Early data security concerns with off-the-shelf generative artificial intelligence solutions prompted the Queensland government to pursue its own purpose-built internal chatbot, QChat.
The internal tool, which is based on OpenAI’s large language model (LLM), is currently being rolled across the public service after three months of testing with select state government agencies.
Since it went live at the start of the month, half a dozen departments have already taken up the tool, which will help reduce the burden of writing briefing papers, aggregating information under different lens of analysis, and helping to refine emails.
Chief digital and customer officer Chris McLaren told InnovationAus.com that the decision to develop QChat was a part of the government’s attempt to ‘learn by doing’, improving the capability of the public service in the process.
“There’s still some unknowns around some of the products, particularly around data in transit, data in storage, and there’s a risk that users and owners and builders may make some incorrect assumptions about what’s happening with that data,” he said.
“Our preference is not to discover that when something bad happens, so we’ve looked closely at Bing Chat Enterprise, at [Microsoft] Copilot. We’re currently not using either of those products.”
Mr McLaren said that advice from the Office of the Victorian Information Commissioner, which asks agencies to consider the adequacy of Copilot’s privacy and information security, as “on the money”.
When generative AI tools exploded into popular consciousness at the start of last year, Mr McLaren said it was an inevitability that it would be used across the public service.
“Rather than blocking and shutting things down, the best thing to do from a policy development and enablement [is to] make something a safe place for people to go and use this capability.”
While the AI platform underpinning QChat is currently powered by OpenAI’s LLM and was developed with support from Microsoft engineers, it has been designed to be compatible with any LLM.
Mr McLaren added that all data inputs to QChat are “stored in our own tenancy onshore” and is fully compliant with privacy laws and the state’s record keeping legislation.
The tool is also compliant with the state’s guidance on the ethical use of generative AI which is aligned with an assurance framework recently agreed at the national level.
The platform “learns how you’re using it and so…will get better at your use cases and it will actually have prompts that are aligned to the way you use it”, Mr McLaren said. This also includes personalised instances of QChat that are tailored to the use-habits of individual public servants.
“If I opened up my instance of QChat and looked at it and we opened up somebody else’s version of QChat, we would actually see slightly different pre-prompts because it’s learning how we’d like to use it.
“If I’m continually asking it to write me a briefing note or do a draft privacy impact assessment or something like that, it’s going to learn to do that a lot.”
Despite allowing public servants to use QChat according to their preferences, there are pre- and post-prompt guardrails to prevent use-cases that don’t comply with the state’s ethical use policy, he said.
This could result in the chatbot refusing or amending a request in circumstances where there has been an attempt to automate decision-making or approval processes.
Staff in government agencies making use of QChat have also been put through “integrated training and onboarding”, according to Mr McLaren, with executive training programs currently being run for non-technical senior business leaders across government.
Development of the platform has required a “fairly modest” resource commitment, with many existing public servants from different government agencies contributing to the Department of Transport and Main Roads project.
Mr McLaren said it was difficult to give a figure on the total headcount of people who worked on the project as there were many safety experts and AI engineers that moved “in and out part time”.
“It’s been really quite inspiring to see our team really lean into this and we’ve honestly been getting a lot of interest from the market, AI engineers, [and] data engineers, who want to come and work with us because they see what we’re doing,” he said.
“I think this is one of those opportunities where the return on investment that we’ve got for the sector and the government is disproportionately advantageous relative to the amount of time, effort, and money that we had to put into it.”
Several other use-cases are also being built on top of the state’s underpinning AI platform, although they are still at an early stage.
