Ethics and Responsible AI - When the Interviewer Asks "What Would You Do If..."

Reading time: ~35 min | Interview relevance: Critical | Roles: MLE, Applied Scientist, Research Scientist, AI Engineer, ML Product Manager, AI Ethics Specialist

The Real Interview Moment

You are in the final behavioral round at a major tech company. The interviewer, a director of applied science, leans forward and says: "Your team has shipped a content moderation model that's performing well overall - 94% precision, 91% recall. But a colleague runs a fairness audit and discovers the model flags content from certain minority communities at 3x the rate of the general population. The product team says the launch timeline can't slip. What do you do?"

You pause. This is not a straightforward STAR question. There is no single "right" answer the way there is for a coding problem. But there are definitely wrong answers - and the interviewer knows exactly what they reveal about your judgment, your values, and your ability to navigate the hardest decisions in AI.

You consider a purely technical response: "I would retrain with balanced data." But you sense the question is bigger than that. You think about saying "I would escalate to leadership," but that sounds like you are passing the buck. You want to demonstrate that you can hold competing priorities in tension - user safety, fairness, business timelines, technical feasibility - and reason through them with maturity.

This chapter teaches you to do exactly that. Ethics and responsible AI questions are increasingly common in AI interviews, and they separate candidates who have genuinely thought about the impact of their work from those who have not. The goal is not to prove you are a philosopher. The goal is to prove you can make sound, principled decisions when the stakes are high and the answers are not obvious.

What You Will Master

• Why companies ask ethics questions and what they are actually evaluating
• The core responsible AI concepts you must understand (bias, fairness, privacy, transparency)
• A structured framework for answering "What would you do if..." ethical scenarios
• ML-specific fairness metrics and when to use each
• Privacy and data ethics - navigating real-world constraints
• Company-specific responsible AI approaches (Google, Meta, Microsoft, OpenAI, Anthropic)
• How to discuss past ethical decisions from your own experience
• The most common ethical scenario questions and how to handle each

Self-Assessment: Where Are You Now?

Level	Description	Target
Unaware	"I have not thought much about AI ethics beyond headlines"	Read everything - this is a growing interview topic and a gap will hurt you
Theoretical	"I understand bias and fairness conceptually but have never dealt with it in practice"	Focus on Parts 3-5 - build practical reasoning skills
Experienced	"I have dealt with fairness or privacy issues at work but need to structure my answers"	Focus on Parts 2 and 6 - framework and scenario practice

Part 1 - Why Companies Ask Ethics Questions

The Shift in AI Hiring

Five years ago, ethics questions were rare in AI interviews. Today, they appear in over 60% of behavioral rounds at major AI companies. Three forces drove this change:

1. Regulatory pressure. The EU AI Act, NIST AI Risk Management Framework, and state-level legislation mean companies face real legal exposure from irresponsible AI.
2. Public incidents. High-profile failures - biased hiring tools, discriminatory lending models, harmful content generation - have made responsible AI a board-level concern.
3. Competitive differentiation. Companies like Anthropic, Google DeepMind, and Microsoft position responsible AI as a core value and hire for it explicitly.

What the Question Actually Evaluates

60-Second Answer

"Companies ask ethics questions to evaluate four things: technical depth (do you understand bias, fairness metrics, and mitigations?), judgment under pressure (can you reason through tradeoffs when there is no clean answer?), stakeholder navigation (can you work with product, legal, and leadership when values conflict with timelines?), and values alignment (do you take responsible AI seriously or treat it as a checkbox?). The best answers demonstrate that you can hold competing priorities in tension and arrive at a principled, defensible decision."

What Interviewers Actually Write Down

Interviewer Notes	Signal
"Identified the ethical issue immediately and named it precisely"	Strong positive
"Balanced user harm against business needs without dismissing either"	Strong positive
"Proposed concrete technical mitigations, not just vague principles"	Positive
"Referenced relevant fairness metrics by name and explained tradeoffs"	Strong positive
"Jumped straight to 'just retrain the model' without analyzing the problem"	Negative - oversimplifies
"Said 'I would escalate to my manager' and stopped there"	Negative - passes the buck
"Dismissed the ethical concern as not their responsibility"	Strong negative
"Could not articulate any fairness or bias concept"	Strong negative for senior roles

Part 2 - The Ethical Reasoning Framework (ERF)

When you get a "What would you do if..." ethics question, do not wing it. Use a structured approach.

