Inclusion: audit outcomes, design fixes
Introduced| English | Meaning |
|---|---|
| bias audit | a documented test of whether a decision system treats groups differently |
| adverse impact | a selection method that screens out a protected group at a higher rate |
Bias lives in outcomes, not intentions
- People processes decide who gets hired, praised, paid and promoted — and each can carry bias invisibly. The bias audit 偏见审计 compares outcomes across groups BEFORE adjusting for criteria.
- The screening audit: one postcode group passes the CV knock-out at 60%, the other at 36% — ratio 0.60, below the four-fifths line.
The four-fifths smoke detector
- Adverse impact 不利影响: a group's selection rate below 0.80 × the top rate. Here 0.80 × 60% = 48%; 36% < 48% — the line is crossed.
- Investigation: "fine-dining experience" as a knock-out functions as a postcode gate. Fix: trial days for all, CV trimmed to legal essentials — the gap closes to 52% vs 46%.
Pass rates 36% and 60%. The ratio, to two decimals?
Difficulty 2/5 · 140 coins + a small bonus on first completion today
$36/60 = 0.60 < 0.80$: the four-fifths line is crossed - a prompt for scrutiny, not a verdict by itself.
The top group passes at 60%. The four-fifths threshold rate, in %?
Difficulty 2/5 · 140 coins + a small bonus on first completion today
$0.80 \times 60\% = 48\%$; the observed 36% sits below it. The detector is beeping.
The comment field carries the bias
- Two crew members with caring responsibilities get "availability concerns" at three times the rate — with IDENTICAL covers and defect rates.
- Fix: comments must cite a measured criterion or be deleted. Make the prose as auditable as the scores.
Which frame - and which fix?
Route HR dilemmas to their ethical frame and audit findings to their process fix.
The knock-out listed "fine-dining experience". The right fix?
Difficulty 1/5 · 100 coins + a small bonus on first completion today
Fix the criterion: make the gate job-relevant and the gap closes (52 vs 46). Quotas treat the symptom and leave the gate biased.
Appraisal comments carry "availability concerns" x3 with identical numbers. The fix?
Difficulty 1/5 · 100 coins + a small bonus on first completion today
Make the prose as auditable as the scores - a comment without a criterion is a bias with a text box.
After the fix, pass rates are 52% and 46%. The ratio, to two decimals?
Difficulty 2/5 · 140 coins + a small bonus on first completion today
$46/52 = 0.88 > 0.80$ - above the line, and more importantly the gate is now job-relevant. The detector is quiet because the fire is out.
Ethical frames, named
- Utilitarian (greatest good), rights, justice — most HR dilemmas are frames colliding. The Saturday chef's second job: flex him (utilitarian) vs contract (rights) vs fairness (justice).
- The designed answer: a swap-board with transparent rules — flexibility real, justice met by ACCESS, not exception.
Match each dilemma reading to its frame.
Difficulty 2/5 · 140 coins + a small bonus on first completion today
Most HR dilemmas are frames colliding - naming them turns shouting into design.
A process is unbiased if the intent behind it was clean.
Difficulty 1/5 · 100 coins + a small bonus on first completion today
Bias lives in outcomes, not intentions - the audit reads the data, not the mission statement.
The bias audit compares outcomes across groups BEFORE adjusting for criteria. Why before?
Difficulty 1/5 · 100 coins + a small bonus on first completion today
Adjustments answer "is the gap explained?"; the raw gap asks "is there one?" Both questions, in that order - never only the second.
Three inclusion slips
- Intent as defence. Clean intent audits nothing; outcomes do.
- One audit, framed. The audit is a habit, not a project.
- The line as verdict or permission. Four-fifths prompts scrutiny — a smoke detector, not a judge.
Recap
- Audit outcomes by group; fix the criterion, not the quota.
- Comments and criteria must both be auditable.
- Named frames = dilemmas designed, not ducked.