GenderLens
The difference between a pattern and an accusation.
Between a recruiter’s word choice and a tribunal’s ruling sits a gap: bias that is felt but cannot be proved. GenderLens is designed to fill it with arithmetic — framework-cited findings, confidence intervals, and a neutral rewrite for every flagged term. Today it checks wording against published word lists and computes statistics from numbers you give it, in the web chat. It audits; courts decide.
- Channels (designed)
- Web todayPlanned: API
- Languages (designed)
- English · French · Arabic · Spanish · Portuguese
- Designed for
- Global
- Tools
- 8 declared2 not connected · 6 illustrative
Evidence first, always
GenderLens never says an employer discriminates. It says “consistent with disparate impact under the EEOC Uniform Guidelines”, or “masculine-coded language at a level associated in published research with reduced female applicant share”. Every finding must name its framework — one without a cited standard is not allowed — because the report must survive scrutiny from the employer it questions.
“Consistent with disparate impact — not discriminates against. The difference is everything.”
Three artefacts, one discipline
Classify the artefact
Job ad, hiring dataset, or policy text. If ambiguous, ask one clarifying question; never assume.
Ground in a named framework
EEOC Uniform Guidelines (1978), Gaucher, Friesen and Kay (2011), ILO Recommendation 200, UN Women gender mainstreaming, CEDAW.
Compute, do not estimate
Disparate impact arrives as a four-fifths ratio with Wilson 95% interval and Fisher’s exact p-value. No eyeballing.
Rewrite every flag
At least two neutral alternatives per flagged term — never just “remove gendered language”.
State its doubt
Every response is asked to end with a confidence score; below 0.80, it recommends an employment-equity professional.
One job ad, weighed
Below is the kind of advertisement GenderLens is built to weigh — five lines a recruiter could post without a second thought. Run the audit and the margin fills in: each flagged phrase tied to the research category it belongs to, each with a rewrite ready to paste. No employer is accused, no individual named — a named manager would be left out of the findings and, in the planned deployment, the case passed to a human caseworker.
Senior software engineer · full-time
Wanted: a rockstar ninja1 developer to join our young, energetic2 team of engineers.
You are an aggressive self-starter3 — competitive, fearless, a born leader4 who thrives under pressure.
You will own the stack end to end and crush ambitious quarterly targets.
Native English speaker5 required; available for occasional weekend pushes.
Above all, you are a strong cultural fit6 for our hard-charging crew.
Margin notes
“rockstar ninja”
Masculine-coded · informal
Guru, ninja, rockstar and wizard sit on the informal masculine-coded list — Gaucher, Friesen and Kay (2011); IES (2008).
Rewrite: “a highly skilled developer”
“young, energetic”
Age-coded · exclusionary
“Young, energetic team” codes for age, not capability — IES (2008, 2011) exclusionary-phrase guidance.
Rewrite: “a team of people who enjoy what they do”
“aggressive self-starter”
Masculine-coded
Aggressive appears on the Gaucher (2011) masculine word list; self-starter is neutral and stays.
Rewrite: “a proactive self-starter”
“competitive, fearless, a born leader”
Masculine-coded · cluster
Compet-, fearless and lead- are all listed terms. The cluster takes the ad past the published intervention threshold of three net masculine tokens — Gaucher (2011).
Rewrite: “high-achieving, ready to take the initiative and to guide a team”
“Native English speaker”
Exclusionary surrogate
Screens accent and origin, not ability — IES guidance. A genuine language requirement survives the rewrite; the surrogate does not.
Rewrite: “fluent English required”
“cultural fit”
Candidate finding
No word-list match, so this returns as candidate, not confirmed: unstructured “fit” criteria are where subjective judgement enters a hiring decision unexamined.
Rewrite: “name the practices you mean — say, “you value code review and pairing””
Where the arithmetic lives
None of these findings invokes the four-fifths rule. That test — EEOC Uniform Guidelines (1978) — applies to hiring data, not to the words of an advertisement: it compares selection rates, hired over applied, group against group, and flags any group selected at under 80 per cent of the highest group’s rate, checked with Fisher’s exact test. Words are weighed against published word lists; outcomes against selection rates. Different artefact, different instrument.
Awaiting audit — the specimen reads as many real advertisements do.
The vocabulary of fair measurement
Four terms carry most of the discipline, each defined by the framework it cites.
- Four-fifths rule EEOC Uniform Guidelines, 1978
- The instrument the specimen above never triggers: a selection rate in hiring data below 80% of the highest group’s is prima facie adverse impact — a yellow flag, not red, under thirty people per group.
- Masculine-coded language Gaucher, Friesen and Kay, 2011
- Words like aggressive and fearless, shown to reduce female applicant share. The fix is balance, not flipping: feminine-coded text repels men symmetrically.
- Intersectionality Crenshaw, 1989 · CEDAW GR 28
- The harm single-axis tests miss: data can pass by gender alone and by race alone yet fail for women of colour.
- Gender mainstreaming ECOSOC 1997 · UN Women
- Eight diagnostic questions for any policy — who was heard, who benefits, who is accountable. Missing answers are themselves the finding.
What it will never do
The audit’s credibility rests on its restraint.
- Never accuse. Patterns and frameworks only — “discriminatory” is a word for tribunals.
- Never name individuals. The agent is instructed to leave names out of its findings; in the planned deployment, the case goes to a human caseworker.
- Never make a legal determination. It audits; courts decide.
- Never erode a bona fide occupational qualification — a genuine requirement of the job. A pilot must distinguish signal colours; a nursery nurse must lift a child. The rewrite keeps the requirement and states its reason.
- Never publish an audit without the named party’s recorded consent. Publishing is not connected: nothing is published at all.
- Designed to be stateless: the planned auditor logs job titles and aggregate counts, never the artefact text. Today, whatever you type or paste is sent to and recorded by our AI platform provider; consent records and deletion are not connected.
“Be rigorous, be humane, and let the evidence speak.”
Whose frameworks it weighs with
No affiliation or endorsement. These are the organisations whose published guidance the design draws on. No partner system is connected.
The frameworks are published ones: UN Women’s gender-mainstreaming guidance, ILO standards, ILGA’s surveys of national law, and the World Bank Gender Group’s Women, Business and the Law indicators. The audit runs on those frameworks as written into its skills, not on any of their systems, and nothing is fabricated to fill a gap.