The Five-Step ERF

Step 1 - Name the Harm. Be specific. "The model is biased" is weak. "The model exhibits disparate impact against protected demographic groups, flagging their content at 3x the base rate" is strong. Naming the harm precisely demonstrates technical understanding.

Step 2 - Assess Severity. Not all ethical issues are equal. A recommendation model that slightly over-indexes popular items is different from a healthcare model that under-diagnoses a specific population. Ask:

• Who is harmed and how many people?
• What is the magnitude of harm (inconvenience vs. material damage vs. safety risk)?
• Is the harm reversible or irreversible?
• Is the harm disproportionately borne by vulnerable populations?

Step 3 - Identify Stakeholders. Map everyone who has a stake in the decision:

• Users who are directly affected
• The product team with launch commitments
• Legal and compliance with regulatory obligations
• Leadership with reputational concerns
• Your engineering team who will implement the fix
• The broader community affected by precedent

Step 4 - Propose Mitigations. Offer a tiered approach - not just one option:

• Immediate: What can you do right now to reduce harm? (guardrails, human review, scope reduction)
• Short-term: What fix can ship in days or weeks? (model adjustments, data augmentation, threshold tuning)
• Long-term: What systemic change prevents recurrence? (fairness testing in CI/CD, bias audits, diverse evaluation sets)

Step 5 - Define Governance. Who decides, and how do you document the decision?

• What threshold would make you block a launch?
• Who has authority to override?
• How will you monitor the issue post-launch?
• How will you document the decision and rationale?

Common Trap

Do not present ethics as a binary choice between "launch with the problem" and "don't launch." Interviewers want to see you navigate the middle ground - phased rollouts, targeted mitigations, monitoring with rollback triggers. Binary thinking signals inexperience with real-world tradeoffs.

ERF Applied: The Content Moderation Scenario

Here is how the ERF applies to the opening scenario:

Name the Harm: "The content moderation model has a 3x higher false positive rate for content from minority communities. This constitutes disparate impact - legitimate speech from these communities is being suppressed at a disproportionate rate."

Assess Severity: "This is high severity. Content suppression affects users' ability to participate on the platform. The harm is concentrated on already-marginalized communities. It is partially reversible (we can restore flagged content) but the chilling effect on participation is harder to undo."

Identify Stakeholders: "Users in affected communities, the product team with launch commitments, our trust and safety team, legal (disparate impact has regulatory implications), and our company's public commitments to equitable AI."

Propose Mitigations: "Immediately, I would add a human review layer for content from the affected demographic segments to prevent false positives while we fix the root cause. Short-term, I would audit the training data for representation gaps and retrain with stratified sampling, adding subgroup-level performance metrics to our evaluation suite. Long-term, I would advocate for disaggregated evaluation as a launch gate - no model ships without subgroup performance within defined bounds."

Define Governance: "I would document this finding, present it to the product lead and my engineering manager, and propose a modified launch with the human review safeguard. If they push back, I would escalate to our responsible AI review board (or equivalent). I would not unilaterally block the launch, but I would ensure the decision-makers have full information about the risk."

Part 3 - Bias and Fairness - The Technical Foundation

You do not need to be a fairness researcher to pass ethics interview questions, but you need to understand the core concepts well enough to discuss them fluently.

Sources of Bias in ML Systems

The Three Fairness Definitions You Must Know

note

These three definitions are mathematically incompatible in most real-world scenarios (the Chouldechova impossibility result). Knowing this - and being able to explain why - is a strong signal in interviews.

