Operationalising Responsible AI
CSIRO had committed to Australia's AI Ethics Principles, but science teams had no practical way to act on them. This was research-stage work with a brief to reach TRL 3: proof that the approach could work, not a finished product. I led the design.
The brief: proof of concept, not product
Technology readiness levels (TRLs) run from 1, basic principles observed, to 9, proven in real use. TRL 3 is a characteristic proof of concept: evidence that the core idea works, before anyone builds it for real.
That set what success meant. The design work wasn't expected to ship, scale or win adoption. It had to answer one question well enough to decide whether the idea deserved further investment: can ethical principles become concrete checks that science teams can act on while they build AI systems?
My role
Design lead. I designed and facilitated the workshops, conducted interviews, then synthesised what we found, and designed the POCs.
The team: Researchers x 11.
The problem
The principles are fairness, transparency, accountability, privacy and human oversight. They're widely endorsed, but they don't tell a science team what to build, test or measure.
The stakes are high in research. Errors can spread through published literature. Bias in training data can quietly favour some research directions over others. In multi-agent systems, it's hard to trace how a recommendation was reached.
What the research changed
The design research I performed with domain experts interrogated the ethical risks of AI systems applied to specific use cases. Scientists interviewed were working in Life Sciences and Materials Science across agriculture, astronomy, and water research.
We also mapped how trust flows between the people involved when a team uses Responsible AI methods. The mapping revealed the complex webs of accountability flows, who teams relied on, and where trust broke down.
The major finding of the design research was that ethical considerations surrounding AI in science was extremely use case specific. The evidence helped bolster the position that there are multiple and somethimes conflicting considerations for the responsible use of AI in Science.
Findings from the work fed into an empirical study, Investigating responsible AI for scientific research.
Key decisions
1. Design each proof of concept to answer one question
At TRL 3, a prototype is an instrument for learning, not an early version of a product. Each proof of concept was built to test a handful of assumptions. For example, 'Can an AI Footprint be measured with targets and controls mapped across the AI lifecycle?'
2. Start from real systems, not the principles
Workshops worked through the risks of specific AI systems rather than discussing the principles in general. A proof of concept only proves something if it's tested against real conditions.
3. Treat a narrow result as a finding
The first proof of concept was built on a CSIRO ethical-AI risk assessment framework. Testing it showed that a risk lens alone was too narrow, so the second widened it to include other forms of governance. At this stage, learning what doesn't work is a result in its own right.
AI Governance proof of concept, with targets and controls alongside the risk register.
4. Make it look real enough to get honest reactions
The demos I created were higher fidelity than TRL 3 strictly needs. My rationale for this was so stakeholders would react to the idea rather than the sketch. They can be thought of as speculative design probes, that suspended disbelief for the viewer so they could imagine that particular future scenario.
AI Footprint mock-up
What the proofs of concept showed
Success at TRL 3 meant teams could test whether their systems mapped to the targets and controls, while stakeholders could discuss tangible approaches worth developing.
What I'd do differently
As this project started before the release of ChatGPT, it was difficult for scientists to imagine what risks might be involved when using AI for science. If I had my time over, I would have honed in on the existing 'narrow AI/ML' use cases that were common within CSIRO. This approach would have given me a better current state view of the AI for sceince landscape.