The page itself gives the judgements. This one gives what they rest on: which technologies bear on the work, how the estimate moved since language models reached the public, and the method behind both.
—— this stretch contains a verified event- - - no event in this stretch — reconstruction only0 = no task exposed, 100 = every task exposed
● 8 verified events for this occupation, plotted at the date it happened — the parts of the line near a marker are anchored to something checkable.
Starts at 34 because audio recording had already taken the capture task in many courts — US federal bankruptcy courts since the 1980s — before this chart begins. It rises slowly with the releases that improved speech recognition on long, multi-speaker audio and with governments testing AI transcripts, and stays in the middle because courtroom audio remains hard for machines, laws tie certification and evidential weight to a person, and attempts to widen recording in California failed.
12022 H234General-purpose text generation reaches the public. Before this point, exposure came from automation that was already deployed — OCR, RPA, machine vision, self-checkout, dispatch algorithms. ChatGPT research preview (2022-11-30) ↗
22023 H136A general model that passes professional exams. First-draft quality crosses the threshold where professional work starts using it. GPT-4 (2023-03-14) ↗
32023 H238Vision input, long context and tool calling. Models can be pointed at documents and connected to systems, which is what moves process work rather than writing work. GPT-4 Turbo:128k 上下文、视觉、工具调用(DevDay) (2023-11-06) ↗
42024 H141The same capability gets much cheaper and faster. Nothing new becomes possible; a lot becomes affordable at volume, which is when deployment decisions change.
52024 H243Reasoning models that work through multi-step problems, and the first models that operate a computer by looking at the screen. The second one is what reaches software-operating jobs. OpenAI o1(推理);同期 Claude 的 computer use 进入公测 (2024-09-12) ↗
62025 H145Agents begin operating real software end to end rather than producing text for a person to paste. This is also when the first public reversals appear — organisations that automated and partly undid it. Claude 3.7 Sonnet 与 Claude Code:混合推理 + 命令行编码代理 (2025-02-24) ↗
72025 H247Long context and tool use become the default rather than a feature. Capability gains continue; the visible constraint shifts from what models can do to liability, procurement and cost. GPT-5(2025-08-07);Claude Opus 4.5(2025-11-24) (2025-08-07) ↗
82026 H149Long-horizon agents land inside specific industry workflows. Adoption becomes sector-specific rather than general. GPT-5.5:「专为实际工作打造」 (2026-04-23) ↗
9Now50The current assessment — this point is the impact index published on the occupation's page, so the curve is anchored to a number the site already stands behind. Worth noting for the flat curves: in the same weeks, a research preview of a shared specification for AI agents to operate physical devices was opened to research labs and manufacturers. That is the first capability class pointed at the physical occupations whose lines here barely move. GPT-6 Astra(2026-09-03);Claude Fable 5.1 / Mythos 5.1(2026-09-01);Model Hardware Standard 研究预览(2026-08-27) (2026-09-03) ↗
A flat line is not a forecast of safety. It says which tasks automation has reached so far — the occupations that moved least here are the ones where the constraint is physical or regulatory, and both of those can change.
LO
