Where we are, and what fairness means tonight
a1 drew the boundary, a2 mapped the law, a3 built the machinery: the vendor gate, the audit and notice processes, the six-part record. What the machinery is missing is a number, because governance with no measurement is a committee. Tonight fairness stops being a value statement and becomes a measurement discipline: a specific division, run on a specific population, on a cadence, by a named person, with a written consequence when it fails. That reframing is the whole session. Values do not survive a hiring wave. Numbers with owners do.
How a screen becomes biased with nobody writing a biased rule 7 min live
This is the part leaders most often get wrong, because they are looking for intent. There is rarely any. A CV is a document, a document contains what it contains, and anything visible on it is available to be scored. Some of what is visible tracks background rather than ability. Score it and you have a proxy, and the proxy does not care that you never meant it.
LiveThe chain, in three links4 min▶
Every proxy failure has the same shape, and once you can see the shape you stop needing to look for villains.
- Link one: it is visible. A CV shows a school, an employer, an address, a name, sometimes a photo, dates that imply an age, a timeline with or without gaps, and prose of a certain fluency. Anything on the page can be read, and anything readable can be scored.
- Link two: it travels with background. Which school you attended reflects where you grew up and what your family could afford. Employer prestige reflects who hired you before, which reflects the same chain one step back. Writing fluency tracks first language and whether anyone edited your CV. None of these are neutral facts about capability. They are facts about a life.
- Link three: it becomes a criterion. Someone writes "strong communicator" or "top-tier background" or leaves the criterion undefined entirely, and raters or models fill the gap with what they can see. The rule was never written. It was inherited.
The uncomfortable part for leaders is that a language model is unusually good at link three. It reads documents, it rewards fluent structured prose, and it has no way to know that fluency was not the job. A model asked to score "fit" with no rubric will produce a confident number built out of exactly these signals, and every individual result will look reasonable when you read it.
The tidy timeline. A screen quietly penalised employment gaps because "consistent progression" was on the criteria list and nobody defined it. The people it filtered hardest were those who had taken parental leave, cared for a relative, recovered from illness or been through a redundancy round. Nobody in the organisation would have written any of those as a screening rule, and everyone had signed off on a criterion that did the same work. The fix was two words: define "progression" as demonstrated growth in scope, evidenced from what the candidate submitted.
Self-studyRemoving the field is not the same as removing the signal2 min read▶
Anonymising a CV is worth doing and it is not a solution on its own. Strip the name and the postcode still implies it. Strip the postcode and the school still implies both. Strip the school and the phrasing of the personal statement still carries a signal. Proxies are correlated in bundles, which is why redaction alone plateaus.
What actually moves the number is the positive side: defining what the job requires, in job-related language, so specifically that a rater has something better to look at than the surface of the document. The b4 simulator makes this visible - "scrub the proxies" is one of five fixes, and on its own it buys back exactly one seat. Structure, evidence and calibration buy the rest. Removing a bad signal helps. Supplying a good one helps more.
The arithmetic, and the three things leaders get wrong 8 min live
The four-fifths rule is one division and you should be able to do it in your head in a meeting. Take each group's selection rate - selected divided by applicants in that group - then divide the lower rate by the higher. Below 0.80 is treated as evidence of adverse impact. That is the entire calculation. Everything difficult about it is what people do with the answer.
LiveKestrel's wave, by hand3 min▶
Not one slate. The wave: 300 applications across the first tranche of Kestrel's 120 roles, split by a characteristic Kestrel already monitors lawfully in aggregate.
| Group | Applicants | Advanced to interview | Selection rate |
|---|---|---|---|
| Group A | 180 | 54 | 30% |
| Group B | 120 | 18 | 15% |
| Ratio = 15% / 30% | 0.50 - fails | ||
0.50 against a 0.80 threshold. Note what this number is and is not. It is a screening signal drawn from the long-standing US selection guidelines, and it starts an investigation. It is not a verdict, it does not by itself establish that anything unlawful happened, and it is not the only fairness measure in existence. It is, however, the one a bias audit reports and the one you should be able to produce for any wave you run.
Nobody ran the division. A team used the same screen for eleven months. Every rejection had a rationale, every rationale looked reasonable on its own, and nobody ever divided one selection rate by the other, because the tool did not offer the number and nobody asked for it. The pattern had been there since month one and took about ninety seconds to see once someone finally looked. It was not a hard number to compute. It was a number nobody owned.
