EXPOSED
Much of the daily work is document- and code-heavy: writing safety analysis reports, assembling NRC licensing submittals, running neutronics and thermal-hydraulic codes, and tracking design changes — all areas where AI already drafts, cross-references regulations, and sets up parameter studies competently. What resists is the accountable engineering judgment: validating that a model matches a physical plant, defending assumptions to the NRC, and signing off on decisions where the failure mode is a radiological release. Plant-side work — outage support, walkdowns, radiation shielding verification, fuel handling oversight — keeps a real physical component that robotics cannot cover.
Mixed — a routine tier and a judgment tier. At 12 rather than 8, MCNP/RELAP5 input decks, 10 CFR 50.59 screening write-ups, and Chapter 15 accident-analysis chapters are all templated enough that AI drafts credible first passes, but reconciling code output against as-built plant configuration, resolving NRC requests for additional information, and interpreting anomalous core-follow or dosimetry data still require someone who has walked the containment.
Some physical or field component. A 9 reflects the split: most weeks are at a workstation, but fuel-shuffle oversight, in-containment shielding walkdowns during refueling outages, ALARA planning at the job site, and hot-cell or reactor-bay work under an RWP put you in a radiological area with a TLD on your badge — controlled environments, not a rooftop or a trench, which is why it isn't 14.
Licensed human required and personally liable. 13 comes from the PE stamp on safety-related design and NRC-recognized roles — the licensing basis submittal names a responsible engineer, and 10 CFR Part 21 defect reporting attaches personally — but a large share of nuclear engineers work under a utility's or vendor's Appendix B quality program without ever stamping a drawing, which caps it well below the 18-20 of a structural PE.
Some relationship component. At 8, your NRC project manager, the plant's licensing lead, and the fuel vendor know you by name across multi-year license amendment and reload cycles, and that continuity matters when defending an assumption — but the deliverable is a calculation package that any qualified reviewer can verify, so the relationship supports the work rather than being the work.
Exists to be accountable for ambiguous calls. 15 is earned on calls like whether a degraded ECCS condition is still within the licensing basis, whether a 50.59 change needs a license amendment, and what margin to accept on peak clad temperature — decisions made in the gap between code output and reality, where being wrong is a radiological release and the regulator will ask you personally to defend the assumption.
Has AI actually changed your work?