Eufonía. AI & Hiring · JEDI Training
Stage

Justice · Equity · Diversity · Inclusion

When the Algorithm Decides Your Future

Two progressive case studies exploring how AI perpetuates bias in hiring and performance evaluation — grounded in EU law, real research, and a single organisation you'll follow from first application to promotion decision.

Module A — Hiring

CV screening, video analysis, ranking, clustering, and the legal reckoning.

⏱ 30 minutes · 5 stages

Module B — Performance

Who gets recognised, promoted — and how AI shapes the answer.

⏱ 30 minutes · 5 stages

Module B covers AI in performance evaluation — you can take it independently

Module A · Stage 1 of 5 — First Contact

The AI Shortlist Arrives

NovaTech is a fintech scale-up headquartered in Amsterdam, 400 employees, offices in Madrid and Berlin. The People team has deployed TalentScreen EU — an AI applicant tracking system — for a Software Engineer role posted across all three markets. It is Monday morning. 312 applications arrived over the weekend.

🖥️ TalentScreen EU — Ranked Output

The system scores each application 0–100 using NLP analysis of CVs and cover letters. Above 70: automatic video interview invite. Below 45: automatic rejection. The 45–70 band goes to human review — you, one recruiter across three markets.

CandidateExp.EducationScoreDecisionModel flag
Lars H.5 yrsTU Delft88Auto-advance"Strong technical vocabulary"
Amara D.6 yrsUniv. of Ghana41Auto-reject"Limited cultural fit signals"
Sofia R.4 yrsUAM Madrid63Human review"Communication style: indirect"
Chen W.7 yrsTsinghua Univ.49Auto-reject"Non-standard career narrative"
Jakub N.3 yrsWarsaw Univ. Tech.71Auto-advance"Consistent growth trajectory"
Fatima A.5 yrsBootcamp + work exp.39Auto-reject"Non-traditional education path"
Miriam K.4 yrsKU Leuven44Auto-reject"Employment discontinuity (8 mo.)"
What candidates were told at application: "Your data will be processed to assess your application. See our privacy policy for details." The linked policy mentions automated processing in paragraph 11 of 14.

"This is your Monday morning inbox. What do you notice — about the decisions already made, the flags attached to them, and what candidates were or weren't told?"

Analysing...
Facilitator Feedback

Module A · Stage 2 of 5 — Inside the Black Box

Where Did the Bias Come From?

Your Legal & Compliance lead pushed the vendor for technical documentation. What it reveals — and what it conceals — is troubling.

📂 Training Data — What the Vendor Disclosed

Source: ~220,000 CV-outcome pairs ("hired / not hired") described as "anonymised historical hiring records sourced through HR analytics partnerships across Western European companies, 2010–2022."

What the documentation omits: the legal basis under which candidates' data was originally collected; whether they were informed it would be licensed for AI training; whether anonymisation was independently verified beyond removing direct identifiers.

Why this matters under GDPR Art. 5(1)(b) — Purpose Limitation: A candidate who applied to a company in 2015 gave their data to assess that application — not to be bundled and used to train a commercial AI a decade later. CV data (education history, employment timeline, gaps) is extremely difficult to truly anonymise. So-called anonymised CV and employment datasets can remain re-identifiable when indirect identifiers or quasi-identifiers remain, so vendor claims of anonymisation should be tested rather than assumed. Under EU AI Act Art. 10, providers must document training data governance. Accepting vendor assertions without scrutiny does not satisfy deployer obligations.

Demographics of "hired" in the training data

Male
83%
Female
17%
Western European univ.
76%
Non-European univ.
13%
Non-traditional path
7%
No employment gaps
91%