LiveThe three errors, in the order leaders make them3 min▶
- Treating a pass as a defence. A ratio of 0.86 tells you one screening signal came back clean on this population. It says nothing about whether your criteria are job-related, evidenced or defensible, and those are what you would actually have to justify. Passing the test is the beginning of the answer, not the answer.
- Running it on one small slate. Twelve interview seats means a single person swings the ratio by a wide margin. A small slate can pass by luck and fail by luck in consecutive weeks, and a team that only ever looks at slates will spend its time chasing noise. Aggregate the wave. Run the check per role for operational awareness and across the whole wave for the number that means something.
- Treating a fail as a verdict. A number under 0.80 is the start of an investigation: which stage, which criteria, which raters, which tool, and does it persist when you aggregate. Teams that treat it as a guilty finding either panic or start explaining it away in the room, and the second is much more common and much more damaging, because it teaches everyone that bringing you a bad number is expensive.
LiveThe trade-off you are braced for may not exist2 min▶
Most executives arrive at this conversation assuming fairness costs quality: that a fairer slate is a slightly weaker slate, and the organisation is being asked to pay a small price for doing the right thing. The b4 simulator, which your HR practitioners run live, is built to settle that assumption with arithmetic.
A slate scores 61 on quality with an adverse impact ratio of 0.09. Five fixes get applied one at a time - structured criteria, scrub the proxies, evidence per rating, calibrated raters, human decision gate - and the slate climbs to 100 quality and a 1.00 ratio. Both numbers move together, every time, in whichever order the fixes are applied. A sixth lever, the impact check itself, changes neither number: it reveals the ratio and blocks the slate. Measuring is not fixing.
The reading for leaders is blunt. Those five defects were not protecting quality, they were costing it. Every seat handed to polish was taken from someone better at the job. Structured criteria, evidence and calibration are selection quality practices that happen to be fairness practices, which is why this is a management argument and not only a compliance one.
The comfortable number and the uncomfortable criteria. One team ran the division, passed at 0.91, and did the second half of the exercise anyway: mark every criterion you could not evidence from what the candidate actually submitted. Four of their nine failed. They were rater impressions with numbers attached, and the ratio had passed on a sample small enough that it would not have kept passing. They had a clean signal sitting on top of a process they could not have defended for ten minutes.
Explainability, from the candidate's side of the table 6 min live
Every governance conversation so far has been internal. Now flip the chair around. A person applied, was rejected, and has asked why. What are they owed, and what does a useful answer sound like? This is the test your whole governance stack has to pass, because an organisation that cannot explain a rejection has not been governing anything - it has been documenting.
LiveWhat a useful answer contains3 min▶
Four elements, in plain language, without legal padding.
- Which criteria applied. The job-related requirements this role was assessed against. Not all of them in exhaustive detail - the ones that mattered for this decision.
- What evidence was considered. The application, the work sample, the interview answers. This is where the record from a3 earns its keep, and where an organisation without one goes quiet.
- Who decided. A role, at minimum, and a named person where your process allows it. This is rail 1 pointed outwards. The candidate should never be told that a system decided, because a system should not have.
- How to ask for a human review. An actual route: a named inbox, a stated response time, and a description of what will happen. An offer of review with no address behind it is worse than no offer.
Notice how much of this is already sitting in the six-part record. If your governance is real, this letter is a formatting exercise. If writing it requires an investigation, the letter is not the problem.
The template that changed the process. A team drafted their rejection template and got stuck on the second line, because they could not say what evidence had been considered for candidates who were cut at the automated stage. There was no evidence. There was a score. Writing a sentence to a stranger did what three governance workshops had not: it made the gap concrete. The template shipped six weeks later, after the stage it described had been redesigned.
Self-studyWhy the two bad answers are bad in different ways2 min read▶
"You were not a fit" is useless. It gives the candidate nothing to learn from, no way to check whether the process worked, and no route back. It is also, quietly, a confession that you may not have a better answer available. Its main cost is trust and reputation, and it is the industry default.
