About & methodology

Is My Job Safe? is a small static tool that matches a job title you type to the closest occupation category in Anthropic's own published research on AI usage, and shows you the real numbers for that category — nothing scored or predicted by this site itself.

Is this relevant to Indian roles?

This result is built on Anthropic's Economic Index — real, published data on how millions of Claude conversations map to occupations worldwide. Job tasks look the same everywhere: what an accountant, a bank clerk, or a software developer actually does day-to-day doesn't change much by country — so these patterns are directly relevant to Indian roles, even though the underlying usage data isn't collected India-only.

Where the data comes from

All of it comes from the Anthropic Economic Index dataset on Hugging Face, released under CC-BY-4.0. Specifically:

Field	Source file
Occupation title, SOC code, exposure score	labor_market_impacts/job_exposure.csv
Occupation hierarchy / category names	release_2025_02_10/SOC_Structure.csv
O*NET task text per occupation	release_2025_02_10/onet_task_statements.csv
Global usage share per task	release_2025_02_10/onet_task_mappings.csv
Automation/augmentation category definitions	release_2025_09_15/data_documentation.md
Task × collaboration-pattern counts (for the automation/augmentation split)	release_2025_09_15/data/output/aei_enriched_claude_ai_2025-08-04_to_2025-08-11.csv

Nothing is fetched at runtime — a build script processes these CSVs offline into one JSON file the site ships with. No API calls, no backend, no per-search cost.

What "exposure" means

The exposure score is Anthropic's own observed_exposure figure, used exactly as published, for the closest matching occupation. It's a measure of how much a role's day-to-day tasks show up in real, measured Claude conversations, weighted toward tasks a language model can plausibly help with — combining theoretical task feasibility with actual observed usage. It is not a probability of job loss, and a low or zero score doesn't mean a job is "safe" — it may simply mean that kind of work hasn't shown up much in Claude usage yet, or doesn't translate well into a text-based AI conversation (hands-on physical work, in-person service, and similar).

What the automation/augmentation split means — and how it's actually calculated

This part is worth being precise about, because Anthropic does not publish a ready-made automation-vs-augmentation percentage per occupation. The only per-occupation figure they publish directly is the exposure score above. The split shown on each result page is derived by this site, using Anthropic's own category definitions, as follows:

1. Every occupation's real O*NET task list is pulled from onet_task_statements.csv.
2. For each of those tasks, real conversation counts by collaboration pattern (directive, feedback loop, task iteration, learning, validation) are pulled from Anthropic's enriched Claude.ai usage file — one week of global data, 4–11 Aug 2025.
3. Patterns are classified into "automation-leaning" (directive, feedback loop) or "augmentation-leaning" (task iteration, learning, validation) — this is the exact split Anthropic documents in their own data dictionary, just applied at the task level instead of only as a single global number.
4. Occupations with fewer than 30 classified conversations across their tasks show no split at all, rather than a number computed from too little data.

This means the split is real data, correctly categorized by Anthropic's own definitions — but it's a one-week global snapshot, not a long-run trend, and about 46% of the 756 occupations don't have enough signal to show a split (they still show an exposure score, which comes from a separate, more robust file).

How job-title matching works

Your input is checked, in order, against: an exact match in a hand-built synonym list of common Indian job-title phrasing (things like "bank PO" or "SDE"); an exact match against a formal O*NET occupation title; and finally a fuzzy text match (combining word overlap and typo tolerance) across both. If nothing scores confidently enough, you get the closest few categories to choose from instead of a guess. All of this runs client-side, in your browser.

Full data documentation

The build script, its full source-to-field mapping, and every judgment call (including the SOC-code-vintage mismatch between dataset releases and how it's handled) are documented in the project's data-pipeline/README.md for anyone who wants to audit or reproduce the numbers.

Contact

Questions about this site or its methodology can be sent to the site operator — add a real contact email here before launch.