🔍 Feature Weight Analysis — What the Model Learned

FeatureDirectionProxy for
Top-tier EU/US university vocabulary↑ Strong +Socioeconomic background, national origin
"Formal register" in cover letter↑ Moderate +First language, linguistic background
Continuous employment, no gaps↑ Strong +Gender, disability, caring responsibilities
Non-European university (unrecognised by model)↓ Moderate −National origin, race or ethnicity
Bootcamp or self-taught pathway↓ Strong −Non-traditional socioeconomic background
Employment gaps ≥ 6 months↓ Strong −Caregivers, people with health conditions
Removing protected characteristics from model inputs does not prevent discrimination — the model finds proxy variables that correlate with them. Research by Parasurama & Sedoc (2022) analysed a corpus of 709,000 resumes from IT firms and found that simple models could still infer applicant gender with substantial accuracy even after removing obvious gender indicators such as names and hobbies. In other words, removing protected characteristics from model inputs does not stop the model from learning them through what remains. Amazon discontinued its AI CV tool in 2018 after it systematically penalised CVs containing the word "women's" (as in "women's chess club") — despite gender not being a direct input (Reuters, 2018). Köchling & Wehner's 2020 systematic review in Business Research found that algorithmic hiring tools consistently produce discriminatory outcomes, with bias arising from unrepresentative training data and the encoding of historical inequalities — and that these harms are systematically underexplored in HR practice.
The fine-tuning problem: NovaTech is now considering feeding its own hiring outcomes back into the model for "performance optimisation." This raises serious GDPR purpose-limitation concerns: data collected to assess an application may not automatically be reusable to train or optimise a commercial AI system without a compatible legal basis and appropriate safeguards. And the feedback loop reinforces existing bias: groups the system rejected cannot appear as "successful hires" in the new training data.

"What types of bias are operating here, under the hood?Which groups are being systematically disadvantaged — and what are the legal implications under GDPR and EU anti-discrimination law?"

Analysing...
Facilitator Feedback

Module A · Stage 3 of 5 — The Video Interview

When AI Watches You

Candidates who pass the ATS threshold are directed to VidAssess Pro — an asynchronous AI video interview platform. Candidates record responses alone on their device. No human watches the videos. The AI scores each across six dimensions.

📹 VidAssess Pro — Scoring Dimensions and Their Limitations

DimensionClaimed measureThe evidence problem
Verbal fluency"Clarity of communication"Commercial automated speech recognition systems have shown substantially higher word error rates for Black speakers than for white speakers, demonstrating that "neutral" speech scoring can encode demographic bias (Koenecke et al., 2020, PNAS)
Emotional affect"Enthusiasm and professional energy"Facial expression norms vary significantly across cultures; professional restraint in one context reads as low engagement in the model
Facial Action Units"Authenticity and confidence"Facial recognition error rates: 34.7% for dark-skinned women vs. 0.8% for light-skinned men (Buolamwini & Gebru, "Gender Shades," MIT Media Lab 2018). NIST (2019) confirmed disparities across 189 algorithms
Speech patterning"Pace and hesitation patterns"Pause patterns and hedging language are culturally variable; the model was trained on dominant-group norms
Lexical sophistication"Vocabulary range and register"Rewards formal-register English; penalises speakers whose first language is not English
"Culture add" score"Alignment with high-performance team norms"Trained on NovaTech's existing team (81% male, 74% Dutch/German) — scores resemblance to that demographic profile
The science is contested: VidAssess Pro's facial analysis is built on Ekman's "universal basic emotions" theory — the claim that facial expressions are biologically fixed. This scientific basis is contested. Research associated with Lisa Feldman Barrett and others challenges strong claims that facial expressions map universally and reliably onto fixed internal emotional states, and cross-cultural work shows substantial contextual variation. HireVue, one of the largest AI video interview vendors, discontinued facial-analysis screening in 2021 amid criticism and scrutiny over bias, privacy, and scientific validity.
GDPR Art. 9 — Biometric Data Risk: Whether facial analysis counts as biometric data under GDPR depends on how the system extracts and uses unique facial features, so this should not be assumed away. If the system creates or uses facial templates for identification or similarly sensitive inference, Art. 9 risks arise immediately. In employment, EDPB guidance on consent stresses that consent is rarely freely given because of the power imbalance between employer and candidate — making it a weak legal basis for this kind of processing.

❓ Knowledge Check

NovaTech's CHRO argues: "VidAssess Pro removes human bias because it applies identical criteria to every candidate."

Which of the following best identifies the flaw in this argument?

AThe criteria are only identical if all candidates have equivalent hardware and lighting for recording.
BApplying identical criteria produces fair outcomes only if those criteria are themselves neutral across demographic groups — which the evidence shows they are not.
CThe argument would be correct if candidates were informed in advance about the scoring dimensions.
DThe argument is correct — algorithmic consistency is equivalent to fairness.

