Why the objects keep the work with people
Ask whether AI will replace museum technicians and the answer starts with the objects themselves. A corroded bronze, a flaking panel painting or a water-damaged photograph exists once. A treatment decision that goes wrong cannot be undone, retrained or rolled back. That is a different risk profile from drafting text or sorting records, and it shapes almost everything about how this job is scored.
Two parts of the daily work show it clearly. Cleaning and stabilizing fragile material means reading how a surface responds under your hand, in that light, in that moment, then changing pressure, solvent or tool. Building mounts, supports and crates for installation and transit means fitting a one-off object with no second copy to practice on. Both are judgment plus touch, and both carry consequences that a museum’s own ethics rules force a named person to own.
There is a second reason, less about dexterity. Conservation work sits inside institutional accountability. Condition reports, treatment proposals and documentation are read by curators, registrars, lenders and insurers. Someone has to sign them and defend the reasoning years later. Software can draft, compare and flag. It does not carry professional responsibility for an irreversible intervention.
What software does, what it assists, what stays in human hands
Some tasks already move. Entering and updating accession and catalog records, and producing first-draft object descriptions or label copy from existing documentation, are text-and-data jobs that current tools handle quickly. Of the task time AI could touch in this job, our split shows the share it can do outright: 0%. That is the part of the role most likely to be quietly absorbed into collections software over the next few years.
A bigger group is assistance rather than substitution. Imaging and documentation benefit from automated comparison of before-and-after photographs, pattern spotting across large condition datasets, and faster literature and provenance searching. Material research is similar: a model can surface comparable treatments and published analyses, while a technician still runs the test and reads the result. The assisted share is 20%.
The rest sits with people, and it is the core of the job: hands-on treatment of damaged or unstable objects, mount-making and physical installation, and the condition judgment that decides whether an object travels at all. That human share is 80%. Overall, coverage — our measure of how much task time AI can handle today — comes out at 17 out of 100, and you can read how that figure is built on the coverage method page.
How strong the evidence is
Weak, and it is worth being blunt about that. Our quality-parity grade for this job is D, which means no study has tested an AI system against trained conservators on this occupation’s real tasks. So there is no parity number here, and anyone claiming one is guessing.
What would settle it is narrow and testable. Blind trials where a model’s material identification is checked against laboratory analysis. Condition assessments from imaging tools scored against reports by accredited conservators on the same objects. Documented treatment proposals reviewed by a panel that does not know which came from software. Until work like that is published, the honest position is unmeasured, not safe and not exposed. The quality-parity method page explains how a grade moves up when real tests appear.
When the picture could shift
Most likely after 2036 (8 in 10 of our scenarios). What that window measures is explained on the replacement-year method page.
Two things could pull it earlier. The first is robot hands: most of the physical work in this job falls into the dexterous-humanoid tier, so real progress on fine, variable-pressure manipulation would matter more here than better language models. The second is budget pressure. This is a small occupation — about 12,310 museum technicians and conservators in the United States, with median pay of $51,440 (BLS, 2025) — and small teams under cost pressure tend to automate the documentation layer first and hire fewer juniors to do it.
Two things hold it back. Irreversibility and professional codes keep a named human in the decision, even where a tool performs well. And cheap software does not buy the thing that is scarce: hours of supervised bench practice. The cost panel above compares tool spend with staffing, but low tool cost is not the binding constraint when the constraint is trained hands.
Good to know: Projected employment change for this occupation is 3.6% between 2025 and 2035 (BLS, 2025), so the near-term story is task erosion and fewer entry-level openings rather than the role disappearing.
How to stay needed in conservation work
Lean into the parts of the job that stay human. Treatment of unstable and damaged material, where you decide and document the intervention. Mount-making and installation for objects that have no standard shape. And travel and loan condition judgment, where you say yes or no and explain why.
Two skills compound. First, analytical imaging and materials testing — learning to run and interpret the instruments, not just read the output. Second, writing that holds up under scrutiny: treatment proposals and condition reports clear enough for lenders, insurers and future conservators. Technicians who can also supervise and review AI-drafted catalog text will be the ones setting the standard rather than competing with it.
What to do: Pick one documentation task you repeat weekly, let software draft it, and spend the saved hours at the bench.
Nearby roles are worth comparing. Curators sit closest on interpretation and collection decisions, archivists on preservation of paper and digital records, and library technicians on cataloging and collection handling. You can see all three alongside this one on the librarians, curators and archivists family page, or in the wider arts and entertainment sector. To weigh two of them directly, use the side-by-side comparison tool; for jobs with a similar balance of hands-on work, see the list of jobs that mostly need a person. The full scoring approach is set out in our methodology.