0:0015:19

Operationalising AI Ethics by Natalie Rouse

15:19 2022

Copying ethics principles from the internet will not make your AI fair

Discriminatory hiring models, facial recognition that fails some groups, deepfakes: the headlines show what happens when AI ethics stays on paper. Most organisations now have principles, yet few have a repeatable way to build them into how models are scoped, trained and run. Natalie Rouse, a general manager leading a data science consultancy across Australia and New Zealand, asks where in the development lifecycle ethics actually has to live.

Learn where to place ethics checkpoints from project scoping to production monitoring

Watch on YouTube

In short

Natalie Rouse, general manager of a data science consultancy in Australia and New Zealand, argues that agreeing on AI ethics principles is not enough and shows how to embed them in the machine learning lifecycle. Using the CRISP-ML process model, she places ethical impact assessments at scoping, subpopulation analysis and performance tolerances in development, and ethics KPIs in production MLOps monitoring.

“Please don't just copy and paste some ethical principles from the internet.”
“As I mentioned back at the start, ethics is not a set and forget, box ticked type of exercise.”

What you will learn

  • Extend a privacy impact assessment into an ethical impact assessment that lets low-risk projects drop out quickly.
  • Define subpopulations in your data so you can measure representation and model performance for each group.
  • Set performance tolerances by asking how accuracy gaps across groups translate into real world outcomes.
  • Document consent, contestability and explainability decisions in a living report before deployment.
  • Embed ethics KPIs and alert thresholds in your existing MLOps monitoring and retraining processes.

Details

Level
Intermediate
Format
Talk
Technical depth
Practitioner
For
Mid-career, Senior individual contributor, People manager
Length
15:19
Published
July 2022

Questions this talk answers

Why is adopting AI ethics principles not enough?

Agreeing on principles does not automatically bridge the gap to the development lifecycle. Careful work is needed to embed them in development and deployment without prohibitive administrative overhead.

Where should ethics checks sit in a machine learning project?

Across all three CRISP-ML phases: an ethical impact assessment and risk register at scoping, subpopulation analysis and tolerance setting during model development, and usage recommendations plus ongoing monitoring at deployment.

How do you check whether a model is fair across groups?

Define subpopulations, review their representation in the training data, and slice performance metrics across them. Defining the groups is hard because demographic attributes like ethnicity and gender can be subjective or only a proxy.

How much accuracy difference between groups is acceptable?

It depends on real world impact: ask whether gaps above a high baseline change outcomes, what happens if one group falls below a certain level, and whether there is a tipping point. Use the answers to set tolerances for production monitoring.

What happens to ethics work once a model is in production?

It continues: embed ethics KPIs in the MLOps framework, use performance thresholds as triggers for alerts or retraining, and give stakeholders visibility of performance.

Topics

Related talks

Browse the library