In short
Most organisations now agree that ethics matters in artificial intelligence, but few have a standard way to embed it in how models are built and run. Natalie Rouse, General Manager of Eliiza, argues that ethical review should sit inside the existing AI lifecycle: an ethical impact assessment at scoping, subpopulation analysis during model development, and ethics KPIs monitored in production. Her core advice is to start with a simple framework now and improve it with every iteration.
Key takeaways
- The five commonly used ethical principles to apply to any AI project
- The most important high-level considerations for embedding ethics into the AI project lifecycle
- How and where to undertake ethical review tasks within the standard project development workflow
- How undertaking an ethical review as part of the development process creates KPIs for ongoing monitoring in production
Principles are not the same as practice
Rouse opens with the familiar headlines: discriminatory hiring models, facial recognition models that perform poorly for some groups, racial profiling and deepfakes. These examples, she says, have pushed individuals and organisations to draft and adopt ethical principles, which have largely converged on fairness, transparency, accountability, contestability, privacy and security, and human-centred values such as autonomy and consent.
The problem is what happens next. Agreeing principles does not automatically connect them to the development lifecycle, and the work of connecting them must avoid burying projects in administration. Rouse also stresses that representation and diversity matter for models to be robust for underrepresented groups, and that the whole effort has to keep evolving.
"And guess what? This is not a set and forget process." (2:39)
She draws on Maya Angelou to describe the journey of learning better and doing better, and encourages teams not to wait for a perfect approach before building one.
Make the framework your own
Rouse's first practical warning is about shortcuts. Ethical principles only earn buy-in when they reflect the organisation's own values, so that ethics and company values pull in the same direction.
"Please don't just copy and paste some ethical principles from the internet." (3:36)
She then highlights several considerations to settle early. Teams should agree and record what an acceptable level of risk looks like, since that shapes both development and how models are managed in production. They need a way to define subgroups or populations within their data, because without them it is impossible to understand how outcomes or accuracy differ between groups of people. And any added process, whether for privacy, ethics or both, must not be prohibitively onerous for the development team.
Mapping ethics onto the CRISP-ML lifecycle
To show where ethical work belongs, Rouse uses the CRISP-ML process model, which evolved from CRISP-DM, the Cross Industry Standard Process for Data Mining. It describes three iterative phases: understanding business requirements and data to validate feasibility, developing and evaluating a model, and deploying a model and stewarding it through its lifetime.
At scoping, she recommends an ethical impact assessment and a risk-level calculation, because a riskier approach might change which solution a team chooses. Many organisations already run privacy impact assessments, and she suggests building on them.
"But going a step further and extending a PIA to be an EIA or ethical impact assessment is an important step up front." (8:28)
The assessment should let low-risk projects with little human or environmental impact drop out quickly, while riskier ones receive proper due diligence. Ethical implications considered here can affect data granularity, explainability requirements, the choice of model and even which data a team has the right to use. The resulting risk register guides the rest of the project.
Subpopulations and the hard question of tolerance
During development, Rouse treats exploratory data analysis as the quantitative heart of ethical AI, since bias comes from unbalanced datasets and underrepresentation. The difficulty lies in defining the groups to measure. Demographic fields such as ethnicity and even gender can be subjective or inadequate, and at best a proxy, and she notes that approaches for classifying image data are still at an early stage.
"It turns out this is a hard question to answer." (10:09)
Once subpopulations are defined, teams can select performance metrics for training and production, then slice results by group as part of the interrogation good practice already demands. The critical step is deciding what a performance gap actually means. If every group is above a high accuracy level, small differences may not matter; if one group falls much lower, or a tipping point exists, the real-world impact could be serious.
"Asking yourself these questions and being really curious about the link between model performance and real world outcomes helps you set realistic tolerances that can be used to monitor the performance of your model in production." (12:38)
From deployment to ongoing monitoring
Before deployment, teams should confirm they have mitigated the risks identified upfront: informed consent, a clear process for contesting outcomes and explainability that is fit for purpose. These decisions belong in a living report for downstream users of the system and the team maintaining it, along with recommendations such as levels of aggregation below which outputs become less accurate.
Production is not the end of the effort. Rouse recommends tracking performance across subpopulations over time, adding thresholds as alerts or retraining triggers, and giving business stakeholders visibility.
"Embedding your ethics KPIs within your existing MLOps framework alongside other performance metrics is the ideal way to do this." (13:58)
Her closing message is simple: begin with a basic framework, learn from each iteration and make ethical performance visible.
"Doing something is infinitely better than doing nothing." (14:30)
Watch the talk
Operationalising AI Ethics by Natalie Rouse (15 minutes): https://www.womentech.net/video/operationalising-ai-ethics-natalie-rouse-0
About the speaker
Natalie Rouse is General Manager of Eliiza, a data science consultancy specialising in well-engineered, practical and ethical AI, ML and data science solutions across Australia and New Zealand. With an engineering background, she has worked in optimisation, data science, analytics and innovation across five continents, most recently leading analytics and AI delivery at Qantas.
Questions this talk answers
- Why is agreeing on AI ethics principles not enough?
- Where in an AI project should ethical review happen?
- How can low-risk AI projects avoid unnecessary ethics bureaucracy?
- How do you decide whether accuracy differences between groups are acceptable?
- What happens to ethics once a model is in production?
Drafted with AI from the talk transcript. Editor review pending.