"The system scored you fourth" is worse, and it is worse in a way that should get a leader's attention. It is transparent, specific, and it names a machine as the decision-maker. That is the exact thing every framework in this course says must not happen, and saying it out loud to the affected person is an admission that the human oversight your governance claims does not exist in this case. Transparency is not the goal on its own. Transparency about a defensible process is.
The EU AI Act's high-risk obligations include transparency and human oversight for exactly this category of system, and where notice obligations apply, as under NYC LL144, candidates are already entitled to know a tool is being used before it is used on them. This is education, not legal advice, and what your organisation owes a rejected candidate varies by jurisdiction - so build the defensible answer as your default and let counsel tell you where more is required.
Run an impact review on your last five roles ★ 12 min · everyone builds
In aggregate, not role by role, and using only monitoring data your organisation already collects lawfully. You are not building a new dataset about individuals for a workshop exercise. If the data does not exist, that is your finding and it is a bigger one than any ratio.
Take your last five completed roles. For each, get applicants per monitored group and advanced-to-interview per monitored group. Aggregate the five before you divide anything.
Compute each group's selection rate, then divide the lower by the higher. Write the number down before anyone in the room interprets it. The order matters, because interpretation is where the explaining-away starts.
Now do it per role as well, and compare. If individual roles swing wildly around a stable aggregate, you are looking at small-sample noise. If the aggregate is low and steady, you are looking at a process.
Second half, and do not skip it even if the number passed: list every criterion used across those five roles and mark each one you could not evidence from something the candidate actually submitted. Those are where the polish got in.
If you could not get the numbers at all, stop and write that down as the primary output of the session. A process nobody can measure is a process nobody can defend, and it is also a process that cannot be independently audited.
Write your answer to "why was I rejected?" ★ 10 min · build your own
A template a recruiter could actually send on a Thursday afternoon without escalating to legal. Short, honest, and with a human-review route that goes somewhere real.
Draft it for one real role, not in the abstract. Use the criteria that role actually used.
Include all four elements: the criteria that applied, the evidence considered, who decided as a role or a name, and how to ask a person to look again.
Now stress-test it. Read every sentence and ask whether the record from a3 would let a recruiter fill it in within two minutes. Any sentence that needs an investigation marks a gap in the record, not a gap in the writing.
Name the review route properly: which inbox, who monitors it, what response time you commit to, and what the reviewer is actually empowered to change. "We will look into it" with nobody behind it is worse than silence.
Read it aloud as if you were the candidate. If it sounds like it was written to protect the organisation rather than to answer the person, rewrite it once more.
Try it yourself - this week ◐ 45-60 min total
- Finish the aggregate impact review across your last five roles, and put a name and a cadence against it so it happens again without you.
- Complete the criteria audit for those roles. Count how many criteria could not be evidenced from what candidates submitted. That count is a better health metric than the ratio.
- Rewrite the worst one properly. One criterion, from a proxy or a vague adjective into something evidenceable. Take the rewrite to the hiring manager who owns the role.
- Finish the rejection template and get one recruiter to complete it for three real past candidates, timed. Anything over two minutes is a record gap.
- Ask whether your organisation has ever explained a rejection to a candidate who challenged one, and what happened. Somebody remembers.
- Bring one unresolved data question to a5, where people data, monitoring and trust get their own session.
Official sources covered
Taught from public regulation and long-standing selection guidance. This course is education for HR professionals, not legal advice, and obligations vary by jurisdiction. No statement here is a legal conclusion about your process. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · Group A: 180 applicants, 54 advanced. Group B: 120 applicants, 18 advanced. What is the ratio?
Rates, then the division. 54/180 is 30%, 18/120 is 15%, and 15 over 30 is 0.50. Comparing raw counts skips the denominator, which is the single most common arithmetic slip in this whole exercise.
2 · Your wave passes at 0.86. What have you actually established?
Error one of the three. Passing is a signal, not a clean bill of health - which is exactly why the exercise has a second half where you mark every criterion you could not evidence. A comfortable number sitting on unevidenced criteria is a process you could not defend for ten minutes.
3 · In the b4 simulator, quality climbed 61 to 100 while the ratio climbed 0.09 to 1.00. What follows for leaders?
Both numbers move together because structured criteria, scrubbed proxies, evidence, calibration and a human decision gate are selection quality practices first. Every seat handed to polish was taken from someone better at the job. The sixth lever, the impact check, moves neither number - measuring is not fixing.