Data Scientist interview question
How do you assess ethical risks in a model?
Assess who can be harmed, representation and label bias, performance gaps by relevant groups, privacy, explainability, misuse, feedback loops, and appeal paths; document controls, owners, monitoring, and stop conditions.
Direct answer
What a strong answer covers
How it works
An ethics assessment asks who is affected by the model’s decisions and how a wrong or biased decision harms them, then checks the concrete places bias enters: who is missing or under-represented in the training data, whether the labels encode past discrimination, whether performance differs across relevant groups, and whether the features proxy for protected attributes. It covers privacy of the data used, whether the explanation is adequate for the stakes, what misuse looks like, and whether the model’s outputs feed back into its own training data. The result is a set of controls with owners, monitoring by group, an appeal path for affected people, and conditions under which the model is switched off.
Example
A résumé screening model was assessed before launch: the training labels came from historical hiring decisions with a known gender imbalance, so the team measured selection rates by gender and found a 30 percent gap. Mitigations included removing proxy features such as graduation year, re-weighting, a human review of every rejection, and a monthly disparity report with a threshold that pauses the model. The appeal process was published to candidates.
Common mistakes
- Declaring a model fair because protected attributes are not among the features.
- Measuring only overall accuracy and never by group.
- Deploying with no way for an affected person to contest a decision.
A useful answer structure
Answer from your own work.
Do not memorise a script: use only examples you can explain honestly in follow-up questions.
Open the desktop workspace and rehearse a clear answer in your own words.