1. Demographic Parity (Statistical Parity)

• Definition: The probability of a positive outcome should be equal across groups. P(Y=1|A=0) = P(Y=1|A=1)
• When to use: When equal representation in outcomes is the goal (e.g., job applicant screening)
• Limitation: Ignores differences in base rates; can require predicting incorrectly to achieve parity

2. Equalized Odds (Separation)

• Definition: True positive rate AND false positive rate should be equal across groups. P(Y_hat=1|Y=1,A=0) = P(Y_hat=1|Y=1,A=1) and P(Y_hat=1|Y=0,A=0) = P(Y_hat=1|Y=0,A=1)
• When to use: When the model should be equally accurate for all groups (e.g., medical diagnosis)
• Limitation: Requires access to true labels for all groups; harder to enforce in practice

3. Predictive Parity (Sufficiency)

• Definition: The positive predictive value should be equal across groups. P(Y=1|Y_hat=1,A=0) = P(Y=1|Y_hat=1,A=1)
• When to use: When a positive prediction should be equally trustworthy regardless of group (e.g., credit scoring)
• Limitation: Can still produce disparate impact in who receives positive predictions

The Impossibility Result

Given:
- Two demographic groups with different base rates (P(Y=1|A=0) ≠ P(Y=1|A=1))
- An imperfect classifier (not 100% accurate)

Then: You CANNOT simultaneously achieve:
  1. Demographic parity
  2. Equalized odds
  3. Predictive parity

Implication: Every fairness decision involves a tradeoff.
The right choice depends on the context and the type of harm you are trying to prevent.

60-Second Answer

"The three core fairness metrics are demographic parity (equal outcome rates across groups), equalized odds (equal accuracy across groups), and predictive parity (equal precision across groups). The key insight is the impossibility result - when base rates differ and the classifier is imperfect, you cannot satisfy all three simultaneously. So the real question is: which type of unfairness is most harmful in this specific context? For a hiring tool, demographic parity matters most because equal access to opportunity is the goal. For medical diagnosis, equalized odds matters most because you need equal accuracy regardless of who the patient is. For lending, predictive parity matters because a positive prediction (approve the loan) should be equally reliable for all groups."

Practical Bias Detection and Mitigation

Detection - What to Measure:

Metric	What It Catches	Tool
Subgroup performance (accuracy, precision, recall per group)	Performance disparities	Fairlearn, Aequitas, What-If Tool
Disparate impact ratio (selection rate of minority / majority)	Outcome disparities (4/5ths rule threshold)	Custom metrics, Fairlearn
Calibration curves per group	Prediction confidence disparities	scikit-learn calibration_curve
Error analysis by demographic slice	Specific failure patterns	Slicefinder, custom slicing
Representation audit of training data	Data imbalances	Pandas profiling, custom scripts

Mitigation - Three Stages:

Stage	Technique	Example
Pre-processing	Resampling, reweighting, data augmentation	Oversample underrepresented groups; collect additional data for minority slices
In-processing	Constrained optimization, adversarial debiasing, fairness regularization	Add fairness constraint to loss function; adversarial network to remove demographic signal
Post-processing	Threshold adjustment, calibration, reject option classification	Set group-specific thresholds to equalize FPR; calibrate probabilities per group

Part 4 - Privacy and Data Ethics

The Privacy Landscape for ML Practitioners

Privacy is not just a legal concern - it is an engineering concern. As an ML practitioner, you need to understand:

What Regulations Require:

• GDPR (EU): Right to erasure, right to explanation, data minimization, purpose limitation, privacy by design
• CCPA/CPRA (California): Right to know, right to delete, opt-out of sale, data minimization
• HIPAA (US Healthcare): De-identification standards, minimum necessary rule, business associate agreements
• AI-specific: EU AI Act risk tiers, NIST AI RMF, emerging state legislation

What This Means for ML:

Key Privacy Concepts for Interviews

Differential Privacy:

• Provides a mathematical guarantee that any single individual's data has limited influence on model outputs
• Epsilon (privacy budget) controls the privacy-utility tradeoff - lower epsilon = more privacy, less utility
• Used by Apple (keyboard predictions), Google (Chrome usage statistics), US Census Bureau

Federated Learning:

• Training happens on-device; only model updates (gradients) are sent to the server
• Raw data never leaves the user's device
• Used by Google (Gboard), Apple (Siri), hospitals (medical imaging)

Machine Unlearning:

• The ability to remove the influence of specific training data from a trained model
• Required for GDPR right-to-erasure compliance
• Approaches: full retraining (expensive), approximate unlearning (SISA, influence functions)

danger

Never say "we anonymized the data so privacy is not a concern." Research has shown that "anonymized" datasets can often be re-identified. In a 2019 study, researchers re-identified 99.98% of Americans in any anonymized dataset using 15 demographic attributes. Instead, discuss defense in depth: anonymization PLUS access controls PLUS differential privacy PLUS audit logs.

