Attorney-General's Dept backs Copilot Chat, Google NotebookLM

The Attorney-General’s Department has set up two controlled AI tools, one each from Microsoft and Google, for staff to use on data classified up to the ‘protected’ level.

Antony Spence, right, speaks at Google Cloud Summit Sydney.

Assistant secretary in charge of IT support, engagement and innovation Antony Spence revealed the department’s approved AI setup at the Google Cloud Summit in Sydney.

The department has stood up Microsoft Copilot Chat and Google NotebookLM for staff over the past year, Spence said.

“These two products are certified at the protected level for us. They're in our own controlled [cloud] tenancies, and what it now allows staff to do is actually use AI with their data.”

The department’s earlier experiments with AI followed the trajectory of other federal government implementations, where tools could be used only on information already in the public domain.

These implementations were of limited utility to staff, as most of their data could not be processed by the tools.

“We went through our own mini hype cycle [with AI]. People got excited about it, then they started using it, they could only use public data, and it kind of just went out of favour,” Spence said.

“Most people went, ‘That's great. I've got actual work to do, and I can't use my data, so come back to me when you've got something else’.”

Copilot Chat and NotebookLM in controlled spaces are that “something else”, and have afforded staff the ability to use AI on data that is contextually relevant to their line of work.

“What we've got now is people being able to actually do their work, have a look at AI opportunities, and have a much more grounded conversation in that space,” Spence said.

The tool choice, and use of the tools, is being guided by the AI plan for the Australian Public Service, which was developed by the Department of Finance.

“This was the framing strategy to ensure that we're engaging with AI and using it safely,” Spence said.

Use of the tools is admissible with reference to a set of “policy guardrails” that the Attorney-General’s Department has put together.

“There's absolutely no point giving people strict rules with AI. AI moves too quickly, and use cases evolve too quickly, and what we're also finding is that if we try to set the traditional public service policies of ‘thou shalt do this, thou shalt not do that’, they age out too quickly, and we just can't keep up,” Spence said.

“So, we leant into our privacy law background. And in privacy law, you don't set rules, you set guardrails. So we now have guardrails for the use of AI. 

“These are policy guardrails - they're not technical guardrails - but what this allows people to do is [access] simple guidance so that they can use their context to play with the tools.”

Users of the tools have to “verify the generated outputs” for accuracy - keeping humans in the loop - and “own the final results”.

Spence said that AI “supplements” work but does not automate it.

“We do not produce anything as final versions that come straight out of AI - and I'm going to go one step further. At this stage, there's no automated decision-making within our organisation,” Spence said.

“We don't create any final versions using AI. It supplements our work, it augments our work, but humans own the start and finish and throughout the process. So we can't say, ‘Sorry, AI created that. I don't know how it came [about’]." 

“The public doesn't accept that; our senators at [senate] estimates won't accept that. It's really important for us that we own whatever comes out of there. That's what all of you expect, and that's what I expect as a citizen as well.”

The department is also recording its uses of AI tooling.

“We keep a record of where we're using AI, how we're using AI, and what it's actually doing for us,” Spence said.

“That does a couple of things. That goes to the core of public service. We have to have a record of what we do. We have strong obligations around records keeping and being accountable to the public, so we've built in some processes by which staff can record their use of AI. 

“We are trying to balance that with practicality, so we don't record every single use of every AI tool ever, in the same way I don't record when I jump on Google Search and I look up some legislation, I wouldn't record that. 

“But if it's creating a draft speech for me, I definitely note that [down]. If it's creating an image, those kinds of things.

“Again, we've deferred this to the line [of business] areas for them to add their context and work out where that makes the most sense for them.”

Users of the tools are also encouraged to report “unexpected anomalies”.

“‘Anomalies’ isn't a bad word. I think anomalies, especially in AI or any new tech space, are unexpected outcomes,” Spence said.

He said anomalies could include gaining access to more data than expected, or instances of a tool not responding “the way you want it to”.

“Being able to report those anomalies allows us to build this big corporate knowledge between our areas,” he said.

“If there are really good use cases for AI, we can learn from that. But I [also] love learning from the ways it doesn't work well, because that allows us to tune the messaging, support people, help them understand where actually maybe AI is not the right tool, or maybe this version of AI is not right - try a different tool, try a different approach.”

Ry Crozier attended the Google Cloud Summit Sydney as a guest of Google Cloud.