Double-header day on my favourite topic — 𝗱𝗲𝘀𝗶𝗴𝗻𝗶𝗻𝗴 𝗵𝘂𝗺𝗮𝗻 𝗼𝘃𝗲𝗿𝘀𝗶𝗴𝗵𝘁.
• Remote presentation at the Symposium on Responsible AI in Public Administration (organised by the wonderful Jason Grant Allen, bringing together law and AI experts in the same room in Singapore)
• Fireside chat with Rasika Mohan at the AI Governance Summit

There’s a joke in the AI community: 𝙩𝙝𝙚 𝙢𝙤𝙢𝙚𝙣𝙩 𝙮𝙤𝙪 𝙛𝙪𝙡𝙡𝙮 𝙪𝙣𝙙𝙚𝙧𝙨𝙩𝙖𝙣𝙙 𝙝𝙤𝙬 𝙖𝙣 𝘼𝙄 𝙬𝙤𝙧𝙠𝙨, 𝙞𝙩’𝙨 𝙣𝙤 𝙡𝙤𝙣𝙜𝙚𝙧 𝘼𝙄 — 𝙟𝙪𝙨𝙩 𝙗𝙤𝙧𝙞𝙣𝙜 𝙖𝙪𝙩𝙤𝙢𝙖𝙩𝙞𝙤𝙣. If you can fully understand how an agent comes up with a plan or an answer, and you have complete control over how it operates, it’s no longer really an autonomous agent — it’s automation.

𝗛𝗲𝗿𝗲’𝘀 𝘁𝗵𝗲 𝗼𝘅𝘆𝗺𝗼𝗿𝗼𝗻 𝗶𝗻 𝗴𝗼𝘃𝗲𝗿𝗻𝗶𝗻𝗴 𝗮𝗴𝗲𝗻𝘁𝗶𝗰 𝗔𝗜:
Governance demands transparency and control, while true agentic AI involves human giving up some understanding and control.

Many oversight mechanisms try, with mixed success, to force understanding and control back into an agent, effectively turning it into automation. That doesn’t work.

𝗧𝗵𝗲𝗻 𝘁𝗵𝗲 𝗵𝗮𝗿𝗱𝗲𝗿 𝗰𝗵𝗮𝗹𝗹𝗲𝗻𝗴𝗲: 𝘄𝗵𝗮𝘁 𝗶𝗳 𝘆𝗼𝘂𝗿 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁 𝗶𝘀 𝘀𝗺𝗮𝗿𝘁𝗲𝗿 𝘁𝗵𝗮𝗻 𝘆𝗼𝘂?
Many oversight mechanisms often assume the oversight role is to catch mistakes. But great leaders manage people far smarter than themselves every day. Oversight here is rarely about catching errors, but it can still be deeply meaningful, just in different ways.

That’s why this line of work at CSIRO’s Data61 excites me the most. The future of oversight is not about hoping the AI slips up so you still have a job. It’s about adding value without being smarter than AI — using tools, scaling yourself, growing your understanding, and shaping outcomes. That’s the real human productivity edge.

Stay tuned.