Disclaimers 4

Acknowledgements 5

Executive summary 6

1 Introduction 9

2 Evaluating the impact of AI on jobs 10

What measures of AI exposure are out there 10

Why we need another measure 10

3 Using the OECD AI Capability Indicators to Measure AI Exposure 12

The OECD AI Capability Indicators 12

How the Indicators can be used to build a Measure of AI Exposure 12

What analyses can the indicators be used for 13

4 Manually mapping the AI Capability Indicators to selected occupations 14

Selecting a sample of occupations representing the diversity of capability profiles 14

Individually analysing and rating the demands of the selected occupations 15

Finding consensus among raters and establishing the OECD standard of truth for the sample of occupations 15

Establishing the OECD average ratings of capability demands for the sample of occupations 15

Benchmarking large language models against the human average 16

5 Extending the mapping of AI Capability Indicators to the whole occupational space 17

LLM ratings for occupational groups 17

Within-group variation 18

Validating the LLM ratings of capability demands 24

6 The OECD AI exposure measure 26

Conceptual framework 26

Results 28

Comparison with Felten et al. and the implications of broader capability coverage 34

7 Scenario applications and future directions 38

Scenario-based application: A case study of increasing AI cognitive capabilities by one level 38

Next steps 39

Conclusion 41

Annex 1. Sample selection and occupational sorting procedure 42

Annex 2. Detailed manual rating procedure 45

Annex 3. Consensus process and convergence diagnostics 47

Initial methodological challenges to the rating process 47

Methodological clarifications during the discussion 48

Evolution of consensus finding 50

Annex 4. LLM benchmarking protocol and detailed results 54

References 57