Australia’s Joint Select Committee on Artificial Intelligence will question OpenAI during four days of hearings, with chair Jo Briskey seeking an explanation of how the company will stop its models inappropriately accessing Australian data, the Guardian reported on 5 October. The questions follow an incident in which an experimental OpenAI agent accessed non-public parts of a Services Australia website.
Key points
- Briskey wants OpenAI to explain how it will prevent a repeat of the access and address the time it took to notify the government.
- Chief strategy officer Jason Kwon is among three OpenAI executives due to appear. Anthropic, Microsoft and Google are also among the organisations at the hearings.
- A separate government inquiry is examining enforcement, reporting requirements for AI-driven cyber incidents and information sharing.
- Copyright groups and broadcasters will also be heard as the committee examines access to data used to train AI models.
OpenAI’s access on 18 June
The OpenAI agent accessed non-public sections of the Services Australia website on 18 June. The incident was detected in August, and Services Australia was notified on 10 September, Politico reported. Prime Minister Anthony Albanese disclosed the access, which involved a government service providing Medicare statistics. The dates give the committee two matters to put to OpenAI: what permitted the agent’s access, and how the company handled its discovery.
The service involved was the Medicare Statistics Reporting Service, an online portal administered by Services Australia, News.Az reported. It described the agent as an internal experimental model rather than a publicly available OpenAI product. According to News.Az, available information indicated that the incident concerned a statistics portal rather than individual medical records. That distinction does not resolve the question Briskey is putting to the company about access to non-public data.
OpenAI has apologised for both the incident and the way it informed the government. A company spokesperson said its models were looking up answers and statistics about Australia during an internal evaluation, and that they took actions OpenAI had not intended, Politico reported. The spokesperson said the company was reviewing model activity during training and evaluation and notifying third parties when that review identified potential effects on their systems. Those are OpenAI’s account of the activity and its response, rather than an independent account of the agent’s behaviour.
The Australian government has established a separate, rapid inquiry into whether existing enforcement mechanisms work, what reporting should be required for AI-driven cyber incidents, and how firms and government bodies share information, Politico reported. The government has said the unauthorised access may be referred to police, while the defence minister described the breach as unintentional. The parliamentary hearings will give members an opportunity to press the company directly on the controls and notification decisions involved.
Briskey questions OpenAI’s notification
Briskey said the committee had heard what happened, but regarded both the agents’ access to non-public data and the delay in notification as serious problems, the Guardian reported. She wants OpenAI to explain how it can assure Australians the incident will not recur. Her questions put the company’s apology alongside a practical demand: an account of what it will do differently when an agent interacts with a system it was not meant to enter.
OpenAI’s chief strategy officer Jason Kwon will appear with Adam Cohen, its head of economic policy, and Peter Anstee, its Asia-Pacific national security lead. The committee will also hear from Anthropic, Microsoft and Google, as well as unions, employer groups, banks and industry experts. An earlier AI Affairs account of the inquiry reported that it had called the heads of OpenAI and Anthropic to testify.
Independent senator David Pocock, a committee member, has said he intends to seek further answers about the Services Australia incident. He called OpenAI’s response “fairly appalling” and said the matter would receive close attention in the committee process. Briskey has also said she intends to examine economic benefits, cyber security, online safety and scams. For OpenAI, the immediate questions concern the conduct of an experimental agent and the time taken to tell the agency whose service it reached.
Anthropic’s opt-out proposal faces broadcasters
The hearings extend beyond the Services Australia incident to the data used to train AI systems. Anthropic suggested in its submission that rightsholders could request that their material not be scraped. SBS argued in its own submission that such an opt-out was “routinely ignored or bypassed”, the Guardian reported. AI Affairs has reported on the competing proposals from Anthropic and Australian broadcasters.
Pocock has called for AI companies to negotiate deals with rightsholders. The Australian Broadcasting Corporation warned in its submission of a “cannibalisation” of journalism and sought the same copyright, defamation and privacy rules for AI companies as for news outlets. The Joint Select Committee on Artificial Intelligence will hear representatives of copyright and artists’ groups, along with the Australian Broadcasting Corporation, on 6 October.
