Last week, Australia’s Fair Work Commission hit international headlines when its general manager released a report showing an estimated 70 percent increase in claims of unfair workplace behavior in the past three years. The increase, it asserts, is in part driven by the proliferation of generative AI assistance tools. The commission, which handles unfair dismissal claims, wage disputes, discrimination, bullying, and workplace sexual harassment, says the surge is directly affecting its ability to provide timely dispute resolution. It is the country’s first major tribunal to publicly frame generative AI as a contributing factor in its workload crisis. It is unlikely to be the last.
At the core of the problem lies the very feature that has driven much uptake of digital technologies dating to the genesis of the internet (and even before that): the simple fact that digital tools have reduced the costs of doing things that used to be very—and sometimes even prohibitively—expensive. Digitization has democratized activities once reserved for the privileged classes wealthy enough to pay the high prices required in an analogue world—from publishing books and releasing recordings of creative endeavors to sending mail. As the unerring laws of economics remind us, reducing the cost of these activities to nearly zero leads to a near infinite increase in their volume. The scarce resource now becomes the human time to view (or review) them.
In the world of judicial and quasi-judicial proceedings, generative AI tools have appeared to democratize access to justice by eliminating the need for aggrieved—or even just malicious—parties to engage the costly services of a lawyer to draft their pleadings. The result, as the Fair Work Commission observes, is an increase in filings. Grievances that previously would not have been pursued because legal costs were a barrier now proceed. Detailed analysis undertaken by the commission reveals that not only are there more cases but the AI-generated ones tend to be longer, contain polished-sounding but generic content, omit key data, and contain hallucinations, meaning they take longer to process than those generated by skilled human lawyers. And that’s before considering the allocation of hearing time to the increased number of cases with some merit that pass through initial vetting.
As an initial step to address the matter, in March the commission issued draft guidance requiring anyone who uses generative AI in preparing documents for cases before it to disclose that fact. The guidance warns that AI-generated information may be incomplete, inaccurate, or fabricated. A new section on the use of generative AI will be built into all commission forms. It is also reportedly trialing a new system in which senior staff help parties try to resolve disputes informally earlier in the process, before cases consume full hearing time.
Unsurprisingly, though, given the limited budgets and human resources the commission must operate under, it is also turning to AI tools to manage increasing workloads It is considering deploying an AI voice agent to help triage calls to its helpline. Australia has already backed AI in other parts of its legal system, including a government-supported chatbot that helps splitting couples divide their assets. The logic is sound: If generative AI is increasing the volume of inbound work, automated triage may be the only way to keep pace without proportional increases in staffing that budget constraints have already ruled out.
But are there other ways of constraining the load? The price of legal advice previously stood as the constraint ensuring the marginal value of cases brought to the tribunal, just as the price of a stamp constrained the volume of mail sent in pre-email times. As the volume of spam emails increased in the early 2000s, Microsoft proposed a contingent “Penny Black” charge on emails sent: Legitimate senders would be refunded or wouldn’t pay if their email was accepted by the recipient; spammers would lose pennies on emails rejected so would bear a cost for their (mis)use of the system. The intention was to reduce the number of spam emails sent in the first place, thereby reserving the (at the time more) constrained resources of the internet for legitimate traffic. A similar “tax” could now be put on AI-generated filings, to be refunded if the case were accepted. Filers would face a price signal similar to that provided by the legal costs constraining existing cases, thereby ensuring that only cases valued over the price of the tax would be pursued. This would preserve scarce tribunal resources for the most valued cases.
As we proceed deeper into an AI economy where scarce but expensive human resources are required to buttress against the excesses of (near-costlessly generated) AI spam, such taxes offer an effective price signal that is otherwise being lost.