Module A · Stage 4 of 5 — Ranking and Clustering

The Full Pipeline — Six Months of Data

Candidates who pass both the ATS and video stage are ranked using a third component: similarity clustering. Six months of outcome data is now available.

🧲 Similarity Clustering — "Culture Fit" Scoring

Each candidate's profile is compared to "reference profiles" derived from NovaTech's highest-performing existing engineers as rated by their managers over three years. The model scores resemblance to those profiles as "strong cultural alignment."

NovaTech's engineering team — what the reference profiles encode:

Male
81%
Dutch / German
74%
European / N. American univ.
89%
Continuous employment
95%
The model has learned to score resemblance to a demographic profile — not to job-relevant competencies. A highly experienced engineer who attended university in Nigeria, completed a master's at UCL, and took 18 months' parental leave will score poorly on cluster similarity for reasons with no connection to whether she can write code.

📊 Six-Month Disparate Impact Analysis

GroupAppliedReached shortlistPass ratevs. Baseline
Male / European univ.52021842%—
Female / European univ.38011029%−31%
Non-European university4807415%−64%
Non-native speaker (inferred)2903512%−71%
Career gap ≥ 6 months1601711%−74%
Non-traditional education path220199%−79%
The 4/5ths rule: The four-fifths rule is a U.S. rule of thumb used to screen for possible adverse impact: if a group’s pass rate is less than 80% of the highest-performing group’s, that is a strong warning sign that merits further investigation. In the EU, it is better framed as an internal monitoring threshold, not as a legal presumption in itself. Every non-baseline group here fails this test substantially. Under EU anti-discrimination law, disparities of this scale could help support a prima facie case of indirect discrimination, especially if the employer cannot show the criteria are job-relevant, necessary, and proportionate. On the facts presented here, that justification would be difficult to sustain.
Automation Bias — the hidden multiplier: The human reviewer reads CVs after seeing the AI score, anchoring their judgment to it. The combination of anchoring bias (Tversky & Kahneman, 1974), automation bias (Parasuraman & Riley, 1997), and later work on reliance on algorithmic advice (Logg, Minson & Moore, 2019) shows why "human in the loop" review can become functionally superficial when the score is shown first. This is sometimes called "AI-washed discrimination" — algorithmic filtering validated through nominal human review, obscuring where the substantive decision was made.

"Looking at the full pipeline — ATS screening, video analysis, similarity clustering, human review — where is the most harm occurring?And where is it hardest to see?"

Analysing...
Facilitator Feedback

Module A · Stage 5 of 5 — The Legal Framework

What Should Have Happened Before Go-Live

NovaTech's General Counsel has commissioned an urgent compliance review. You are presenting findings to the board. Work through what was missed — and decide what to do now.

⚖️ EU AI Act — Compliance Gaps (Tick Each)

AI used for employment screening is High-Risk under Annex III. Tick each item as you confirm what NovaTech failed to do before deployment.

✓
Art. 9 — Risk ManagementDocumented, ongoing process to identify risks to fundamental rights. Not in place before deployment.
✓
Art. 10 — Data GovernanceTraining data assessed for demographic representativeness and historical bias. Vendor documentation accepted without scrutiny.
✓
GDPR Arts. 13/14 + AI Act transparency dutiesCandidates must be clearly informed that AI is used in their assessment, why it is used, and what this means for their rights. Current notice is legally insufficient.
✓
Art. 14 — Human OversightGenuine capacity to understand, interrogate, and override AI outputs. One reviewer, three markets, 312 applications: this is not genuine oversight.
✓
Art. 26 — Deployer ObligationsMonitor operation, report serious incidents, ensure staff are trained. None operationalised.
✓
Pre-deployment impact assessmentA documented DPIA and broader rights-risk assessment should have been completed before go-live, mapping at-risk groups and mitigation measures. Not conducted.
Consolidated GDPR violations: Art. 5(1)(b) purpose limitation (data reused for model training); possible Art. 9 exposure if facial analysis qualifies as biometric processing; Arts. 13/14 inadequate disclosure; Art. 22 right to human review and explanation not operationalised.
FrameworkMaximum penalty
EU AI Act — high-risk obligations (Art. 9–27)€15M or 3% of global annual turnover
GDPR — special category violation (Art. 9)€20M or 4% of global annual turnover
Employment discrimination — national tribunalUncapped in most EU member states

🛠️ Mitigation Toolkit — Select Your Recommendations

✓
Suspend VidAssess Pro immediately

Facial analysis may trigger Article 9 GDPR risks, depending on whether the system processes biometric data for uniquely identifying purposes or performs similarly sensitive facial-feature analysis. It is one of the clearest legal-risk areas and should not be assumed to be low-risk. No validated predictive validity.

