{
  "source": "Cooked Index — occupational AI risk register",
  "page": "https://cookedindex.com/jobs/physical-scientists-all-other/",
  "methodology": "https://cookedindex.com/methodology",
  "notice": "Verdicts are re-examined as evidence accumulates. Re-fetch before relying on this; the page above always carries the current score.",
  "scored_at": "2026-08-11",
  "model": "claude-opus-5",
  "occupation": {
    "title": "Physical Scientists, All Other",
    "soc_code": "19-2099",
    "category": "Science",
    "us_employment": 22300,
    "median_annual_wage": 122570
  },
  "verdict": "EXPOSED",
  "risk_resistance": 42,
  "contested": false,
  "near_boundary": false,
  "dimensions": {
    "task_resistance": 11,
    "embodiment": 9,
    "liability_shield": 3,
    "trust_premium": 7,
    "judgment_accountability": 12
  },
  "reasoning": {
    "task_resistance": "Sample prep, XPS or SEM runs, instrument calibration and drift checks, and deciding which follow-up experiment actually discriminates between two hypotheses stay with you — but the literature triage, the Python that reduces your spectra, the statistics, and the first draft of the methods and results sections are now hours of assisted work rather than days, and those pieces are close to half of a typical week, which is what lands this at 11 instead of the 15+ a bench-only role would earn.",
    "embodiment": "Cruise-based water sampling, radiosonde launches, contaminated-site fieldwork and glovebox or vacuum-chamber work put real hands-on hours in the job, but the modal person in this bucket does that in a temperature-controlled lab or on scheduled campaigns, not daily in uncontrolled conditions — a 9 reflects field-and-bench work that is intermittent and mostly staged, unlike a well-site geologist who is outdoors by default.",
    "liability_shield": "There is no licence to practise as a materials or atmospheric scientist: a PhD and a lab safety certification get you to the bench, and when a result is wrong the institution, the PI, or the accrediting body (ISO 17025, ASCLD for forensic labs) absorbs it — the 3 rather than 0 is only because forensic examiners can be cross-examined personally on their own findings.",
    "trust_premium": "Peer review is blind and papers are read for the data, not the author, so most output travels without you attached; the 7 comes from the parts that don't — program managers who fund you again because your last deliverable was clean, and the internal client who brings a failure analysis to you specifically because you told them the truth about the last one.",
    "judgment_accountability": "Calling a peak real versus instrument artifact, deciding a contaminated run gets discarded rather than reported, and setting detection limits and uncertainty bounds are genuinely yours and often unfalsifiable in the short term — but much of the work runs on written protocols, SOPs and validated methods, and the consequential calls usually go up to a PI, program office, or lab director before anyone acts, which caps this at 12."
  },
  "rationale": "This is a catch-all bucket (materials and surface scientists, oceanographic and atmospheric specialists, forensic and government lab scientists), and the modal worker splits time between instrument-based lab or field measurement and desk work: literature review, data reduction, statistical modeling, and report and grant writing. The desk half is the most automatable part of the job — AI already drafts methods sections, writes analysis scripts, and screens literature faster than a postdoc. What holds is hands-on experimental design, running and calibrating physical instruments, field sampling, and being the person who signs off on whether a result is real.",
  "outlook": "Headcount holds roughly flat but the job reshapes: fewer hours on analysis and writing, more on instrumentation, field deployment, and standing behind conclusions.",
  "what_would_raise_it": {
    "levers": [
      {
        "dimension": "task_resistance",
        "change": "As AI absorbs literature screening, script-writing and report drafting, the residual day becomes experimental design, instrument calibration/troubleshooting, and adjudicating anomalous results — a genuine two-tier split in this bucket. Watch for lab staffing plans where the analyst tier is cut and remaining scientists are reassigned to method development and validation.",
        "plausibility": "already happening",
        "would_add": 3
      },
      {
        "dimension": "liability_shield",
        "change": "Forensic and environmental sub-populations are the route: if ISO/IEC 17025 accreditation bodies (A2LA, ANAB) or ASCLD/ASB standards add an explicit requirement that a named qualified analyst personally review and sign any AI- or algorithm-derived result before report release — mirroring the technical-reviewer signature already required for casework — the signature becomes non-delegable. Also watch state environmental labs' NELAP certification rules and expert-witness admissibility fights over black-box analysis under Daubert.",
        "plausibility": "plausible",
        "would_add": 5
      },
      {
        "dimension": "judgment_accountability",
        "change": "If agency reproducibility/validation policies (e.g. EPA method validation, NOAA data-quality directives, or a federal research-integrity rule) require a named individual to attest that an AI-assisted analysis was independently verified, the role formally owns the 'is this result real' call rather than informally holding it.",
        "plausibility": "plausible",
        "would_add": 3
      }
    ],
    "ceiling_note": "Trust premium has no realistic route: buyers here are agencies, courts and internal R&D clients who pay for accredited output, not for a human per se. The catch-all nature of the SOC also means gains concentrate in the forensic/regulatory-lab slice; industrial materials scientists likely see none of it."
  },
  "adjudication": null,
  "employment_history": {
    "points": [
      {
        "y": 2017,
        "emp": 17320,
        "wage": 103990
      },
      {
        "y": 2018,
        "emp": 17380,
        "wage": 107230
      },
      {
        "y": 2019,
        "emp": 17550,
        "wage": 109910
      },
      {
        "y": 2020,
        "emp": 19050,
        "wage": 107210
      },
      {
        "y": 2021,
        "emp": 19680,
        "wage": 104100
      },
      {
        "y": 2022,
        "emp": 19820,
        "wage": 107970
      },
      {
        "y": 2023,
        "emp": 20820,
        "wage": 112280
      },
      {
        "y": 2024,
        "emp": 22580,
        "wage": 117960
      },
      {
        "y": 2025,
        "emp": 22300,
        "wage": 122570
      }
    ],
    "from": 2017,
    "to": 2025,
    "change_pct": 28.8,
    "comparable_from": 2019,
    "spans_soc_revision": true
  },
  "pivots": [],
  "license": "https://cookedindex.com/terms"
}