Which real-world jobs does anime actually show? This project scans 21,988 anime and counts how often each occupation appears — a complete lookup table keyed to ISCO-08, the UN/ILO's standard occupation classification. The headline finding is the absence list: the jobs anime never names at all.

🔎 Live explorer: https://elmiram.github.io/anime-professions/  ·  search any profession, see which anime feature it, and how it trends by decade.

📊 The deck: a 12-slide fun tour of the findings (PDF).

A few things it found:

• Teachers are the most-named profession (in ~1,370 anime, 7.6%) — anime lives at school.
• In the 2020s, shopkeepers overtook teachers for #1, on the back of the isekai/slice-of-life boom.
• If you count fantasy roles, being royalty (1,792 titles) out-appears every real job.
• Of the ~150 jobs a character could plausibly be named as, anime never names 17 — even across millions of lines of dialogue: financial analysts, systems analysts, driving instructors, printers, fast-food workers, ships' engineers… the back-office and service tail. (See The occupation list below.)
• Your niche jobs, quantified: software developers 41, insurance reps 18 (Trigun's leads are literally insurance agents), pharmacists 15, tram/bus drivers 5.

A hobby data project. It builds a taxonomy of occupations, detects mentions of them in anime, classifies each mention with an LLM, and produces per-occupation counts, a decade-by-decade breakdown, precision/recall estimates, and three CSVs. A small local/static web app explores the results.

Real occupations are kept strictly separate from a fantasy stratum (ninja, mage, adventurer, royalty, samurai…) — the fictional roles are counted but never folded into the real-job totals.

The occupation list — 436 trimmed to ~150. Occupations are classified against ISCO-08, the UN/ILO's list of 436 official job codes. Most of those can't be named in a story — nobody announces "I'm a mineral-processing-plant operator" — so a zero for them would say nothing about anime. So the list is trimmed to the ~150 jobs a character could plausibly be identified as (the "renderable" set); the other ~286 are reported as untestable, not absent. Anime names 133 of the 150 — the 17 it never names are the headline finding, because they could plausibly appear yet don't. (Every anime is still scanned for all ~150; the trimming is of job categories, not titles.)

Detection itself is text-first, in three layers:

1. Corpus + protagonists — every anime's plot summary (via AniList).
2. Supporting characters — AniList character bios, with a role field (main / supporting / background) so a doctor who appears in six episodes still counts as representation.
3. Dialogue — ~128,000 Japanese subtitle files (from a community mirror), so a job mentioned only in passing gets caught too.

Detection is a two-stage pipeline:

• Stage 1 — high recall. A keyword lexicon (English + Japanese surface forms mapped to ISCO codes) casts a wide net over the text, accepting false positives.
• Stage 2 — precision. Every ambiguous candidate sentence is judged in context by an LLM (Claude Haiku): does a character here really hold this job? This kills the classic traps — negation ("I'm not a doctor"), hypotheticals ("if I were a pilot"), polysemy ("officer", "band"), and fantasy homonyms (a mecha pilot is not an airline pilot).

A stronger model (Claude Opus) then grades a stratified sample of the confirmed hits to estimate precision per occupation, and AniList tags provide a recall benchmark. Occupations whose precision falls below a floor are flagged unreliable in the output rather than reported as fact.

• Named, not shown. We detect jobs that are named in text. A profession shown on screen but never labeled — a festival food stall, a hotel receptionist standing in frame — is invisible to the method. The "never-named" list is exactly that: never named.
• Unreliable flags are real. ~a quarter of measured occupations are flagged unreliable (e.g. "aircraft pilot" is mostly mecha pilots) — treat those counts as upper bounds.
• Franchise vs entry. Counts deduplicate to franchise by default (Detective Conan's ~28 films = 1), but every entry is kept as its own row so you can expand or collapse.
• Denominator is printed next to every percentage (of ~18,127 titles with scannable text).

• Anime metadata, titles, character bios, tags: AniList (GraphQL API).
• Occupation taxonomy: ISCO-08 (International Labour Organization).
• Dialogue: a community archive of Japanese anime subtitles.

Subtitles are copyrighted works. This project derives counts and locators only — it never stores or redistributes subtitle text. The exported dataset and the web app contain derived statistics, titles, and AniList ids; no subtitle text, and not even the raw AniList descriptions. Non-commercial fan project.

Two ways to run the same UI:

python ap.py web            # local server at http://127.0.0.1:8000
python ap.py export-static  # precompute a static ./site for GitHub Pages (no backend)

The static build turns every API response into JSON and answers the same calls client-side, so it hosts anywhere as plain files.

Requires Python 3.11+ and an Anthropic API key (for the LLM stages).

python3 -m venv .venv && ./.venv/bin/pip install -r requirements.txt
echo "sk-ant-..." > api_key.txt        # or export ANTHROPIC_API_KEY

python ap.py collect          # AniList corpus + characters + tags (free, ~15 min)
python ap.py franchise        # cluster franchises
python ap.py extract          # Stage-1 lexicon candidates (free)
python ap.py estimate         # exact candidate count + projected batched cost (free)
python ap.py adjudicate       # Stage-2 Claude Haiku via Batch API (spends)
python ap.py measure          # precision/recall (Claude Opus gold + tag recall)
python ap.py analyse          # coverage report + zero-mention list
python ap.py export           # occupation_counts.csv, title_occupation.csv, evidence.csv

# Optional Layer C — subtitle dialogue:
python ap.py subtitles-prep   # clone the subtitle mirror + estimate
python ap.py subtitles-match  # match subtitle folders -> AniList ids
python ap.py subtitles-scan   # extract candidates (text never stored)
python ap.py adjudicate       # adjudicate the new subtitle candidates

Cost/scale reference: the full run analysed 21,988 anime and 128k subtitle files with ~101,000 Claude Haiku judgments (+ ~2,000 Opus spot-checks) for roughly $34 total, in about a day of mostly-unattended batch processing. Collection is resumable (checkpointed) and respects rate limits.

ap/                package
  taxonomy.py      ISCO-08 hierarchy (436 unit groups) + renderable/fantasy overlay
  lexicon.py       English surface forms per occupation + ambiguity flags
  jp_lexicon.py    Japanese surface forms (Layer C)
  db.py            SQLite schema + resumable checkpoints
  anilist.py       AniList GraphQL client
  collect.py       corpus + characters + tags + franchise clustering
  extract.py       Stage-1 lexicon candidate generation
  adjudicate.py    Stage-2 Claude Haiku / Batch API + free cost estimate
  measure.py       precision/recall (Claude Opus gold + tag recall)
  layerc.py        subtitle scanning (smart middle; text never stored)
  analyse.py       console coverage report + zero-mention list
  export.py        three CSVs
  webapp.py        local explorer (stdlib server + JSON API)
  export_static.py precompute the static site
ap.py              CLI entrypoint
web/index.html     the explorer frontend (vanilla JS, inline SVG charts)

Built with the Claude API. Data © their respective providers (AniList; ILO for ISCO-08). Code released for reference; the data shown is derived and non-commercial. Not affiliated with AniList, MyAnimeList, or any studio.