Home › How to read an exposure score
Explainer
How to read an exposure score
- Exposed to what? “Could be done by a language model”, “is done by Claude users”, and “is being replaced” are three different claims. We keep them separate.
- Per task or per job? Jobs are bundles. An occupation with 60% exposed tasks is not 60% gone. The other 40% may be the part that gets you paid.
- Whose rating? On our job pages the human raters and GPT-4 often differ by tens of points. If a source gives you one number, ask where the disagreement went.
- How old? The “could” ratings use early-2023 model abilities. Usage figures come from sampled periods in 2025. Neither is today.
- Whose data? Usage numbers come from one vendor’s product and mostly US users. The task lists are written for US jobs. Outside the US, task-level usage data barely exists.
- Does anyone say “will”? Exposure studies describe potential. Even the studies that look for effects on employment report limited evidence so far. Treat any date or percentage for “replacement” as an opinion.
What a useful score looks like
One that shows its tasks, its sources and their dates, its uncertainty, and what it leaves out. If you can’t click through to the underlying tasks, don’t trust it. Every job page here does that.