The Australian Securities and Investments Commission (ASIC) has discontinued an experiment to create machine-generated summaries of public submissions using generative AI, after the results were found to be of lower quality compared to those produced by human staff.

The pilot of an in-house generative AI system was developed in tandem with Amazon Web Services using the Meta Llama 2 large language model.

ASIC has been exploring the use of artificial intelligence (AI) to streamline its operations, as revealed during a hearing of the Australian Senate Select Committee on Adopting Artificial Intelligence (AI). Joe Longo, Chair of ASIC, and Graham Jefferson, Digital and Legal Transformation Lead, shared insights into the organization's AI pilots and their outcomes.

ASIC has been using machine learning for some time and has around 20 algorithms registered in its inventory of machine-learning technology that support the work that it does.

It recently completed a pilot focused on using a standalone large language model to summarize submissions received by ASIC.

"The pilot was a success, but the results from the large language model wouldn't be something that we would want to use going forward," said Jefferson.

"We did a comparison between what the large language model generated and what our staff generated, as a blind comparison - a proper experiment, I guess. The results weren't sufficiently good for us to want to use that summary technique in that particular way."

“Basically, we took the large language model, established it within an [offline] environment, took the submissions that we wanted summarised, passed them through the model and then compared the results from the summarising technology, the 50-odd summaries, to human summaries of the same submissions.

Jefferson noted that the AI-generated summaries were generic and lacked the nuance captured by human employees.

"What we found was that in general terms … the summaries were quite generic, and the nuance about how ASIC had been referenced wasn't coming through in the AI-generated summary in the way that it was when an ASIC employee was doing the summary work."

Longo said, “In ASIC's world we're often having to read and absorb submissions because we consult heavily with the market, and so we ran a pilot to see whether the technology could 'read' all those submissions and come up with an analysis that was accurate, to save hundreds of hours of human time to do that.”

He described the AI-generated summaries as "bland" and not misleading.

"It really didn't capture what the submissions were saying, while the human was able to extract nuances and substance," said Longo.

ASIC has also participated in a whole-of-government Microsoft Copilot pilot coordinated by the Digital Transformation Agency (DTA).

"We had 150 ASIC staff using that for about a month, and we're now reviewing the results of that experience, asking surveys and getting feedback from our staff about what was good, what wasn't and how it worked," Jefferson told the Select Committee.

Calissa Aldridge, ASIC’s Executive Director, Markets, described the take-up of AI by Australia’s banks and financial services industry as “cautious but exploratory.”

“There is a lot of focus on internal capability, productivity enhancements and middle- and back-office functions, looking at how it can be used to improve their own compliance, looking at client onboarding and those types of practices,” said Aldridge.

“There's a lot of caution around automated decision-making and the engagement and communication directly through to clients. Obviously, we see the use of chatbots and other types of AI and that has been deployed quite broadly, but there is a lot of caution around automating decisions in other parts of client-facing businesses. We do see, in some areas, some communications that are being automated.

“One of the big international investment banks has been talking globally for some time about how they're using AI to automate communications to clients. When there's significant volatility in the market, for example, or there's a significant event, they want to be able to communicate very quickly to all their clients and give them an update on their portfolio. We're seeing some of those sort of things.

“We've seen the use of generative AI for some time for a range of other initiatives in markets, like using it for portfolio rebalancing, for creating synthetic data to help design systems, for testing, for informing algorithmic trading, and for sentiment analysis and predicative capabilities to look at what's happening in social media to help inform some of those algorithms. Those developments have been in place for some time. But, in terms of that consumer-facing use, there has been significantly more caution,” said Aldridge.