The recruiter can't tell you what your job looks like in ten years. The published research can get you closer.

Writing reports, building recommendations, and synthesizing data is core LLM territory — half this job’s tasks show measurable exposure. The 2013 model rated it low-risk because "analyze and recommend" work wasn’t what that generation of automation research was built to flag.

The single highest-exposure occupation in this curated set — 63% of tasks touched by LLMs plus supporting software, because building models and writing up analysis is close to what LLMs do natively. The 2013 model, working from a completely different definition of "automatable," rated it almost immune (3.5%).

Planning documents, forecasts, and coordination memos are language-heavy — 45% task exposure in the LLM study. The 2013 model scored this job almost immune (1.2%) because spreadsheet-and-memo planning work doesn’t fit a model built around physical/procedural automation.

Report writing, pattern analysis, and briefing production are the core of the job — real, meaningful LLM exposure (40%) in the 2023 study. Frey & Osborne’s 2013 appendix never scored "Intelligence Analysts" as a distinct occupation (it wasn’t broken out as its own line in their 702-job list), so there’s no comparable 2013-era number — we’re not going to borrow one from a neighboring title and pretend it fits.

Curriculum writing, lesson plans, and SOPs are exactly the kind of drafting work LLMs are rated highly exposed on — 59% of this job’s tasks show up as touched in the OpenAI-funded study. The 2013 model, built years before generative AI existed, called this job nearly automation-proof (1.4%) because it was scoring physical/procedural roboticizability, not writing.

Job postings, policy memos, and HR correspondence are classic LLM-exposed writing work (59%). This occupation doesn’t appear anywhere in Frey & Osborne’s original 702-job appendix, so there’s no 2013-era comparison point for it — we’re not inventing one.

Healthcare administration runs on reports, compliance paperwork, and scheduling — meaningful LLM exposure (37%). The 2013 model considered management occupations essentially un-automatable (0.7%): judgment-heavy people-management didn’t score as automatable under that model’s criteria.

Safety programs, inspection reports, and compliance paperwork are language-heavy — 36% exposure in the 2023 study. The 2013 model rated it low-risk (17%) under this same legacy SOC code, before it was renumbered 19-5011 in the 2018 federal taxonomy update — a bookkeeping change, not a different job.

Design documentation, spec writing, and calculation work show real LLM exposure (41%). The 2013 model rated engineering design low-risk (10%) — creative technical problem-solving didn’t fit that era’s definition of automatable.

The sharpest split in this dataset. The 2013 industrial-automation model rated this job 84% computerizable — hands-on testing and measurement looked highly proceduralizable to that model. The 2023 LLM-specific study rates it only 33% exposed: wiring, testing, and troubleshooting physical hardware isn’t something a chatbot does, no matter how good it gets at writing.

Documentation, scripting, and config-file work sit squarely in LLM territory (51% exposure). The 2013 model — filed under this occupation’s old SOC number, 15-1142, since renumbered 15-1244 in 2018 — rated it almost automation-proof (3%), because hands-on server-room work didn’t fit that era’s model.

Flying an aircraft isn’t a language task, so LLM exposure reads low (22%). The 2013 model called it closer to a coin flip (55%) — that paper was written during the early wave of serious autonomous-flight R&D and treated flight operations as plausibly roboticizable within a couple of decades.

Both studies agree on this one: flying a commercial aircraft is not exposed to language-model automation, nor was it rated a near-term robotics-automation target back in 2013.

Another sharp divergence, and a genuinely useful one: the 2013 model rated aircraft maintenance 71% computerizable, treating repetitive procedural work as automatable by future robotics. The 2023 LLM study rates it just 6% exposed — turning a wrench on a turbine engine is not a language task, no matter how good the chatbot gets.

Patrol work is physical, situational, and legally accountable in ways language models don’t touch. Two studies, a decade apart, using completely different methods, both land in the same place: low exposure.