EXPOSED
The daily substance of this job — literature review, prototype code, ablation scripts, results tables, paper and grant drafting — is exactly the text-and-code work frontier models already do at usable quality, and this occupation is uniquely exposed because it builds the tools that automate it. What resists is choosing which problems are worth attacking, designing experiments whose results are actually informative, and owning claims that a lab, a funder, or a product organization will bet on. The modal worker is a PhD researcher in industry or a federally funded lab, not a tenured professor; no license protects the role, and headcount concentrates around the people who set research agendas rather than execute them.
Headcount grew steadily across the period.
Median pay $122,840 → $140,300 -8.6% in real terms
This line is counted by the Bureau of Labor Statistics — the one figure on this page that isn't a judgement of ours. Headcount moves on demand, offshoring, demographics and the business cycle, and automation is one term among several, often not the loudest.
So a falling line is not evidence that AI did it, and a rising one is not evidence that it won't. Both happen in this register: some occupations resist automation and shrink anyway, others are highly automatable and keep growing. The marked year is 2020.
BLS projection, 2024–2034
+19.7% 40,300 → 48,300 on the projections basis
Exposed, but growing
AI can already do a lot of these tasks, and the BLS still expects +19.7% more of these jobs by 2034. Demand for the output is growing faster than the work is being automated away — the mechanism BLS gives for software developers, and the combination people most often misread as an error.
Different clocks. The score is what current AI could do to this work today. The projection is how many of these jobs will exist in 2034. Everything between the two — how fast employers actually adopt, whether demand grows in the meantime — is why they can point opposite ways without either being wrong.
~3,200 openings a year on average, including replacing people who leave.
ScientistCryptologistApplied ScientistResearch EngineerComputer ScientistResearch ScientistComputer SpecialistLanguages ResearcherInformation ScientistComputational LinguistComputational ScientistComputer Vision ScientistDigital Solutions ManagerMachine Learning EngineerSite Reliability EngineerMachine Learning ScientistResearch Computer ScientistScientific Programmer AnalystComputational Theory ScientistMachine Learning Data ScientistProgramming Languages ResearcherControl System Computer ScientistMachine Learning Software EngineerMulti-Disciplined Language Analyst
AI Engineer
Holding it up: judgment & accountability . Weakest point: liability shield .
Mixed — a routine tier and a judgment tier A 12 reflects the split inside a single working week: the implementation half — writing training loops, sweeping hyperparameters, running ablations, building the results table, drafting related-work sections and NSF/DARPA narrative boilerplate — is already being done by coding agents at rates a first-year grad student can't match, while the half that holds is formulating a hypothesis that isn't already saturated, deciding what baseline would actually falsify it, and recognizing when a benchmark gain is measurement artifact rather than signal; it sits below 14 because that formulation work is itself text, not an act the tools are structurally barred from.
Fully desk- and screen-based A 3 covers the little that isn't a laptop and a cluster queue: standing up GPU nodes, physical robotics or HCI apparatus for the minority of researchers in those subfields, and conference travel — nothing that requires your hands in an uncontrolled setting, since the SLURM job runs whether you're in the building or not.
No licence, no signature requirement A 2 is close to floor because no jurisdiction licenses computer scientists: your PhD is a credential, not a practice permit, and when a published result fails to replicate the consequence is a retraction and reputational damage, not a board action — the only formal exposure is IRB approval and export-control or classification rules on specific projects, which attach to the institution rather than to you personally.
Exists to be accountable for ambiguous calls A 15 is where it lands because you are the person who says a research direction is worth eighteen months of a team's headcount and compute budget, that a safety or capability claim is sound enough to publish under an institution's name, and that a negative result should kill a program — calls with no procedure to follow, made on incomplete evidence, and traceable to you when they turn out wrong.
The verdict above describes this occupation as a whole. Almost nobody does the typical version of a job — tick what's actually in your week and see how your own mix sits.
Your task mix speaks to task resistance (12/20 here) — how much of the day's work current AI already does. That is the dimension the boxes above are about.
It cannot move the other three. Liability shield (2/20) is whether the law requires a licensed human to sign. Trust premium (8/20) is whether buyers specifically pay for a person. Judgment and accountability (15/20) is whether the role exists to own consequential calls. Those are facts about the occupation's standing, not about which tasks are in your week — a paralegal who does only trial exhibits still holds no licence. Together they are 25 of this occupation's 40 points (63%).
Embodiment (3/20) is also a property of the work rather than the worker, but we don't tag individual tasks as physical or not, so the picker can't tell you anything about it. That's a limit of this tool, not a claim.
Did we get the list right? Tell us what's missing — the tasks are written from the outside, and you're reading this from the inside.
