Selection, Artificial Intelligence, Ethics, Bias and Inclusion
Competency-based interviews, evidence, rubrics and responsible judgement with AI
4 original sources
01. Summary
Start here
If you only remember five things from this topic:
- A pleasant impression is not evidence of competence. Role → competencies → observable indicators → questions → rubric → decision.
- Behavioural = real past. Situational = hypothetical. Follow up without suggesting the answer.
- Structured interview + rubric make candidates comparable. They reduce bias; they do not remove it.
- AI needs human oversight: a person who can understand, question and change the decision.
- Three readings: McKinsey (value, but tools are not enough), Moser (judgment is not reckoning), Snyder and Mann (algorithmic divide).
Each section has an exam paragraph: a short answer in English that you can adapt in the exam.
From the role to a selection decision
- RoleThe outcomes and competencies it needs
- IndicatorsBehaviours you can observe
- QuestionsAbout the past or about a hypothetical situation
- RubricCriteria fixed before you score
- DecisionReasoned and open to review
A pleasant impression is not evidence of competence. "I liked the candidate" says nothing about what the person did. Each question must be connected to something the role needs.
Competency and observable indicator
- Competency = knowledge + skills + behaviours.
- Observable indicator = what you would see a person do if the person has the competency.
| Competency (Activity 2, Talent Acquisition Specialist) | Observable indicators |
|---|---|
| Communication skills | Listens actively, adapts the explanation, checks that the other person understood |
| Problem-solving and decision-making | Identifies the problem, compares alternatives, uses data |
| Teamwork and collaboration | Listens to other opinions, shares responsibilities, manages disagreement |
| Ethics, fairness and inclusion | Applies role criteria, reviews bias, protects information |
Role: shop assistant in a clothes shop. Competency: customer orientation. Write three observable indicators.
Model answer
- Greets the customer and asks what the customer needs before offering a product.
- Proposes alternatives when the size or colour is not available.
- Solves a complaint without arguing and offers a specific solution.
Each indicator is a behaviour you can see, not an adjective ("nice", "kind").
Behavioural and situational interviews
| Type | It asks about… | Example |
|---|---|---|
| Behavioural | A real past experience | "Tell me about a time when you managed a disagreement. What did you do?" |
| Situational | A hypothetical scenario | "Two team members disagree about the final selection. How would you organise the decision?" |
A behavioural question gives evidence of what the person did. A situational question shows how the person thinks. It does not prove that the person has done it.
- Context"What was the situation?"
- Own responsibility"What was your part?"
- Action"What did you do, exactly?"
- Result"What happened? How did you check it?"
- Learning"What would you do differently?"
Competency: problem-solving for an IT support technician. Write one behavioural question and one situational question.
Model answer
- Behavioural: "Tell me about a time when a user had a problem you did not know how to solve. What did you do and how did it end?"
- Situational: "On a Friday at 6 p.m. the email of the whole office stops working. Which steps would you follow?"
Possible follow-up: "What did you do, and what did the team do?" · "How did you know it was solved?"
A behavioural question asks the candidate to describe a real past experience, for example "Tell me about a time when you managed a disagreement". It gives evidence of actual behaviour. A situational question presents a hypothetical scenario, for example "What would you do if two colleagues disagreed?". It shows reasoning, but it does not prove that the person has acted in that way. A good interview combines both and uses neutral follow-up questions about context, action, result and learning.
Structured interview and rubric
Structured interview: all candidates answer about the same competencies, with the same questions and the same criteria. This makes the answers comparable.
Equal treatment does not mean you ignore accessibility. If a person needs a different format or more time, adapt the conditions. Do not change the competency you assess.
| Points | Level | Evidence |
|---|---|---|
| 1 | Poor | No relevant evidence, or inadequate behaviour |
| 2 | Below expectations | Limited evidence or an ineffective answer |
| 3 | Meets expectations | Adequate evidence and a reasonable approach |
| 4 | Above expectations | Solid evidence and effective behaviour |
| 5 | Excellent | Complete evidence, initiative and good judgement |
- RecordWrite down behaviours, decisions and results.
- RelateLink each piece of evidence to a competency.