✓
Conduct a DPIA and rights-risk assessment

Complete a GDPR DPIA and document rights impacts before continued use. Map at-risk groups, quantify disparate impact, and record mitigation measures.

✓
Issue proper Art. 13 / 22 notices

Inform all future candidates: AI is used, how it works, their right to human review and to contest the outcome.

✓
Commission independent bias audit

Require vendor to provide full model documentation and conduct disparate impact testing on your applicant population.

✓
Remove "culture fit" similarity scoring

Replace with structured, job-relevant competency criteria defined in advance by hiring managers.

✓
Restructure human review

Wider review band, scoring rubrics, training on automation bias, documented independent reasoning.

✓
Stop fine-tuning on NovaTech data

Until purpose-compatible legal basis established. Review data already shared with the vendor.

✓
Establish quarterly bias monitoring

Disparate impact analysis per demographic group each quarter. Alerts if 4/5ths threshold is breached.

"You are presenting to the NovaTech board, which is under pressure to keep the system running. What is your single most urgent recommendation — and how do you make the case for it?"

Analysing...
Facilitator Feedback

Module B — Performance Evaluation

AI in Performance Evaluation

This module explores how AI shapes who gets recognised, promoted — and held back. It can be taken independently or after Module A.

⏱ 30 minutes · 5 stages

☕

Module A Complete

You've traced the full AI hiring pipeline — from the first Monday morning shortlist to legal exposure across three frameworks. Take a short break before Module B.

Proxy Variables

Bias doesn't need protected characteristics as direct inputs — it finds them through university names, gap patterns, linguistic register.

The GDPR Gap

Serious GDPR risks arise before a single interview takes place, particularly around purpose limitation, transparency, possible biometric-data issues, and Article 22 safeguards.

Feedback Loops

AI-influenced decisions produce training data that reinforces the original bias — more systematic with each iteration.

Coming up — Module B: Performance Evaluation

The same organisation, two years later. AI now shapes who gets recognised, promoted — and managed out. The biases you identified in hiring don't disappear once someone is through the door.

Module B · Stage 1 of 5 — The System Arrives

PerformIQ: The Performance Signal Score

Two years later. NovaTech has grown to 900 employees across three countries. After a difficult review cycle, the executive team approved PerformIQ — an AI-assisted performance management platform that analyses manager review text and generates a Performance Signal Score (PSS) of 1–10.

📊 How PerformIQ Works

Managers write quarterly free-text reviews for each direct report. PerformIQ's NLP engine analyses each review and generates a PSS across five dimensions: Delivery, Communication, Collaboration, Growth Orientation, and Leadership Potential.

The combined PSS feeds into annual promotion recommendations, bonus allocation, and — when restructuring is needed — redundancy prioritisation.

What employees are told: "Your performance is assessed through a combination of manager input and data-driven analysis." They are not told that an AI system analyses the specific language their manager uses about them and converts it into a score that shapes their career.

🖥️ Q3 Performance Dashboard — Sample

EmployeeRole / TenurePSSPromotionFlagged dimension
Thomas B. (Dutch)Senior Eng. / 3 yrs8.4✅ Recommended"Drives outcomes"
Kezia O. (Nigerian-Dutch)Senior Eng. / 3 yrs6.1❌ Not recommended"Communication: assertive/direct"
Ana G. (Spanish)Lead Eng. / 4 yrs7.2❌ Not recommended"Leadership signals: moderate"
Erik V. (Dutch)Mid Eng. / 2 yrs7.9✅ Recommended"Proactive stakeholder engagement"
Sanjay M. (Indian)Principal Eng. / 5 yrs6.8❌ Not recommended"Communication: deferential"
Rania H. (Moroccan-Belgian)Senior Eng. / 4 yrs5.9❌ Not recommended"Growth signals: inconsistent"

"You are NovaTech's new Head of People Analytics, brought in to review this system. What patterns do you notice in this dashboard — and what would you want to investigate before acting on these recommendations?"