No occupation passed every test: close enough to computer and information research scientists on skills and subject matter, at least 10 points more resistant, no big jump in training, no new licence, no pay cut, and not shrinking on its own. That happens for 223 of the 654 occupations here that aren't SAFE, and it is worth stating plainly rather than leaving the section off.
The usual reason is that exposure travels with the skill profile. The jobs most similar to yours tend to be exposed for the same reasons yours is, so the near neighbours don't clear the gap — and the ones that do are a different kind of work, not a transfer of what you already know. Read that as a limit of this method, not a verdict that you're stuck: it only compares whole occupations, and it cannot see specialisation, industry, or anything you'd bring that isn't in a federal skill survey.
Here is that claim on your own job rather than in the abstract. These are the three occupations closest to this one by skill and subject matter — the places the work would most naturally transfer — with what the register scores them:
That is the whole problem in three lines. The nearest work is not meaningfully safer, so there is no move here that trades a similar skill set for a better verdict. This is not us running out of ideas — it is what the neighbourhood looks like.
The moves above are yours to make. This is the other half: what would have to change in the world for the occupation itself to score higher. None of it is in any one person's gift, but it is where the floor actually comes from. Scores here are not a one-way ratchet. Only two of the five dimensions — task resistance and embodiment — track what machines can do. The other three track law, what buyers will pay for, and who is answerable, and those move in both directions, often in response to the same pressure AI creates. If every lever below landed, this occupation would score around 56/100, still EXPOSED.
Task-mix shift is genuine here: if agentic coding and literature tools absorb the ablation-script/results-table/related-work tier, the residual job is problem selection, experiment design that yields informative negative results, and adjudicating whether a benchmark gain is real. This raises task_resistance for the surviving (smaller) headcount even with no institutional change.
Named-researcher accountability for model release decisions: if internal frontier-safety frameworks (Anthropic RSP, OpenAI Preparedness, Google DeepMind FSF) harden into audited sign-offs where a specific research lead attests capability-threshold findings, the role owns a consequential call under ambiguity.
Export-control and clearance gating: if advanced-compute or model-weights work requires cleared US persons (BIS rules, DoD/IC contracts), buyers are paying for a specific human body rather than a capability, and that premium is structural rather than sentimental.
A statutory attestation duty attaching to a named individual — e.g. EU AI Act GPAI obligations or a US state frontier-model law (California SB 53-style transparency reporting) implemented so that a designated technical officer or research lead personally signs capability and risk evaluations, with penalties for false attestation.
Frontier-model safety evaluation and red-team science becoming a distinct research function with no self-referential automation path — e.g. work under the US/UK AI Safety Institute evaluation agreements, or EU AI Act Article 55 systemic-risk model evaluations, where the evaluator cannot be the system under evaluation.
Research-integrity exposure: if federal funders (NIH/NSF) or journals require a named human to attest that results were not fabricated by generative tools, with debarment consequences, a weak personal-liability hook appears — weak because it is administrative, not licensure.
The limit. No licensure body exists for this occupation and none is being proposed, so liability_shield has a low realistic ceiling; the attestation routes above attach to a handful of senior leads, not to the 37k. Trust premium cannot rise on craft grounds — funders and product orgs buy results, not human authorship. The dominant risk is not scored here: headcount concentration, where the judgment tier survives but employs far fewer people.
| Washington-Arlington-Alexandria, DC-VA-MD-WV | 3,420 | $161,090 +15% |
| San Jose-Sunnyvale-Santa Clara, CA | 2,210 | $218,420 +56% |
| New York-Newark-Jersey City, NY-NJ | 2,100 | $170,890 +22% |
| San Diego-Chula Vista-Carlsbad, CA | 1,680 | $121,530 -13% |
| San Francisco-Oakland-Fremont, CA | 1,670 | $171,440 +22% |
| Boston-Cambridge-Newton, MA-NH | 1,620 | $170,510 +22% |
| Seattle-Tacoma-Bellevue, WA | 1,620 | $211,270 +51% |
| Baltimore-Columbia-Towson, MD | 1,070 | $156,750 +12% |
| San Jose-Sunnyvale-Santa Clara, CA | 2,210 | $218,420 +56% |
| Seattle-Tacoma-Bellevue, WA | 1,620 | $211,270 +51% |
| Portland-Vancouver-Hillsboro, OR-WA | 660 | $206,220 +47% |
Reported that Amazon carried out staff reductions within its artificial general intelligence (AGI) division.
La Jornada reports Amazon cut jobs within its artificial general intelligence division.
CNBC reported that Amazon cut jobs within its artificial general intelligence division.
Reuters reported that Amazon cut jobs within its artificial general intelligence organisation.
CNBC reports Meta is beginning layoffs this week, framed around the company's AI restructuring priorities.
Has AI actually changed your work? One tap, anonymous, and the running tally is public. Nothing else is asked of you.
Rather than check back: get the digest and we'll tell you what changed — or watch a single occupation from its own page.