Analytical Approach
This paper does not present new empirical data or a novel scoring algorithm. Instead, it employs a critical review and synthesis methodology to:

1. Deconstruct the "GPTs are GPTs" scores: Analyze the specific constraints of the original methodology (temporal, geographic, and ontological) and demonstrate how these constraints compound when applied to policy questions.
2. Map the diffusion of scores: Trace how the 2023 scores have been utilized in academic and policy literature from 2024 to 2026, highlighting the disconnect between the scores' design and their application.
3. Survey emerging research: Systematically categorize recent methodological innovations that attempt to address the limitations of static scoring. The authors organize these into four families:
   • Dynamic and benchmark-based measures.
   • Ensemble methods.
   • Task-framework extensions.
   • Worker-centered metrics.
4. Propose a tripartite framework: Synthesize the findings to argue for a shift from purely predictive (ex-ante) measurement to a combination of signals, resilience frameworks (ex-post), and collective imagination.

Key Contributions
The paper makes three primary contributions to the discourse on AI and labor:

1. Diagnosis of the "Static Score" Gap: The authors rigorously detail three specific limitations that static exposure scores fail to capture:

   • Temporal Limitations: Scores are fixed snapshots of model capabilities (e.g., GPT-4 in early 2023). As frontier model capabilities advance (e.g., a 26% increase in the Epoch Capabilities Index by 2026), the scores become increasingly decoupled from reality. Furthermore, the scoring rubric itself is sensitive to the specific model used, leading to divergent results when replicated with newer models.
   • Geographic Limitations: The reliance on O*NET (a U.S.-centric taxonomy) introduces systematic distortions when applied globally. Issues include language bias in task descriptions, the exclusion of "data worker" categories (often offshored to the Global South), and the failure to account for regional differences in work formalization and education.
   • Ontological Boundaries: The assumption that work can be decomposed into discrete, scorable tasks ignores tacit knowledge, relational work, and workflow interdependencies. The paper notes that the rubric produces the most instability in roles requiring judgment and trust, precisely where policymakers need the most reliable estimates.

2. Synthesis of Emerging Methodologies: The paper surveys four distinct research directions that move beyond static scoring:

   • Dynamic Measures: Approaches like the "Remote Labor Index" and dynamic occupational exposure indices that track actual AI agent performance on real-world tasks or update scores against frontier models in near real-time.
   • Ensemble Methods: Research (e.g., Frank et al., 2025) demonstrating that combining multiple exposure frameworks yields better predictive power for unemployment risk than any single score, as individual scores are often weakly or negatively correlated.
   • Task-Framework Extensions: Models that account for task interdependencies (Demirer et al., 2026) and distinguish between "expertise-raising" and "expertise-lowering" automation (Klein Teeselink & Carey, 2026), showing that aggregate exposure masks economically meaningful heterogeneity.
   • Worker-Centered Metrics: Frameworks that incorporate worker preferences (e.g., "Red Light" zones where workers do not want automation) and adaptive capacity (e.g., Manning et al., 2026), revealing that high exposure does not always equate to high vulnerability.

3. Strategic Framework for Policy and Research: The paper proposes that better measurement alone is insufficient. It advocates for a three-pronged approach to navigating uncertainty:

   • Signals: Improved ex-ante exposure scores.
   • Actions: Ex-post resilience frameworks (e.g., "no regret" policies, scenario-based triggers) that prepare institutions for shock regardless of specific forecasts.
   • Imagination: The political work of collectively articulating and advocating for desirable futures, countering technological determinism.

Results and Findings

• The Capability Gap: Frontier AI capabilities have grown significantly (approx. 26%) since the original scores were calculated, rendering the static scores increasingly obsolete as predictors.
• Rubric Sensitivity: Replicating the original methodology with newer models (ChatGPT-5, Gemini 2.5, Claude 4.5) results in a 3.6-fold divergence in mean exposure scores, proving the original rubric is not stable across rating instruments.
• Geographic Distortion: Applying U.S. taxonomies to non-U.S. contexts (e.g., Denmark, UK) requires translation layers that narrow the scope of analysis or introduce assumed equivalences that may not hold.
• Ensemble Superiority: Combining five different exposure frameworks explains 29.8% of variation in unemployment risk on its own, significantly outperforming single-score approaches.
• Adaptive Capacity: High exposure does not linearly map to vulnerability; workers in highly exposed occupations often possess higher financial buffers and transferable skills, suggesting that exposure scores alone misidentify who is actually at risk.

Significance and Claims
The paper modestly claims to be a "bridge" rather than a final solution. Its significance lies in:

• Correcting the Evidence Base: It urges policymakers to widen the evidence base beyond the 2023 static scores, engaging with dynamic, ensemble, and worker-centered metrics to answer questions about harm and adaptation more reliably.
• Reframing the Goal: It argues for a shift in the research-policy relationship from "prediction" (forecasting specific outcomes) to "preparedness" (building resilience against uncertainty).
• Shared Responsibility: It posits that closing the gap between research and policy is a shared task. Policymakers must engage workers as "epistemic partners" and move toward scenario-based policy. Researchers must build data infrastructure that is globally representative, versioned, and decision-relevant.
• Political Agency: The paper concludes that exposure scores describe technical feasibility, not inevitability. The ultimate significance of the work is to empower stakeholders to engage in the "deliberate, political work of reimagining what futures are worthy of building towards," rather than accepting a deterministic trajectory.

The authors emphasize that while better measurement is necessary, it will not close the gap between research and policy alone; it must be paired with ex-post governance frameworks and the collective work of articulating desired futures.