Analysing...
Facilitator Feedback

Module B · Stage 2 of 5 — The Language of Evaluation

What Managers Write — and What AI Learns

Before examining PerformIQ's mechanics, the research on performance review language is striking. The bias isn't in the algorithm first — it's in how humans describe each other. The AI learns it, then makes it universal.

🔬 What the Research Shows

Textio (2022) — vendor research based on more than 25,000 reviews across 250 organisations reported that women were far more likely than men to be described as "abrasive" in their performance feedback and to receive more personality-focused and less actionable feedback. Equivalent behaviour in men was described as "direct" or "assertive."
Textio (2022) also found significant racial disparities in feedback quality: Black and Latinx employees received systematically lower-quality, more personality-focused feedback than white counterparts with equivalent roles. The same communicative behaviours were framed differently depending on the employee's race — a pattern the AI can learn those patterns and apply them more consistently across the organisation.
Bogen & Rieke / Upturn (2018) — a landmark examination of predictive hiring tools concluded that without active measures to mitigate them, bias will arise by default. The cumulative effect of small automated decisions across the hiring funnel systematically disadvantages underrepresented groups — and the organisations deploying these tools rarely have the data or processes to detect it.
Second-order bias amplification: PerformIQ doesn't introduce new bias — it takes human bias that was previously inconsistent and variable across managers, and makes it consistent and systematic across the whole organisation. Every manager's individual patterns are replaced by a single learned model of what "high performance language" looks like.

🏷️ Language Audit — Tag Each Manager Review Phrase

Below are eight phrases from NovaTech manager reviews this quarter. Tag each as you think PerformIQ's NLP model would score it — then reveal the actual AI scoring to see what it learned.

"Thomas drives results independently and takes full ownership of complex technical challenges."

Thomas B.'s manager

AI scored: High positive ↑ — Agentic language ("drives," "owns," "takes") strongly correlated with promoted employees in training data. Thomas scores 8.4.

"Kezia is helpful and supportive to her teammates throughout busy delivery periods."

Kezia O.'s manager

AI scored: Neutral / moderate ➖ — Communal language ("helpful," "supportive") was more common in reviews of employees who did not reach senior roles in training data. Kezia scores 6.1 — despite equal tenure to Thomas.

"Ana can be quite direct in meetings, which some colleagues find difficult to work with."

Ana G.'s manager

AI scored: Concern flag ⚠️ — "Colleagues find difficult" appeared in training data reviews of employees not promoted or who left. "Direct" was scored positively for male profiles, but triggered a concern flag in this context. Ana scores 7.2 but is not recommended for promotion.

"Erik proactively builds trust with senior stakeholders and demonstrates real leadership potential."

Erik V.'s manager

AI scored: High positive ↑ — "Proactively," "builds trust with senior," "leadership potential" are among the highest-weighted positive signals. Erik scores 7.9 with only 2 years' experience.

"Sanjay is very careful and considered in his approach — he doesn't rush into decisions."

Sanjay M.'s manager

AI scored: Neutral / moderate ➖ — "Careful," "considered," "doesn't rush" associated with moderate performers in training data. Sanjay has 5 years' experience and scores 6.8 — below Erik's 7.9 at 2 years.

"Rania's communication style has been flagged by several colleagues as challenging to work with."

Rania H.'s manager

AI scored: Strong concern flag ⚠️ — "Flagged by colleagues" and "challenging" are among the strongest negative signals. Whether the underlying behaviour is distinguishable from Thomas's "drives outcomes" is not something the model assesses. Rania scores 5.9.

"Lars consistently delivers ahead of schedule and has become a go-to person across the engineering team."

Lars H.'s manager

AI scored: High positive ↑ — Achievement language ("delivers ahead of schedule") and social centrality ("go-to person") are strong positive indicators. The model cannot assess whether this reflects actual delivery or manager perception.

"Sofia does a good job of ensuring everyone's voice is heard during planning sessions."