Privacy Scenario: Interview Question

Q: "Your team wants to build a recommendation system using user behavior data. A privacy review reveals that the training data contains sensitive health-related browsing patterns. How do you proceed?"

Strong Answer Structure:

1. Acknowledge the sensitivity: "Health-related browsing data is among the most sensitive categories. Even if we are not a health provider, users have a reasonable expectation that their health interests are private."

2. Assess what you actually need: "First, I would question whether we need this data at all. Can we build an effective recommendation system using less sensitive signals? Data minimization is both a privacy principle and a practical risk reduction strategy."

3. If the data is necessary, propose technical safeguards:

   • "Apply k-anonymity or l-diversity to the dataset before training"
   • "Use differential privacy during training to limit the model's ability to memorize individual patterns"
   • "Run membership inference attacks on the trained model to verify individual records cannot be extracted"
   • "Implement purpose limitation - this model should only be used for recommendations, not shared with other teams"

4. Governance: "Document the decision, get sign-off from our privacy team, set up quarterly audits, and create a data deletion pipeline for users who exercise their right to erasure."

Part 5 - Deployment Ethics and Real-World Impact

The Deployment Checklist

Before any ML system goes to production, a responsible practitioner considers:

Question	Why It Matters
Who benefits from this system and who might be harmed?	Identifies asymmetric impact
What happens if the model is wrong? What is the worst case?	Calibrates appropriate confidence levels and human oversight
Can users understand why the system made a decision about them?	Right to explanation; user trust
Is there a human override mechanism?	Safety net for high-stakes decisions
How will we monitor for drift and degradation?	Models decay; bias can emerge over time
What is the rollback plan?	Ability to quickly undo harmful deployments
Did we test on the populations who will actually use this?	Prevents deployment bias

Explainability - When and How Much

Not every model needs to be fully interpretable. The level of explainability should match the stakes:

Generative AI Ethics - The New Frontier

If you are interviewing for roles involving LLMs, expect questions about:

Hallucination and Factuality:

• "How do you ensure your LLM-based system does not provide false medical/legal/financial information?"
• Strong answer: retrieval-augmented generation, citation grounding, confidence thresholds, human review for high-stakes outputs

Content Safety:

• "How would you design a content safety pipeline for a user-facing LLM?"
• Strong answer: input classifiers, output filters, red-teaming, constitutional AI principles, RLHF alignment, monitoring dashboards

Copyright and Training Data:

• "A customer claims your model reproduces their copyrighted text verbatim. How do you respond?"
• Strong answer: memorization testing, deduplication in training data, opt-out mechanisms, content filtering

Dual Use:

• "Your model could be used to generate misinformation or assist in harmful activities. How do you think about this?"
• Strong answer: threat modeling, use case restrictions, monitoring for misuse patterns, responsible disclosure

Common Trap

Do not take an absolutist position on generative AI ethics. Saying "we should never deploy LLMs because they hallucinate" signals that you cannot navigate nuance. Instead, discuss risk-appropriate safeguards: "For a creative writing assistant, hallucination is a feature. For a medical advice system, it is a critical safety risk. The safeguards should match the stakes."

Part 6 - Company-Specific Responsible AI Approaches

When interviewing, tailor your answers to the company's stated values and approach.

Google

• Framework: AI Principles (2018) - seven principles including "be socially beneficial," "avoid creating or reinforcing unfair bias," "be built and tested for safety"
• Tools: Model Cards, Fairlearn integration, What-If Tool, Data Cards
• Key emphasis: Disaggregated evaluation, inclusive design, red-teaming
• In interviews, reference: Subgroup analysis, Model Cards for model documentation, inclusive evaluation datasets

Meta

• Framework: Responsible AI team, five pillars - privacy, fairness, robustness, transparency, accountability
• Tools: Fairness Flow (internal), open-source fairness research
• Key emphasis: Fairness in ranking systems, content moderation, responsible open-source release
• In interviews, reference: Fairness in large-scale systems, content integrity, balancing openness with safety

