Why One Exposure Ranking Is Never Enough
The OECD occupational exposure map is most useful when read as a set of capability profiles, not as a single ladder from safe to threatened. This chapter examines what those profiles reveal across occupational groups and why the OECD’s broad frontier produces different rankings from a narrower measure centred on language and information processing. The practical goal is to choose the measure that matches the question being asked, while avoiding the false precision of a universal exposure ranking.
A single ranking is attractive because it seems to settle a difficult question quickly. Put occupations in order, identify the top and bottom, and the future appears legible. But a ranking can conceal the reason an occupation is near the top. It may be close to AI in routine language work, in visual inspection, in equipment operation, or in another capability entirely. Those mechanisms imply different forms of assistance, different constraints on whole-role substitution, and different evidence a manager or worker would need before acting.
The occupational space has several centres, not one peak
The OECD results show a recognisable pattern, but not a simple high-skill versus low-skill hierarchy. Current AI capabilities are closest to occupations with standardised, codifiable, and repeatable information or equipment-mediated tasks. Office and Administrative Support has the lowest major-group total gap, at 0.8, followed by Production at 1.8, Food Preparation and Serving at 2.4, and Sales and Related at 2.6. In the index’s logic, lower remaining gaps indicate greater potential exposure because AI is closer to the measured capability profile.
At the other end, Community and Social Service has a gap of 6.4, Legal and Educational Instruction each 5.8, Healthcare Practitioners and Technical 5.6, Management 5.5, and Protection 5.2. These groups tend to combine capabilities for contextual interpretation, responsibility, social interaction, complex judgement, or physical action in settings that do not follow a fixed script. The numbers do not imply that their workers will avoid AI. They indicate that current systems remain farther from the full configuration of capabilities the occupations require.
The domain averages explain why the group pattern looks this way. The largest average gaps are in Social Interaction, Problem Solving, and Metacognition and Critical Thinking, each at 1.1. Manipulation remains at 0.7 and Robotic Intelligence at 0.6. Smaller average gaps appear in Creativity at 0.1, Vision at 0.2, Language at 0.4, and Knowledge, Learning and Memory at 0.5. The interpretation is not that AI has solved creativity in every meaningful sense. The report cautions that the upper end of its creativity scale describes exceptional or world-class creativity, which whole occupations rarely require; that compresses the active range and makes the average gap look unusually small.
The broader pattern is therefore about kinds of work. Structured language, codified knowledge, routine classification, and predictable monitoring are relatively close to current capabilities. Social understanding, open-ended contextual reasoning, accountability, fine dexterity, and adaptation in dynamic environments remain farther away. These are not permanent human monopolies.