A lot of attention has been paid to the mischief done by “agentic” large language models (LLMs) on their own initiative. But the more immediate and commonplace danger is that AI will multiply and indulge existing human weaknesses. AI does not have to be super intelligent to provide an excuse for secretive and poor-quality government decision-making.
Wed 11 Nov 2026 16.33 AEDT
Photo: AAP Image/Dan Himbrechts
A lot of attention has been paid to the mischief done by “agentic” large language models (LLMs) on their own initiative, like the OpenAI agent that “breached” the Medicare website while scraping the internet for data and the swarm of OpenAI agents that hacked a rival AI company. A sufficiently powerful and intelligent AI could escape the limits imposed by its creators and any attempt by government regulators to constrain it.
But the more immediate and commonplace danger is that AI will multiply and indulge existing human weaknesses. AI does not have to be super intelligent to provide an excuse for secretive and poor-quality government decision-making. Consider the earlier case of Robodebt, where human responsibility and accountability were abandoned in favour of supposedly objective and impartial processes.
To be clear, the problem is not that AI would make a “mistake” like Robodebt. Even one of the earlier ChatGPT chatbots can spot the logical flaw within Robodebt that made it unfair and illegal. It was humans, in their cruelty and indifference, who decided on Robodebt, but it was automated data-matching that made it possible.
AI magnifies and multiplies human fallibility. Whereas humans have a track record they can be held accountable to, AI is a black box.
Government agencies have already adopted AI to justify their arguments, without proving that there is any substance behind those claims. The NSW Productivity Commission claims that emerging technologies like AI “could potentially increase productivity growth in NSW to 2.0 per cent a year and lift the growth rate of real Gross State Product to 3.0 per cent a year.” The modelling was “powered by Faethm AI”, a reference made many times throughout the report but not satisfactorily justified or explained. The appeal to AI seems intended to awe and deflect, not enlighten.
Similarly, Tasmanian Planning Minister Felix Ellis supported a contentious luxury hotel’s bid for major project status on the basis of the company’s job creation claims. Treasury questioned those claims, but the Department of Premier and the State Planning Office said “that it had deemed the job numbers ‘reasonable’ after using Microsoft’s Copilot AI assistant.” It is not clear how Copilot is equipped to do that kind of analysis.
Management consultants at Deloitte used AI to produce a report “riddled with artificial intelligence-authored errors” for government. Deloitte only offered a $98,000 refund for the $440,000 paid for the report.
Ironically, governments are now cutting out the middleman and going straight to generative AI to provide consultant-like services. This may be an improvement on paying expensive management consultants, but it would be even better to bring the work (and expertise) in-house to be done by humans.
Quality research, analysis, explainers and factchecks from experts you can trust.
AI is also being offered as a poor substitute for more staff and better funding. Hospital staff are reportedly using ChatGPT to calculate medication doses because they are understaffed and overworked. Chatbots are “helping teach” psychology classes at Macquarie University. These shortcuts paper over deeper problems.
AI detection of seatbelts being not worn or improperly worn has led to a great deal of unnecessary grief in Queensland and Western Australia, with court time consumed in dismissing “inaccurate” and unreasonable fines.
AI is also used to frighten and encourage compliance. Last year, the Albanese Government claimed, without evidence, that the freedom of information system was being overwhelmed by many “AI bot-generated requests” and that AI “means it is possible for someone to disrupt an agency completely”, in this case the eSafety Commission. On that basis, the government sought to introduce new fees for FOI requests, process FOI requests more slowly and defy the findings of the Robodebt Royal Commission by making government decisions even less accessible than they were during Robodebt.
In fact, the requests in question were generated by real people, using a form “which is more or less 1990s technology”. It would be extremely concerning if a few hundred FOI requests somehow brought the $43 million-a-year eSafety Commission to its knees.
Fortunately, this attempt by the Albanese Government to make government more secretive failed. But as concerns about AI grow, governments will be tempted to use it as a bogeyman.
AI has already given humans, with all their frailty and flaws, an excuse for poor service, secretive government, unaccountable modelling and laziness. This extends even to their attitude to AI itself. Last year, almost 40% of federal agencies had not filed their AI transparency statements by the deadline.
While there is no one solution to these problems, they are at least familiar ones because they are the product of human weaknesses, not an alien, synthetic intelligence. A good start would be to hold ministers and public servants to account for government use of AI, to fund public services properly and to finally put an end to the use of management consultants for core government work.
Bill Browne is Director of the Democracy & Accountability Program.