Microsoft

• Framework: Responsible AI Standard (2022), six principles - fairness, reliability, privacy, inclusiveness, transparency, accountability
• Tools: Responsible AI Dashboard, Fairlearn (Microsoft-funded), InterpretML, Counterfit (adversarial testing)
• Key emphasis: Impact assessments before deployment, Responsible AI reviews for sensitive use cases
• In interviews, reference: Impact assessments, Responsible AI tooling, Azure AI content safety

OpenAI

• Framework: Safety-first approach, iterative deployment, usage policies
• Key emphasis: Alignment research, red-teaming, preparedness framework for frontier models
• In interviews, reference: RLHF for alignment, safety evaluations, responsible capability scaling

Anthropic

• Framework: Responsible Scaling Policy, Constitutional AI
• Key emphasis: AI safety research, interpretability, honest and harmless AI
• In interviews, reference: Constitutional AI, scalable oversight, interpretability research, responsible scaling commitments

Amazon

• Framework: AI Service Cards, Fairness in ML
• Tools: SageMaker Clarify (bias detection and explainability)
• Key emphasis: Customer obsession applied to AI fairness, practical bias detection at scale
• In interviews, reference: SageMaker Clarify, bias detection in production ML pipelines, customer impact

Part 7 - The Hardest Ethical Scenario Questions

These are the questions that separate senior candidates from everyone else. Practice each one.

Scenario 1: The Accurate but Unfair Model

"Your model achieves 95% accuracy overall. When you disaggregate by race, accuracy is 97% for the majority group and 82% for the minority group. The product team says 95% overall accuracy meets the launch bar. What do you do?"

Framework for Answering:

1. Name the harm: 15-point accuracy gap is disparate impact
2. Quantify: How many users are affected? What does an error mean for them?
3. Challenge the metric: Overall accuracy is misleading when it masks subgroup disparities
4. Propose alternatives: Require minimum subgroup accuracy thresholds alongside overall accuracy
5. Suggest mitigations: Targeted data collection, subgroup-specific thresholds, additional human review for minority group predictions
6. Escalate appropriately: This needs to be a documented decision, not a quiet launch

Scenario 2: The Data You Should Not Have

"During an exploratory analysis, you discover that your training dataset contains personally identifiable information that should have been stripped during preprocessing. The model has already been trained and is performing well. What do you do?"

Key Points:

• This is not optional - PII in training data is a compliance violation
• Report immediately to your data privacy team and manager
• Assess the risk: Can the model memorize and reproduce the PII?
• Determine the remediation: Retrain from scratch on properly cleaned data, or use machine unlearning if available
• Conduct a root cause analysis: Why did the preprocessing pipeline fail?
• Document everything - this is the kind of incident that auditors will ask about

Scenario 3: The Pressure to Ship

"Your manager tells you to skip the fairness evaluation because the product launch is in two days and the test suite takes 8 hours to run. They say the model was tested on a diverse dataset, so it should be fine. What do you do?"

Key Points:

• "Should be fine" is not a fairness evaluation
• Propose a middle ground: run a reduced fairness test (key subgroups, key metrics) that takes 2 hours
• If even that is impossible, propose launching with a smaller rollout (1% of traffic) while the full test runs
• Document the risk: send a written note saying "We are launching without fairness evaluation. Here are the risks."
• Know your non-negotiables: if the domain is high-stakes (healthcare, criminal justice, lending), push harder

Scenario 4: The Dual-Use Discovery

"You are building a medical imaging model to detect tumors. During development, you realize the same architecture could be used to identify individuals from medical images, creating a surveillance risk. How do you handle this?"

Key Points:

• This is a dual-use concern - beneficial technology with potential for misuse
• Continue the beneficial work but document the dual-use risk
• Restrict access: the model weights should not be publicly released without safeguards
• Add architectural guardrails: strip identifying features from inputs before processing
• Engage your organization's ethics review board
• Consider responsible disclosure if the risk is novel

Scenario 5: The Uncomfortable Correlation

"Your fraud detection model uses zip code as a feature. You discover that zip code is highly correlated with race in your data. Removing it drops precision by 8%. What do you do?"

