The Queensland and South Australian state governments are forging ahead with practical guidance and early sanctioned uses of generative AI that run on top of the Azure OpenAI service.
Details of the two states’ use cases, and leadership in the space, are included in this year’s State of IT, an annual review of state and territory ICT programs by iTnews, which officially launches today.
Given the rise of generative AI technology in the past year, this year’s State of IT asks the states and territories where they are at with policies, experiments and production use cases.
An emerging theme from a technology perspective is the use of – or recommendation to use – the Azure OpenAI service as a kind of ‘private ChatGPT’.
This is the initial basis of QChat, an internal Queensland government GPT tool powering experimentation at multiple agencies, as well as of EdChat, a public pilot run by South Australia’s Department of Education to test how the technology could help students with their studies.
QChat
The Queensland government has stood up what it is calling QChat, which is initially based on the Azure OpenAI service but is intended to be technology agnostic in terms of the technology that powers it in future.
QChat briefly broke cover at a Queensland government-run ICT forum last month, where executive director of data and information service Nathan Bines described it as “an internal government GPT”.
“It’s private, based on the OpenAI Microsoft service for now, but we’re building a frontend that should be technology agnostic,” Bines said.
“The idea is that public servants can access these tools, starting with just the chat interface. It’s secure, it’s logged, we have single sign-on for government, we can understand usage, we can report on it.
“No data is travelling outside our environment so they can upload briefing notes and those sorts of things.”
Bines said that QChat is in private preview “with a few departments”, but that “the intent over time is to roll it out” more broadly.
Speaking to iTnews, chief customer and digital officer Chris McLaren said that the government is also building safe ways to plug into the centrally-run generative AI capabilities and tooling.
The working name for this is the QG AI platform; this would set up some AI products “for identified core and common use cases”, according to the October presentation, as well as guardrails for access and capability augmentation.
“We need to test and learn essentially,” Bines said.
“We know there’s products coming on all of the vendors’ roadmaps. We need to make sure that we understand where we can play a role.”
McLaren said that “the real excitement is going to happen when we dive into use cases.”
“There’s a huge opportunity in regtech – we’re doing some interesting things with Safety Queensland around using digital and AI to optimise that process.
“There are really interesting use cases to remove some of the burden on tasks within government and uplift the productivity of employees, and then [move to] customer- and possibly cyber-focused uses.”
McLaren said that an AI community of practice numbering around 400 people already exists.
“[They] are all leaning in and getting involved in how we can use this for better service delivery, better policy development, regulation reduction and so on,” he said.
South Australia’s EdChat
South Australia’s Department of Education ran an eight-week trial of an Azure OpenAI-backed tool called EdChat that was completed in August.
The forward-thinking trial sought to test how generative AI could positively aid learning opportunities for more than 1500 students and teaching by 150 educators from eight schools.
The trial is a sharp contrast with other jurisdictions where generative AI use in schools is banned.
The SA-based trial is interesting in part because responses are tailored to the year level of the student engaging with the bot.
“If I log in as a year eight student … it will deliver results suitable for a year eight student, which might be a different level of sophistication than if a year 11 or 12 student asked for information,” Dr Eva Balan-Vnuk, executive director of ICT and Digital Government in the state’s Department of the Premier and Cabinet (DPC) told iTnews.
The bot also offers students guidance on how to craft better prompts, familiarising them with use of generative AI tools.
An expansion of EdChat is now on the cards, according to a case study recently published by Microsoft.
Specific Azure guidance
The popularisation of generative AI has produced a wide spectrum of government responses.
While many err on the side of being overly generic, South Australia has gone the other way – releasing specific, targeted guidelines [pdf] on how to engage safely with generative AI and large language models (LLMs) in a government context.
That specificity includes a nudge towards Microsoft’s Azure Open AI service, which is establishing credentials as a safe space for enterprise and government experimentation.
The state’s Office of the Chief Information Officer (OCIO) first provided general guidance to SA government agencies at the end of May, before issuing the more detailed guidelines.
The specificity of the guidance is by design, Dr Eva Balan-Vnuk said.
“Agencies were saying, ‘Please give us guidelines because our colleagues are asking how we can use this responsibly and safely’,” Dr Balan-Vnuk said.
“We really thought there was some value in the specificity. We have been as specific as possible because that makes it very grounded for people. I think people can really understand what their obligations are and what’s safer for how they use this.”
The guidelines note that “to reduce the information confidentiality risk associated with the use of open access LLM AI, agencies should consider using Microsoft’s Azure Open AI service.”
“The Microsoft Azure OpenAI Service gives customers advanced LLM AI with OpenAI GPT-4, GPT-3, Codex, and DALL-E models with the security and enterprise service of Azure,” it states.
“Moreover, Azure OpenAI does not use enterprise data to train the model.”
However, the guidelines add that as Azure OpenAI is not available within the Australia region … use cases would need to be carefully considered for out-of-region [cloud] access, given that the inputs are stored in region cache for 30 days as part of [Microsoft’s] data abuse detection and monitoring facility.”
About Author
