Which STEM work is susceptible to automation

Which jobs are most exposed to displacement

Pick the question you actually want answered — the ranking changes a lot depending on which one it is. Click any row to load that occupation's tasks and dimension profile at the bottom of the page.

The handoff frontier

Tractability (right) is whether AI can lead the work; resistance (up) is whether it will be permitted to. The curve is the frontier — the resistance a given tractability can currently overcome. Work below it has crossed; work above it is still human-held. Framework from Watson (2026); axis positions are relative to other STEM occupations, and the curve is calibrated to this corpus, not derived from theory.

Where cognitive leadership sits, and where it could

Watson's six-stage scale. The light dot is how many occupations sit at each stage given what is actually deployed today; the dark dot is where current AI capability could already put them. The distance between the two is pending handoff. He expects two crossings to carry most of the strategic weight: proposing action to taking it, and human veto to after-the-fact audit — the second being a crossing by erosion, with no event to observe.

Where each occupation sits

Each dot is one occupation, placed by the importance-weighted average of its tasks. Right = a machine could do more of this work. Up = more of it requires a human to do it or answer for it. The dividing lines are the medians across these STEM occupations, so a quadrant label means "relative to other STEM work", not "safe" or "doomed". Dot size = number of scored tasks. Click a dot to load that occupation's tasks below.

Does this agree with anyone else?

Left: agreement with human annotators is the evidence the index measures what it claims. Right: the pre-LLM measure disagrees, and should — Frey & Osborne scored the routine/manual gradient, while LLMs land hardest on non-routine cognitive work they called safe. A tight line on the right would have been the warning sign.

The occupation network

Occupations linked to their three most similar peers by shared subtasks (575 of 4,656 edges — the full graph is a hairball). Position is force-directed, so clusters are groups of occupations that do the same kind of work. Colour is susceptibility. Click a node to load its tasks.

Which subtasks have the most leverage

A susceptible subtask used by 50 occupations matters far more than one used by a single job. Up and to the right = automatable and widespread — the activities whose automation would touch the most of STEM work.

Capability is ahead of deployment

The left dot is what today's deployed technology can actually do; the right dot is what a current model could do in principle. The gap is where change is pending rather than finished — these are the near-term candidates, not the already-automated ones.

Weighted by how many people actually do the work

These charts are at the 6-digit SOC level, not the O*NET level used above. BLS reports employment per SOC, and several O*NET occupations roll into one SOC — attaching SOC employment to each O*NET row would count the same workers repeatedly (3.4 million registered nurses, five times over). Susceptibility is averaged across an SOC's O*NET occupations before the employment is attached once.

Subtasks ranked by susceptibility

The shared vocabulary underneath every task. Jobs is how many occupations use this subtask — a highly susceptible subtask used by 50 occupations matters far more than one used by a single job. Click a column to sort.

Tasks — click an occupation above

Task-level scores, inherited from the subtasks each task maps to. Sorted by susceptibility.

Bars are this occupation's seven rated dimensions; the vertical tick is the median across all 268 STEM occupations.