📚 It’s incredible how much has changed in just a year! When I delivered the opening talk at last year’s Generative AI for Government Summit, we were still experimenting cautiously and exploring the technology’s potential.

This year, in the second installment of the conference, I had the chance to share CSIRO’s Data61 lessons learned from real-world deployments. My talk, titled “Harnessing Responsible AI for a Fair and Effective Public Service,” focused on the growing maturity of generative AI in industry and government.

🥇 Three key takeaways from this year:

1. 𝐒𝐭𝐚𝐤𝐞𝐡𝐨𝐥𝐝𝐞𝐫-𝐃𝐫𝐢𝐯𝐞𝐧, 𝐌𝐮𝐥𝐭𝐢-𝐏𝐫𝐢𝐧𝐜𝐢𝐩𝐥𝐞 𝐓𝐫𝐚𝐝𝐞-𝐨𝐟𝐟𝐬: Improving privacy can sometimes increase fairness, but in other cases, it may reduce it. Balancing transparency, privacy, accuracy, robustness, and fairness is a complex trade-off, but our science aims to quantify these factors, enabling stakeholders to make informed decisions. https://lnkd.in/gGwBTMaZ

2. 𝐌𝐞𝐚𝐧𝐢𝐧𝐠𝐟𝐮𝐥 𝐇𝐮𝐦𝐚𝐧 𝐎𝐯𝐞𝐫𝐬𝐢𝐠𝐡𝐭: Human oversight shouldn’t become a “liability sponge” or “accountability sink.” As AI capabilities surpass human abilities in some areas but remain unreliable occasionally, we must provide humans with the right methods and tools to oversee AI effectively. Our research at Data61 focuses on empowering meaningful oversight. https://lnkd.in/gPhid9tX

3. 𝐒𝐲𝐬𝐭𝐞𝐦-𝐋𝐞𝐯𝐞𝐥 𝐀𝐬𝐬𝐮𝐫𝐚𝐧𝐜𝐞 𝐚𝐧𝐝 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐓𝐡𝐫𝐨𝐮𝐠𝐡 𝐒𝐲𝐬𝐭𝐞𝐦𝐚𝐭𝐢𝐜 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧: AI learning extends beyond model training—it also involves learning in the scaffolding, out-of-model components, and external data (e.g., RAG or instructions, context data) unique to your organisation. By conducting carefully architected system-level evaluations, organisations can harness the evaluation infrastructures to create a fast learning feedback loop for their AI systems (even using third-party AI models). https://lnkd.in/gzVUmKi4

You can see a selected set of slides here. Across these lines, we are actively seeking government and industry partners to co-innovate and trial our technologies. If you’re interested, let’s connect! 🤝

Slides on dropbox