- ScoreApply the rubric.
- ReviewJustify differences and check for bias.
Question: "Tell me about a time when you managed a disagreement in your team." Answer: "Two colleagues argued about the schedule. I spoke with each of them separately. I compared the two proposals with the customer's delivery dates. We agreed on one schedule with an adjustment from the other. The project was delivered on time." Which score would you give? What would you ask first?
Model answer
Evidence: listening (spoke with each one), use of criteria (customer dates), agreement (one schedule) and result (delivered on time).
Indicative score: 4 (above expectations). Solid evidence and effective behaviour.
Before you score, ask: "Which part of the agreement did you propose?" and "How was the relationship between the two colleagues afterwards?" If the answer shows initiative, it could be a 5.
In a structured interview all candidates are assessed on the same competencies, with the same questions and the same criteria, so their answers can be compared. A rubric defines the levels of performance before the interview, for example from 1 (poor) to 5 (excellent). The interviewer first records the evidence, then relates it to the competency, scores it and reviews the score. In this way the decision is based on observed behaviour and can be explained and reviewed.
Bias, fairness and inclusion
| Bias | How it appears | How to control it |
|---|---|---|
| Affinity | You prefer a person with your background or hobbies | Go back to the role criteria |
| Halo | One excellent answer raises all the scores | Score each competency separately |
| Confirmation | You only look for what confirms the first impression | Look for contrary evidence too |
| Stereotype | You infer ability from a personal characteristic | Check the relation with the role |
| Historical data | A system learns from a process that excluded people | Review the data and the definition of "success" |
| Automation | You accept a system's score without review | Real human review, with power to change the decision |
The same questions for everybody help. They do not remove all bias.
Case: an unfair request to exclude a candidate
In the activity, a manager wants to reject a qualified person because of age or origin.
- Keep the role criteriaAssessment is based on competencies, not on personal characteristics.
- Question the requestIn a professional way.
- DocumentWrite down what happened.
- Use the correct channelsHR, compliance or the channel the company provides.
- Protect the informationKeep the candidate's data confidential.
Bias distorts selection decisions. With affinity bias the interviewer prefers people who are similar. With the halo effect one strong quality raises all the scores. With confirmation bias the interviewer only looks for evidence that supports the first impression. Stereotypes infer ability from personal characteristics. Role criteria, separate scores for each competency and a rubric reduce these risks, but they do not eliminate them. Fair selection assesses the competencies of the role and nothing else.
AI in selection
AI can process information and support decisions. But a prediction is not an understanding of the person or the context.
- PurposeWhich specific problem does it solve? Is it necessary to automate this decision?
- RepresentationWhich data does it use and what does it leave out?
- CriterionIs its measure of "success" related to the role?
- ImpactWho benefits, who can be excluded and which errors matter?
- ResponsibilityWho reviews, corrects and can stop or change the decision?
Human oversight means that a person can understand, question and change the decision. To sign what the machine says is not oversight.
A system recommends interviewing only the candidates with more than 87 points. It was trained with the hires of the last ten years. What would you check before you use it?
Model answer
- Purpose: what does the score represent? Does it predict an important outcome of the role?
- Data: past hires can contain bias. If the company hired mainly men, the system can learn that "men = success".
- Errors: who is below the threshold? Which information is lost?
- Responsibility: is there a person who reviews and can change the decision?
- Judgement: is it sensible to automate this step?
The same threshold for everybody does not make the system fair if the training data was unfair.
AI can help recruiters to process applications faster, but it brings two risks. Historical data bias appears when a system learns from past decisions that already excluded some groups. Automation bias appears when people accept a score because a machine produced it. Before using a tool, the company should examine its purpose, the data it uses, its criterion of success, its impact and who is responsible. Human oversight is essential: a person must be able to understand, question and change the decision.
AI in workforce planning (McKinsey, 2024)
The article studies travel and logistics companies (airlines, hotels, logistics providers). They face labour shortages, rising labour costs and productivity that grows slowly. They are rich in data but plan with spreadsheets and simple models.
- Digital demand reductionRemove work before it reaches frontline employees: self-service, chatbots, automation of repetitive tasks. 15–20% fewer calls.