Key Points:

• Zip code is a textbook proxy variable for race
• Keeping it creates a model that discriminates by race, even if race is not an explicit feature
• Explore alternatives: can you replace zip code with features that capture the same fraud signal without the racial correlation? (e.g., transaction velocity, device fingerprinting, behavioral patterns)
• If no replacement achieves comparable performance, this is a fairness-utility tradeoff that needs to be decided at a level above individual contributors
• Propose running the model both ways and measuring the fairness impact explicitly

danger

Never say "zip code is not a protected attribute, so it is fine to use." Proxy discrimination - using features that correlate with protected attributes - is both ethically problematic and increasingly regulated. The EU AI Act and several US frameworks explicitly address proxy discrimination.

Part 8 - Building Your Ethics Story Portfolio

You need 2-3 prepared stories from your own experience where you navigated an ethical or responsible AI challenge. If you have not encountered one directly, draw from adjacent experiences.

Story Template

SITUATION:
What was the context? What system were you building?
What ethical issue did you encounter or proactively identify?

TASK:
What was your responsibility? What were the competing pressures?

ACTION (the core of your answer):
1. How did you identify and name the issue?
2. Who did you involve in the decision?
3. What analysis did you conduct?
4. What solution did you propose or implement?
5. How did you handle pushback (if any)?

RESULT:
What was the outcome for the product?
What was the outcome for users?
What process or cultural change resulted?
What would you do differently in hindsight?

If You Have Never Faced an Ethical Issue Directly

Draw from related experiences:

• Data quality issues where you advocated for higher standards
• Testing gaps where you pushed for more comprehensive evaluation
• Feature decisions where you considered user impact beyond engagement metrics
• Documentation where you pushed for transparency in model limitations
• Peer review where you raised concerns about a colleague's approach

Frame it as: "While this was not a fairness issue per se, I applied the same principled reasoning - I identified a risk to users, quantified it, proposed mitigations, and ensured the decision was documented."

Part 9 - Responsible AI Review Processes

A common interview question for senior roles is: "How would you design a responsible AI review process for your team?"

The Three-Layer Review Architecture

When Each Layer Triggers

Model Risk Level	Layer 1 (Automated)	Layer 2 (Peer)	Layer 3 (Board)
Low risk (internal tools, non-user-facing)	Required	Optional	Not required
Medium risk (user-facing recommendations, content ranking)	Required	Required	Optional
High risk (healthcare, finance, hiring, content moderation)	Required	Required	Required
Critical risk (safety-critical, legal liability)	Required	Required	Required + external audit

Building Model Cards

Model Cards (introduced by Mitchell et al., 2019) are standardized documentation for ML models. Every model your team deploys should have one.

MODEL CARD TEMPLATE:

Model Name: [Name and version]
Date: [Training/deployment date]
Owner: [Team and individual]

1. MODEL DETAILS
   - Architecture: [e.g., XGBoost, BERT fine-tune, custom transformer]
   - Training data: [Source, size, date range, known limitations]
   - Input features: [List with descriptions]
   - Output: [What the model predicts, confidence format]

2. INTENDED USE
   - Primary use case: [What the model is designed for]
   - Out-of-scope uses: [What the model should NOT be used for]
   - Users: [Who interacts with model outputs]

3. PERFORMANCE
   - Overall metrics: [Accuracy, precision, recall, F1, AUC]
   - Disaggregated metrics: [Performance broken down by
     demographic groups, content categories, user segments]
   - Performance across conditions: [Edge cases, low-data
     scenarios, adversarial inputs]

4. FAIRNESS ANALYSIS
   - Protected attributes evaluated: [List]
   - Fairness metrics applied: [Which ones and why]
   - Results: [Disparities found, if any]
   - Mitigations applied: [What was done to address disparities]

5. ETHICAL CONSIDERATIONS
   - Known risks: [Potential harms if misused or if model fails]
   - Data privacy: [How user data is handled]
   - Human oversight: [Where human review is required]

6. LIMITATIONS AND RECOMMENDATIONS
   - Known failure modes: [Where the model performs poorly]
   - Monitoring plan: [What is tracked post-deployment]
   - Update schedule: [How often the model is retrained]

Building Data Cards

Data Cards (Pushkarna et al., 2022) document the datasets used for training and evaluation.

