Who drafts and who still signs? A multi-instrument map of generative AI exposure across 50 UK civil and environmental occupations

This paper maps generative-AI exposure across 50 UK civil and environmental occupations as a task-allocation problem: where a machine draft may enter a workflow, and where a person still has to check and sign. Scores are derived from 20 recorded runs for each occupation with ChatGPT, DeepSeek and Gemini 3 Pro (3,000 observations). The ranking is compared with category-level judgements from 52 practitioners, ONS automation probabilities, Felten et al.’s AI Occupational Exposure index and a language-modelling proxy. Annual Population Survey weights are applied once to each of 16 SOC 2020 unit groups. Digital documentation and estimation has the highest unweighted family mean (4.00); field survey and site investigation has the lowest (2.43). At unique-SOC level the LLM ranking is positively associated with Felten AIOE (Spearman ρ = 0.563, 15 groups) and is too imprecise to be separated from no association with the ONS baseline (ρ = −0.179, 14 groups). The main prompt already rewards standardised text and data work. Professional liability and statutory sign-off are discussed as a hypothesis, not estimated.