Monday 25 September 2017 9:45am
University of Otago researchers are warning of the potential pitfalls in government departments using computer-based risk prediction models, as has been recently revealed through a controversial new tool used by the Accident Compensation Corporation (ACC) to profile and target clients.
The ACC uses a computer model to assist staff managing claims, however, details of what the model does, and how it is used, are somewhat sketchy, says spokesperson for the University’s Artificial Intelligence and Law in New Zealand Project, Associate Professor James Maclaurin.
The most concrete information comes from a short press release issued by ACC indicating that the tool is used to make three types of prediction:
- Which clients are likely to need help and should be called proactively,
- Which type of case owner should assist the client, and
- How long we should expect a claim to be managed.
“This somewhat vague description leaves open the possibility that ACC uses these predictions to minimise treatment times, either by intervening in patients’ treatment, or (more seriously) by declining applicants with long predicted treatment times,” says Associate Professor Maclaurin.
“The tool makes predictions about future ACC cases using a database of information about 364,000 past claims that were lodged between 2007 and 2013. ACC stresses that details about individual cases are kept private both from ACC staff and from other agencies,” says Department of Computer Science’s Professor Alistair Knott, another researcher on the project.
Management still appears to be ultimately under human control: an ACC ‘case owner’ makes the final decision about each case.
“But ACC workers find themselves in a situation increasingly common in our society: their decisions are guided by advice generated automatically by a machine, based on a large set of data extending far beyond their own experience.
“We are in the same position when we use Google’s navigation system in our cars, or choose a book based on Amazon’s recommendations. In these cases, having a computer in the decision-making loop seems innocuous enough. It seems less innocuous when it guides the agencies whose decisions have serious consequences for people’s lives. We set out six questions that such agencies must be able to answer about the algorithms they use.
“Of course it is a fundamentally a good thing for people’s decision-making to be informed by statistics. Systems like the ACC tool can be quite accurate, but because they don’t reason in the way humans do, it is essential that Governments and companies relying on them are able to answer the following questions:
- How accurate is the tool, in fact? Predictive tools are easy to evaluate: if the public is to have confidence in the ACC tool, and the courts are to evaluate its use, the agency should give a public account of how it was evaluated. Without divulging personal details, this should include a thorough description of the data set on which it was assessed.
- Can the agency explain the way this tool works so that clients could appeal particular decisions?
- Does the use of this tool distort the way the agency pursues its stated policy objectives?
- By ‘passing the buck’ to the machine, is the agency to ducking its responsibility to make fair and humane decisions about treatment of New Zealanders in need?
- Does this tool implicitly discriminate against individuals on problematic grounds such as age, ethnicity or gender? Importantly, there is simply no way to remove this bias without compromising the accuracy of its predictions. There is a real risk that the ACC tool unfairly discriminates against some clients. This possibility needs to be explored in an evaluation of the system.
- Are employees effectively trained in the use of the system? While the predictive system is probably intended to be used as a guide, to supplement the case owner’s own knowledge and judgements, it is easy to fall into ‘autopilot’ mode when guided by a system—especially if it is fairly accurate. (The dangers of autopilot mode are particularly clear if the system’s errors are prone to bias, of the kind just mentioned.) The charge of ‘falling into autopilot mode’ is often levelled at judges using the US predictive system.
“Predictive analytics technologies show great potential in informing public decision-making, but it is important for these technologies to be evaluated and scrutinised when used in the public domain. It may be that ACC has addressed the issues we raise in its own internal training and evaluation processes. But we are calling for ACC to provide a public account of how it uses its predictive tool, so as to maintain the integrity of its decision-making,” says Faculty of Law Professor Colin Gavaghan.
For more information, contact:
Associate Professor James Maclaurin
Associate Professor Colin Gavaghan
Associate Professor Alistair Knott
The authors are investigators on the Artificial Intelligence and Law in New Zealand Project at the University of Otago, funded by the New Zealand Law Foundation.
A list of Otago experts available for media comment is available elsewhere on this website.
Electronic addresses (including email accounts, instant messaging services, or telephone accounts) published on this page are for the sole purpose of contact with the individuals concerned, in their capacity as officers, employees or students of the University of Otago, or their respective organisation. Publication of any such electronic address is not to be taken as consent to receive unsolicited commercial electronic messages by the address holder.