Home › Possible vs. observed
Explainer
Possible vs. observed
Read two articles about AI and jobs and you will get two different numbers for the same job. Usually they are not contradicting each other. They answer different questions.
Three questions
- Could it? Researchers read each task in a job description and ask whether a language model could do it or cut the time in half. That is the Eloundou et al. “GPTs are GPTs” approach. Across all 17,938 tasks here, human raters said yes for 47% (alone or with extra software); GPT-4 as rater said yes for 56%.
- Is it happening? Count what people actually ask a chatbot to do and match it to tasks. In the Anthropic Economic Index data, only 9% of tasks show measurable use. This is a lower bound for AI in general: it is one company’s product.
- Is it changing employment? Look at unemployment, hiring, wages. Anthropic’s study looked for this among highly exposed US workers and found no systematic rise in unemployment since late 2022, plus suggestive but tentative evidence of slower hiring of workers aged 22 to 25 into exposed jobs.
A caution about the second question
Anthropic’s observed measure is built on top of the first: a task counts as covered only if it was rated feasible and shows enough Claude use. In our joined data, 1,651 of the 1,654 tasks with measurable use were also rated as something an LLM could do or speed up by GPT-4. So “observed” cannot exceed “could” by construction, and part of the gap you see is that arithmetic, not a finding. What is a finding is how small the covered share is next to the feasible share.
Why the first two differ so much
Anthropic’s own paper gives the clearest example: for computer and maths occupations, about 94% of tasks were rated feasible, but around 33% are covered by observed Claude use. Plausible reasons include that capability is not the same as reliability, that organisations adopt slowly, that law, liability and trust matter, and that the early-2023 ratings may already be out of date in both directions. The data cannot tell these apart, and neither can we.
How to use this
- For a headline number, ask which of the three questions it answers.
- Treat “could” as the ceiling of attention, not of impact.
- Treat “observed” as a floor for one product, growing with time.
Sources: Eloundou et al. (OpenAI/UPenn) via GitHub; Anthropic, Labor market impacts of AI (March 2026); Anthropic Economic Index data. Percentages computed on our joined dataset.