Why curator work keeps a person in the room
Ask whether AI will replace curators and the answer sits in the mix of tasks, not in the job title. A curator writes, researches, catalogs and plans. A curator also decides what an institution should own, what it should show, and what it should say about objects that carry contested history. Software is useful for the first list. The second list is judgment under public scrutiny.
Two tasks make the point. Acquiring and appraising objects means handling the thing itself, checking provenance paperwork, and taking responsibility for a call that can cost a museum money and reputation. Negotiating loans with other institutions means trust between named people, plus insurance, condition terms and politics. Neither is a text problem.
The other half of the role is more exposed. Catalog records, collection descriptions, wall text drafts, grant paragraphs and visitor reports are all writing and structuring work, and language models do that quickly. That is where task erosion shows up first, usually in the hours a junior curator used to bill against a collection backlog.
What AI does, what it assists, and what it leaves alone
Our task split sizes the AI-does group at 24% of curator task time. The clearest cases are catalog and metadata entries, where a model can read existing records and produce consistent descriptions, and first-draft text for labels, press notes and funding applications. A person still signs off, but the blank page is gone.
The assisted group covers 27% of task time. Provenance research is the obvious example: search tools surface auction records, archival mentions and comparable objects far faster than manual lookups, and the curator checks every claim. Audience and attendance analysis is the other. Software can sort visitor data by exhibition and hour; deciding what that means for next season is still a curatorial call.
Work that stays with people accounts for 49% of task time. Acquisition and deaccession decisions sit here, because they commit an institution for decades. So does loan negotiation, along with community and donor relationships, staff supervision and the final interpretive line an exhibition takes. Coverage, our answer to “Can AI do it?”, comes out at 34 on a 0 to 100 scale; the coverage method page explains how that share is built.
What the evidence actually shows
There is no published test that puts AI systems against working curators on curatorial tasks. Our quality grade for this job is D, and a grade of D means exactly that: not measured. We give no parity number here, because inventing one would be worse than admitting the gap.
What would settle it is narrow and testable. A blind comparison of catalog records written by a model and by a trained cataloger, scored by collections managers. A provenance research task with a known answer, run against both. An exhibition proposal judged by a review panel without knowing the author. Until work like that exists, the honest position is uncertainty, and our quality parity method treats it that way.
Market data is firmer. The Bureau of Labor Statistics counts about 12,150 curators in the United States, with median pay of $63,420 and projected employment growth of 4.9% over 2025 to 2035 (BLS, 2025). That is steady demand in a small occupation, which matters: small fields hire infrequently, so even modest task erosion can show up as fewer openings rather than layoffs.
When the picture could shift
Most likely between 2034 and 2047 (8 in 10 of our scenarios). The replacement year method sets out what that window is measuring and how the scenarios are drawn.
Two things could pull the date earlier. Budget pressure is one: running collection software and language tools costs a fraction of a salary line, and small museums feel that gap first. Wider adoption of shared collection platforms is the other, because standardized records make automated description far easier to deploy across institutions.
Two things hold it back. Part of the job is physical, including object handling, condition checks and gallery installation, and that fraction would need dexterous humanoid robotics that is not in service today. Accountability is the second brake. When attribution, repatriation or a donor agreement goes wrong, boards and the public want a named curator who made the decision, not a model output.
What to do: keep a written record of the judgment calls you make, not just the outputs you produce, because the calls are the part that is hardest to hand over.
How curators stay needed
Lean into the work in the human column. Own acquisition and deaccession recommendations end to end, including the research trail behind them. Run loan negotiations and the relationships they depend on. Take the interpretive decisions in an exhibition, particularly around contested objects and community input, and be the person who defends them in public.
Two skills compound. First, verification: knowing how to check a machine-written provenance claim against primary records, fast. Second, supervision of AI-assisted output, so a department can use drafting tools without publishing errors on a gallery wall.
Nearby roles are worth comparing if you are planning a move. Look at archivists, museum technicians and conservators, and librarians and media collections specialists, all in the librarians, curators and archivists family. The arts and entertainment sector page shows how the wider field scores, and the jobs that mostly need a person list gives a sense of what keeps work with people.
This job’s Still needs a human score is 68 out of 100 (higher is safer). You can put curators and a neighboring role side by side on our comparison tool, or read how every figure on this page is built in the methodology.