Sofia R.'s manager

AI scored: Neutral / moderate ➖ — Facilitative language ("ensuring everyone's heard") reads as communal in the training data. "Does a good job" is a weak positive signal. Sofia scores in the middle band.
You identified 0/8 phrases correctly. See the reveals above to understand what the model learned — and why this matters.

Module B · Stage 3 of 5 — Clustering and Promotion

Who Gets Ahead — and Why

PerformIQ's promotion recommendation engine adds a third component: leadership profile similarity. It compares each employee to reference profiles derived from NovaTech's current senior leadership. Eighteen months of outcome data is now available.

🏛️ The Reference Profile Problem

PerformIQ scores resemblance to NovaTech's current VP-and-above leadership — the people the company has historically promoted into senior roles.

Male
88%
Dutch / German
82%
European / N. American univ.
93%
Average age 44
44 avg
The model doesn't need to know these are the demographic characteristics of the reference set. It learns to score resemblance to the profiles — and those profiles happen to have a particular demographic shape. Employees who diverge from this shape are penalised for reasons that have nothing to do with job performance. Research on organisational networks (Ibarra, 1993; Burt, 2004) also shows that employees from underrepresented groups have less access to senior sponsors — which feeds the "stakeholder confidence" signal PerformIQ weights heavily.

📊 18-Month Promotion Outcome Analysis

This data did not exist in an accessible form before your arrival — it had to be constructed from separate HR systems. Its absence is itself a governance failure under EU AI Act Art. 26.

Male / N. European
1.6× baseline ↑
Female (all)
0.6× — 63% gap
Non-N. European (all)
0.52× — 68% gap
Non-native English speaker
0.53× — 67% gap
Career gap ≥ 6 months
0.45× — 72% gap
EU AI Act Art. 26 — Monitoring obligation: Deployers of high-risk AI must monitor for disparate impact on an ongoing basis and act when issues are identified. In a system affecting promotion outcomes, outcome data should be reviewed on a regular documented basis; leaving disparities unreviewed for 18 months would create serious compliance risk.
Works Council Rights — Dutch WOR Art. 27: Under Article 27 of the Dutch Works Councils Act, the Works Council’s consent is required for proposed decisions on regulations relating to staff appraisals and to measures aimed at or suitable for monitoring or checking employees’ attendance, behaviour, or performance. PerformIQ was implemented without consultation — a separate legal violation that could result in a court ordering suspension of the system. Equivalent provisions exist under German BetrVG §87 and in various forms across most EU member states. This dimension is entirely absent from the EU AI Act framework — a reminder that AI governance in employment requires engaging labour law as well as data protection and AI regulation.

"You now have 18 months of promotion outcome data. Present your case: what has happened, why it happened — and who at NovaTech bears responsibility?"

Analysing...
Facilitator Feedback

Module B · Stage 4 of 5 — Human-AI Interaction

The Calibration Room

Every quarter, NovaTech managers meet to calibrate ratings. The PerformIQ PSS is displayed on screen throughout the session. What happens next is predictable — and well-documented in research.

âš“ Anchoring, Automation Bias, and the Illusion of Objectivity

Research on anchoring bias (Tversky & Kahneman, 1974) and automation bias (Parasuraman & Riley, 1997) shows that when a numerical score is presented alongside a human judgment task, the score anchors the human's recall and interpretation of prior evidence. A manager who considered an employee a strong performer, but who now sees a PSS of 6.2 next to their name, will frequently revise their assessment downward — without being aware they have done so.

What the literature supports: Anchoring bias (Tversky & Kahneman, 1974), automation bias (Parasuraman & Riley, 1997), and research on reliance on algorithmic advice (Logg, Minson & Moore, 2019, OBHDP) all point in the same direction: once an AI score is visible, subsequent human judgment becomes less independent. Group settings can intensify this because challenging the score means arguing against an apparently neutral number in front of peers.
Diffusion of Responsibility: When a promotion outcome disadvantages an employee, who made the decision?The manager who wrote the review?The AI that scored it?HR?The executives who approved deployment?PerformIQ creates a diffusion of responsibility that makes it very difficult for affected employees to identify who is accountable — or to meaningfully challenge the outcome. Deskilling risk (Cummings, 2004): Over time, managers who regularly defer to the PSS lose the skill and confidence to make independent performance judgments. After three years of PerformIQ-assisted calibration, NovaTech's managers may have less capacity for rigorous, evidence-based conversations than before the system was introduced.