- Resource and capacity planningMore accurate demand forecasts and dynamic scheduling. More than 90% accuracy in call-centre staffing; overtime costs down 15–20%.
- Hiring and people developmentIdentify candidates, speed up hiring, plan training with simulations (digital twin). An airline cut pilot-training bottlenecks by 5–10%.
- Day-of decision makingGive local teams standards, frameworks, playbooks and real-time data. Daily route optimisation: 15% less travel time.
- Continuous improvementClear standards, visible metrics, coaching and change management. A railroad moved from resistance to use of its routing tool.
Tools and analytics alone are not sufficient. The best tool fails without process changes, performance management and change management. Culture is "the true engine" of the transformation. The article also warns that AI assessments can introduce bias in hiring.
The railroad case shows the people side. The routing tool had low usage because of technical problems and fear of job displacement. The company answered with a root-cause analysis, more training and a clear change story. Results: 40–50% more trained dispatchers and 25–30 percentage points more usage.
Change management starts with a compelling change story: why the change is needed, what changes, what is expected, and training so that employees can adopt it. Leaders role-model the change.
McKinsey argues that AI can transform workforce planning in travel and logistics companies, which suffer labour shortages, rising costs and weak productivity. It proposes five actions: digital demand reduction, resource and capacity planning, hiring and people development, day-of decision making and continuous improvement. Its main lesson for People Management is that tools alone are not sufficient. They only create value with new processes, performance management and change management, because employees must trust and adopt them.
AI, judgement and ethics
What Humans Lose When We Let AI Decide (Moser, den Hond and Lindebaum, MIT Sloan, 2022)
Judgment · human
Considers the social and historical context and the possible outcomes.
Uses imagination, reflection, examination, valuation and empathy.
Has an intrinsic moral dimension.
Reckoning · machine
Processes data with calculus, computation and rule-driven rationality.
Is constrained by the narrow objectives for which it was designed.
Has "an air of precision", but it is a selective representation.
Not just "bias"
To call the limits "bias" suggests AI can one day remove them. The authors say the problem is deeper: AI will never be capable of judgement, only of reckoning.
The map is not the territory
AI decides on "technical images": incomplete representations of the world. Drivers in Marseille followed the GPS into the harbour.
Real harm
A flawed AI system for child-benefit fraud harmed about 40,000 families in the Netherlands. The government resigned in January 2021.
"All models are wrong, but some are useful" (Box)
If you do not know how a model is wrong, it is dangerous to use it.
Call for inaction
Stop automating everything that can be automated. This trend is technological solutionism (Morozov).
Call to action
Learn what AI is. Managers should judge case by case whether, how and why to use it, with vigilance and doubt.
The Looming Algorithmic Divide (Snyder and Mann, Wharton)
The algorithmic divide goes beyond the digital divide. It comes from (1) unequal access to technology and AI literacy and (2) cognitive biases in AI models trained on available data. In the US, 15 cities hold two-thirds of AI assets and capabilities.
Biases cannot be eliminated
Confirmation, survivor and availability bias are part of human cognition. To remove one bias often introduces another. Computing power amplifies them.
Alter ego AI
Each AI creates a counterpart AI (dual-sided AI). Human agency must stay central.
Guidepost 1 · AI readiness
Educate leaders and employees (AI literacy), build trust, pilot quick wins with human-centric design.
Guidepost 2 · AI governance
Responsible-AI policies with legal, compliance, HR, procurement and IT. Example: Rolls-Royce's Aletheia Framework, 32 dimensions.
Guidepost 3 · Sustainable investment
Transparency, an internal AI sandbox, explainable AI, data protection, less energy use.
The three readings together
| Reading | Central question | What it gives you |
|---|---|---|
| McKinsey | How do we get value from AI in workforce planning? | Tools + processes + people + adoption |
| Moser, den Hond and Lindebaum | What do we lose if reckoning replaces judgement? | A model is a limited representation. Decide case by case |
| Snyder and Mann | Who has access to AI and who carries the risks? | Readiness, governance and fair distribution of benefits |
Moser, den Hond and Lindebaum distinguish judgment from reckoning. Human judgment considers context and consequences and has a moral dimension. Reckoning is the calculation that algorithms perform on a limited representation of reality. The danger is that we accept the precise answer of the machine as the correct one and stop using our own judgment. The authors recommend not automating everything that can be automated and deciding case by case, with vigilance and doubt, whether AI should be used.