DATA CARD TEMPLATE:

Dataset Name: [Name and version]
Purpose: [What this dataset is used for]
Owner: [Team responsible for maintenance]

1. COMPOSITION
   - Size: [Number of examples]
   - Features: [List with types]
   - Labels: [How labeled, by whom, quality assessment]
   - Temporal range: [Date range of data]

2. COLLECTION PROCESS
   - Source: [Where the data came from]
   - Consent: [How consent was obtained, if applicable]
   - Sampling: [How examples were selected]
   - Preprocessing: [Cleaning, filtering, transformations]

3. DEMOGRAPHIC REPRESENTATION
   - Breakdown by protected attributes: [If available]
   - Known underrepresentation: [Which groups are underrepresented]
   - Efforts to improve representation: [What was done]

4. KNOWN ISSUES
   - Label noise: [Estimated error rate]
   - Biases: [Known biases in collection or labeling]
   - Missing data: [Patterns of missingness]

5. ETHICAL REVIEW
   - PII present: [Yes/No, and how handled]
   - Sensitive content: [Categories of sensitive content]
   - Consent status: [Verified/Unverified/Not applicable]

Part 10 - Ethics Questions by Company Type

Different types of companies face different ethical challenges. Tailor your preparation accordingly.

Big Tech (Google, Meta, Microsoft, Amazon, Apple)

Focus areas: Bias at scale, content moderation, privacy in large datasets, dual-use concerns, regulatory compliance across jurisdictions.

Typical questions:

• "How do you ensure fairness in a model that serves billions of users across diverse cultures?"
• "How do you balance personalization with privacy?"
• "What is your framework for deciding when a model should not be deployed?"

What they want to hear: That you can think at scale. A fairness issue affecting 0.1% of users is still millions of people. You understand the reputational and regulatory exposure. You can propose solutions that work in production at scale, not just in research.

AI-Native Companies (OpenAI, Anthropic, Cohere, Mistral)

Focus areas: Alignment, safety evaluation, red-teaming, responsible release, capability development vs safety research balance.

Typical questions:

• "How do you evaluate whether a model is safe to release?"
• "What is your view on open-source release of powerful models?"
• "How do you think about the tradeoff between model capability and safety?"

What they want to hear: Deep engagement with alignment and safety concepts. Not just "safety is important" but specific technical approaches: RLHF, Constitutional AI, red-teaming methodologies, evaluation frameworks for harmful outputs.

Healthcare and Biotech

Focus areas: Patient safety, regulatory compliance (HIPAA, FDA), clinical validation, explainability for clinicians, health equity.

Typical questions:

• "How do you ensure a diagnostic model does not perform differently across demographic groups?"
• "How do you validate an ML model for clinical use?"
• "What is the appropriate level of human oversight for an ML-assisted medical decision?"

What they want to hear: That you understand the stakes are life and death. You know the regulatory landscape. You insist on clinical validation, not just technical validation. You default to human oversight and conservative deployment.

Finance and Fintech

Focus areas: Fair lending, regulatory compliance (ECOA, FCRA), model explainability, adverse action notices, anti-money laundering.

Typical questions:

• "How do you ensure a credit model does not discriminate against protected groups?"
• "How do you provide explanations for model-based lending decisions?"
• "What is your approach to model risk management?"

What they want to hear: That you know fair lending regulations. You can discuss disparate impact analysis. You understand that in finance, every model decision that affects a consumer must be explainable and auditable.

Startups

Focus areas: Moving fast without creating ethical debt, building responsible practices from scratch, limited resources for fairness testing.

Typical questions:

• "We are a small team. How do you build responsible AI practices without a dedicated ethics team?"
• "How do you prioritize ethics work when you are also trying to ship product?"
• "What is the minimum viable responsible AI process?"

What they want to hear: That you are pragmatic. You can build lightweight processes that scale. You do not need a 50-person ethics team to do the right thing. You can bake fairness checks into the development workflow without slowing the team down significantly.

Practice Exercises

Exercise 1: ERF Application (30 minutes)

Apply the Ethical Reasoning Framework to each of these scenarios. Write out all five steps for each.