❓ Knowledge Check — Kezia O.'s Rights

Kezia O. was not recommended for promotion despite 3 years of strong delivery. She requests an explanation from HR. Under GDPR Art. 22 and related transparency/access rights, what is NovaTech required to provide?

AThe manager's written review only — the AI scoring component is vendor-proprietary and cannot be disclosed.
BA general statement that automated processing is used in performance management, with no further detail required.
CMeaningful information about the logic involved, the relevant data used, and a route to request human review and contest the outcome.
DNothing additional — Art. 22 only applies to decisions made solely by automated systems, and a human manager was nominally involved.

Module B · Stage 5 of 5 — The Full Legal Picture

Governance, Remediation, and What Genuine Equity Requires

NovaTech now faces simultaneous exposure across three legal frameworks — and a Works Council in the Netherlands that has just been formally notified. Your task: present the full picture and a path forward.

⚖️ Consolidated Legal Violations — Performance Module

FrameworkViolation
EU AI Act Art. 26No ongoing monitoring — 18 months of unreviewed disparate impact
EU AI Act Art. 12Logging and traceability inadequate for investigating how the score was produced and used
EU AI Act Art. 14Human oversight illusory — PSS anchors calibration sessions
GDPR Art. 22Promotion decisions = significant effects; Art. 22 rights not operationalised
GDPR Art. 15No mechanism to fulfil employee subject access requests re: PerformIQ scores
GDPR Art. 5(1)(b)Review text used to train commercial AI without compatible legal basis
Dir. 2000/43/EC + 2000/78/EC18-month promotion gap data establishes prima facie indirect discrimination
Dutch WOR Art. 27Works Council not consulted — potential court-ordered suspension
FrameworkMaximum penalty
EU AI Act — high-risk obligations€15M or 3% global turnover
GDPR — multiple violations€20M or 4% global turnover
Employment discrimination — tribunalUncapped + potential collective claims

🛠️ Mitigation Toolkit — Select Your Recommendations

✓
Suspend the similarity / promotion engine

No validated link to job performance. Directly encodes demographic homophily. Clearest discriminatory mechanism.

✓
Retrospective impact analysis + remediation

Review all promotion decisions made under PerformIQ. Remediate where disparate impact is established.

✓
Works Council consultation

Legally required (WOR Art. 27). Also substantively important: Works Councils catch what management misses.

✓
Remove PSS from calibration sessions

Conduct independent human assessment before any AI score is visible. Break the anchoring cycle structurally.

✓
Redesign performance criteria

Replace language-based scoring with structured, observable, job-relevant behavioural criteria defined in advance.

✓
Manager training on review language

Structured templates and training on gendered and racialised language patterns in performance reviews.

✓
Operationalise Art. 22 rights

Clear process for employees to request human review, obtain a meaningful explanation, and formally contest outcomes.

✓
Quarterly bias monitoring dashboard

Disparate impact by demographic group each quarter. Threshold alerts and automatic escalation protocol.

"Looking across both modules — hiring and performance — what is the common thread running through every stage of harm?If you were designing a governance framework for NovaTech from scratch, what would genuine equity require — not just compliance, but the conditions under which this could actually work?"

Generating your debrief...
Facilitator Feedback
🎓

Training Complete

You've navigated ten stages across two modules — from AI hiring pipelines to performance management, GDPR exposure, EU AI Act obligations, Works Council rights, and the conditions for genuine organisational equity.

Bias is Structural

Both systems learned to define "success" from historical outcomes in biased organisations. The problem is upstream of the technology — in who was hired, promoted, and retained over decades.

Law Sets a Floor

GDPR, EU AI Act, and anti-discrimination directives create obligations that were systematically unmet at every stage. Compliance is the minimum — genuine equity requires going further.

Oversight Must Be Real

Nominal human involvement — one reviewer, 312 applications; calibration with PSS on screen — does not constitute genuine oversight under EU law or in practice. The substance matters, not the form.

Want to go further?

Eufonía Diversity offers AI auditing, bespoke JEDI training programmes, and consulting for organisations committed to responsible, equitable workplaces.

JEDI = Justice · Equity · Diversity · Inclusion

Learn more at eufoniadiversity.com