Fair hiring starts with fair language.

How to use it

1. Paste a job advert into the home page — or give us a public URL and we'll fetch the readable text.
2. Click Analyse. You'll get a bias score, a list of flagged phrases grouped by category, and a neutral suggestion for each one.
3. Revise your draft using the suggestions. Paste the updated version back in and re-run. The score moves as you edit.
4. When you're happy, copy the findings or take a screenshot. Nothing is stored on our side — the text never leaves your browser tab except for the moment of the scan itself.

How it works

Under the hood, Unbias JD runs a curated lexicon across six bias categories: masculine-coded, feminine-coded, age, ability, ethnicity / culture, and elitism. Each term carries a severity (low, medium, high) and a neutral rewrite suggestion. The bias score is a function of how many high-severity terms appear relative to the length of the text — longer adverts are not automatically penalised.

The English lexicon is informed by Gaucher, Friesen & Kay (2011), "Evidence that gendered wording in job advertisements exists and sustains gender inequality," and by guidance from the UK Equality Act 2010. The Japanese lexicon is informed by 厚生労働省 "公正な 採用選考" guidance, which discourages age, gender, physical, and background-based language in recruitment.

A contextual LLM pass via the Vercel AI Gateway is available for picking up implicit or framing-level bias that a lexicon cannot see. It is disabled in the public build to keep the tool free; it may be enabled later as the project grows.

Limitations

• A lexicon will miss context. "He" is flagged; "he or she" still reads as gendered but the phrase as a whole may not trigger. Use the tool as a first pass, not the final word.
• Unbias JD is a writing aid. It is not legal advice and does not certify compliance with any specific regulation.
• The score is directional. Going from 70 to 90 means your draft is measurably more neutral; the absolute number is not a grade.

Privacy

No accounts, no cookies used for tracking, no database. Text is sent to the server only for the duration of the scan and is not persisted anywhere. URL fetches are proxied through an SSRF-protected endpoint that blocks private IP ranges and cloud metadata.