Decisions of the courts around bail applications have recently become a hot topic, being front-page news on a regular basis.
Largely, this reporting concentrates on decisions in which the courts appear to have got things wrong, based on the fact that defendants on bail are being regularly re-arrested for similar offences. That is not the case – the decisions being made are correct based on what the courts have before them.
Moreover, being arrested is not equivalent to guilt rarely gets a mention in the news stories, but the evidence certainly suggests that a significant portion of people given bail are re-offending while on bail. The reasons for this are no doubt complex, but the problem for the legal system is that the public seems to be coming to the view that the system is getting it wrong – and they are losing faith in it.
This is an over-simplification of what is occurring, and few people appreciate the difficulty facing magistrates and judges when making decisions on bail applications. They are working under heavy workloads and with limited information; perfection in such circumstances is impossible.
That said, what if there were a tool that could overcome these pressures, and allow courts to make better decisions on bail applications? To incarcerate only those who are likely to reoffend, and let low-risk defendants back into the community?
There is in fact such a tool, although it involves the current bogeyman of the legal profession: AI. The problems of AI are well known and oft discussed in the legal world, but there are benefits as well, and getting to more accurate bail decisions may be one of the best of them.
Almost 10 years ago, a team led by Sendhil Mullainathan (currently Professor at the Massachusetts Institute of Technology) trained AI models to make bail decisions, based on 758,027 real bail decisions. The team had access to the current offence, prior offence and prior failures to appear, and the only demographic used was the age of the person charged. They also knew whether the person was released, failed to appear or was rearrested.
When tested, the results of the model were impressive. In short compass, the model outperformed human decision-makers. Depending on the level of risk set, the model could reduce crime rates by 24-42 per cent, because the model would release fewer people who actually reoffended.
Even more importantly, given Australia’s concerning rates of incarceration or First Nations people, the model – which avoids demographics that can be highly correlated with race, such as postcode – jailed 41 per cent fewer people of colour than actual judges, based on the same risk rate.
Is such a model perfect? Of course not; and this model was based on data from the United States, a very different legal system from any here in Australia. That said, a model trained on Australian data – presuming it showed the same rate of success – would be a powerful tool in the hands of judges and magistrates tasked with the difficult prospect of making bail decisions under time and volume pressure.
Lawyers have a general aversion to AI, and given its long record of hallucination it is easy to see why. It would be irresponsible, however, not to investigate the possibilities of this technology; if it speeds up bail applications, helps to get them right, and reduces the number of First Nations people being incarcerated, why wouldn’t we at least look?
(NB: this article is the personal opinion of the author and draws on independent research, including the book Noise, by Daniel Kahneman, Oliver Sibony and Cass R Sunstein)
© Shane Budden 2026
Share this article