Electronic Frontiers Australia is disturbed to learn that the Australian Government is incorporating ChatGPT 4.o into a pilot program for AI use in the public service.
The Department of Finance previously announced that it was exploring AI use cases across government; however, no specific mention was made as to whether this AI was to be “generative AI” or “agentic AI.”
In July, ITNews announced the use of Chat GPT 4.o within the Commonwealth Public Service but failed to articulate how the AI would be trained or deployed. This was clarified here to small degree where the author raised considerations of ethics, AI and ambitions of an AI rollout across the whole government.
ChatGPT 4.o, at this stage, is still under development and is not yet ready, in EFA’s opinion, to assist with government decision-making.
Part of the fundamental problem with adopting AI to make decisions is that the law is designed for human decision-makers. If an AI tool forms part of government decision-making in any way, it could expose the government to legal actions similar to the Robodebt administrative law actions.
Administrative lawyers have expressed concern that transparency is required in relation to decisions made within government. If a decision is challenged under the Administrative Decisions (Judicial Review) Act 1977, the decision-maker must be able to articulate the facts, matters, and circumstances that were taken into consideration; those that were excluded; and the extent to which the affected party was afforded procedural fairness/natural justice, such as receiving timely information about the decision-making process and having opportunities to interact with the decision-maker.
In cases where a decision, or part of a decision, is made by AI, the decision-maker would need to be able to explain the weightings assigned by the algorithm to various considerations, as well as the precedents the AI relied on and the data it was trained with.
AI, including ChatGPT 4.o has been notoriously bad at interpreting legislation and the law.
In fact, lawyers who have used generative AI to assist in preparing submissions, pleadings, or other legal documents have faced serious consequences, including being struck off (or, as the American Bar would say, disbarred). Australia is already over-represented in global instances of AI hallucinations in legal matters.
Bearing in mind that each and every decision which may affect a member of the public is a potential legal case, it remains unclear how the Commonwealth Government has developed such strong confidence in ChatGPT 4.o, or whether it has sought advice from the Solicitor-General regarding the potential liability arising from its use. The concern is this: if using ChatGPT 4.o is such a good idea, then why isn’t it being proudly promoted by the Commonwealth Government? And why can’t the public access its procurement documents (or at least use cases that aren’t subject to classification restrictions?
EFA asks: will the government be transparent and accountable in its use of ChatGPT 4.o? Where are the procurement documents? What use cases are currently in deployment? What high risks were identified during deployment, and how were they mitigated?
Let’s not forget Robodebt, while the ink is still wet on the government’s settlement cheque…