How to write the answer
- Role and competencyWhat must be assessed and why it matters.
- InstrumentBehavioural question, situational question or other evidence.
- CriteriaIndicators and rubric before you score.
- Bias and inclusionWhich risks exist and how you review them.
- AI and judgementWhat the tool can do and which responsibility stays with the person.
- LimitsWhat we know, what is missing and what we cannot conclude.
| Idea | Sentence you can write |
|---|---|
| Evidence | "A pleasant impression is not evidence of competence." |
| Behavioural | "A behavioural question asks for a real past example." |
| Structure | "A structured interview applies the same questions and criteria to all candidates." |
| Halo | "To reduce the halo effect, each competency is scored separately against the rubric." |
| Historical data | "If the system was trained on biased past hires, it will reproduce that bias." |
| Oversight | "Human oversight means that a person can understand, question and change the decision." |
| Judgement | "AI is capable of reckoning, not of judgement, so the final decision must stay with a person." |
| Adoption | "Tools alone are not sufficient; they need process changes and change management." |
A logistics company wants to hire 20 warehouse managers. It receives 1,500 applications. It wants to use an AI program that filters the CVs and to interview only the top 60. Write an answer with the six steps.
Model answer
- Role and competency: team leadership, shift organisation, problem-solving, safety.
- Instrument: structured interview with behavioural questions ("Tell me about a time when you reorganised a shift at short notice") and situational questions ("A truck arrives two hours late. What do you do?").
- Criteria: indicators for each competency and a 1–5 rubric fixed before the interviews.
- Bias: affinity and halo in the interviews. Historical data in the AI program: if the company hired mainly men, the filter can exclude women.
- AI and judgement: the AI can order the CVs, but a person must review a sample of the rejected candidates and must know which criteria the program uses.
- Limits: we do not know which data trained the program. This must be checked before it is used.
Links with the other topics
| Topic | Connection |
|---|---|
| T1 | Selection applies the same rule as diagnosis: evidence first, conclusion later. |
| T2 | The competencies come from job design; the AI readings extend workforce planning. |
| T4 | The interview is also a moment of truth for the candidate; the same biases return in performance appraisal. |
Typical exam mistakes
Friendly = competent
You need an observed behaviour.
Behavioural = situational
Past vs "what would you do".
Leading questions
"You did it well, didn't you?" is not neutral.
Absence = incapacity
Maybe you need another question.
Same questions = no bias
It helps, but it does not remove all bias.
Precise = fair
A precise score can come from unfair data.
Oversight = signature
Oversight is the power to understand and change the decision.
Judgment = reckoning
Judgement looks at context and morality. Reckoning calculates.
Cheat sheet
| They ask… | Remember… |
|---|---|
| Selection sequence | Role → competencies → indicators → questions → rubric → decision |
| Observable indicator | What you would see a person with the competency do |
| Behavioural | Past experience and real behaviour |
| Situational | Hypothetical scenario: "what would you do?" |
| Follow-up | Context → your action → result → learning |
| Structured interview | Same competencies, questions and criteria for everybody |
| Rubric | 1 poor · 2 below · 3 meets · 4 above · 5 excellent |
| Human biases | Affinity, halo, confirmation, stereotype |
| Machine biases | Historical data, automation |
| Five AI questions | Purpose, representation, criterion, impact, responsibility |
| McKinsey: five actions | Demand reduction, capacity planning, hiring and development, day-of decisions, continuous improvement |
| Judgment / reckoning | Human, with context and morality / calculation on data |
| Technological solutionism | To automate everything that can be automated |
| Algorithmic divide | Unequal access to AI and AI literacy, plus biased models |
| Three guideposts | AI readiness, AI governance, sustainable investment |
02. Test
12 easy, 14 medium and 14 hard. The order changes every time.
03. Cards
Tap the card to flip it and mark whether you know it. Your progress is saved.