1. Your sentiment analysis model performs 20% worse on African American Vernacular English (AAVE) than on Standard American English. The model is used for customer service ticket routing.

2. Your company acquires a smaller company and wants to merge their user data with yours to improve recommendations. Users of the acquired company did not consent to this use.

3. A government agency wants to use your facial recognition API for law enforcement. Your company's terms of service do not prohibit this, but you have concerns.

Exercise 2: Fairness Metric Selection (20 minutes)

For each scenario, identify which fairness metric is most appropriate and explain why:

1. A resume screening tool for entry-level software engineering positions
2. A medical diagnostic tool for detecting diabetic retinopathy
3. A recidivism prediction tool used in bail decisions
4. A content recommendation system on a social media platform
5. A credit scoring model for mortgage applications

Exercise 3: Your Ethics Story (30 minutes)

Using the story template above, write out a complete ethics story from your own experience. Then:

• Identify the weakest part of your story and strengthen it
• Prepare for three follow-up questions the interviewer might ask
• Practice delivering it aloud in under 3 minutes

Exercise 4: Company Research (20 minutes)

Choose two companies you are interviewing with. For each:

• Find their published responsible AI principles or framework
• Identify one specific tool or process they use
• Prepare one question you would ask them about their approach
• Tailor your ethics story to align with their stated values

Interview Cheat Sheet

Quick Reference: Ethics Question Types and Approaches

Question Type	What They Want	Your Approach
"What would you do if you found bias in your model?"	Structured reasoning, not panic	ERF: Name, Assess, Stakeholders, Mitigate, Govern
"How do you think about fairness in ML?"	Technical depth + practical wisdom	Explain 2-3 fairness metrics, the impossibility result, and context-dependent choice
"Tell me about a time you navigated an ethical challenge"	Real experience, not theory	STAR format with emphasis on your reasoning process
"How would you design a responsible AI review process?"	Systems thinking	Pre-deployment checklist + monitoring + incident response + documentation
"What is your view on [controversial AI topic]?"	Nuanced thinking, not ideology	"It depends on context. Here are the tradeoffs..."

Five Phrases That Signal Maturity

1. "The right fairness metric depends on the context and the type of harm we are trying to prevent."
2. "I would want to understand the downstream impact before making a technical recommendation."
3. "This is a decision that should not be made by any single individual - it needs input from multiple stakeholders."
4. "We can mitigate this risk without blocking the launch, but we need to monitor and set rollback criteria."
5. "I would document this decision and the rationale so there is a clear record if it is ever questioned."

Five Phrases That Signal Immaturity

1. "Fairness is too subjective to measure, so I just focus on accuracy."
2. "That is a legal/policy problem, not an engineering problem."
3. "We anonymized the data, so privacy is handled."
4. "The model is not biased because we did not use protected attributes."
5. "I would just retrain with more data."

Spaced Repetition Checkpoints

After Reading (Day 0)

• Can you name the five steps of the Ethical Reasoning Framework?
• Can you explain the three fairness metrics and the impossibility result?
• Can you describe three sources of bias in ML systems?

After 3 Days

• Apply the ERF to a new scenario you have not seen before. Time yourself - can you structure a response in under 2 minutes?
• Without looking, list the key privacy concepts (differential privacy, federated learning, machine unlearning) and explain each in one sentence.
• Recite your prepared ethics story aloud. Does it flow naturally?

After 1 Week

• Explain the difference between pre-processing, in-processing, and post-processing fairness mitigations with an example of each.
• For a company you are interviewing with, explain their responsible AI framework and how your experience aligns.
• Have a friend or peer ask you a "What would you do if..." ethics question you have not prepared for. Evaluate your response against the ERF.

After 2 Weeks

• Teach the fairness impossibility result to someone who is not in ML. If you can explain it clearly, you understand it.
• Take a model you have built and run a fairness audit (even a simple one). Document the results.
• Practice answering the five hardest scenario questions from Part 7 with a study partner.

What Comes Next

You now have the frameworks, technical knowledge, and practice tools to handle ethics and responsible AI questions in any AI interview. In the next chapter, Leadership and Influence, you will learn how to discuss technical leadership, driving adoption of best practices (including responsible AI practices), and influencing decisions when you do not have formal authority - a skill that is essential for senior AI roles.