Why I’m not giving robojudges a plug
Trying to turn law into robust code is far from black and white
Some years ago, a friend and I talked about setting up a website for dealing with straightforward and mundane legal cases where the law worked as flowchart, in a IF THIS; DO THAT kind of process.
The idea was to save costs, anything legal being ridiculously expensive and to speed up the process with less bureaucracy and form filling, as justice delayed is justice denied.
Better access to justice was another drawcard, although it would require internet access which at the time wasn’t ubiquitous as it is today.
Efficiency and progress would be ours once more, along with greater fairness and less human-introduced bias; this seemed a very worthwhile project.
For cases that didn’t fit in with what we had envisaged in code, there would be a button to call for human help. The software even had a snazzy name: Robojudge.
“Oh hell. Robojudge is awful,” said Brenda Wallace, developer extraordinaire who, unlike myself and my friend, has actually turned law into computer code.
Wallace is right. Robocop, Robodebt, and other variations of the name evoke cold and inhumane machine systems that crush people with no empathy. It makes for dramatic movie scripts but isn’t acceptable in real life.
That said, as Wallace points out, legislation is a set of rules. Computer code is also a set of rules. Law gets turned into computer code all the time. Have you used an accounting program? Bought things off the web, or at a store with computerised point of sale systems?
Lots of everyday apps have the law embedded in them, ditto when you apply for government entitlements online.
A rules engine is how Wallace describes her team’s work for the Department of Internal Affairs, and you can find examples of it on the web at nz.openfisca.org. The work inspired other countries including Singapore.
It’s not a black and white process though. For example, finding out what that exact intent might have been for an older law still in effect is often difficult as the experts who wrote it some decades ago are now retired or gone.
A programmer looking at a legal outcome that severely disadvantages poor people might feel that this can’t be right, but there’s nobody to ask.
Maybe that was the intention, maybe it wasn’t.
Times change along with people’s politics and sentiments, but legislation that stays static and married to the era it was created in can be
What could possibly go wrong if you take an AI . . . and feed it with decades worth of legal data from a justice system that’s biased against people of colour, the poor and minorities?
problematic as any lawyer will tell you.
With new law, one big issue for turning it into good computer code is that the people affected by it aren’t present “when the system is actually implemented by a software nerd at her keyboard, working for a government agency”.
This leads to edge cases being missed, and people risk falling through the cracks, along with absurdities like multiple definitions of marriages and spouses, depending on which law is referred to.
“It's less of a computing project and more of an overarching concept to ensure legislation can be implemented and does what was intended, by bringing the people affected and the people who implemented the rules into the same room and actually working together instead of each being isolated from each other,” Wallace said.
Technology marches on at a stunning pace, and we now have automated systems that can take colossal amounts of data, more than humans could cope with, and run it through algorithms to create rules.
That’s artificial intelligence and machine learning with neural networks of course. Make no mistake, AI/ML is capable of amazing things that simply aren’t humanly possible. That’s after all the reason why we build machines, to do things we can’t perform ourselves.
Better rules are needed, ones that are tested and constantly improve and it’s tempting to think AI/ML is the silver bullet.
Scientists in the AI/ML camp really do believe that’s possible, going as far as positing Minority Report- style pre-crime facial recognition software for law enforcement.
Said to have 80 per cent accuracy and no racial bias, the notion of using facial recognition to catch criminals before they do their deeds alarmed over 1500 other researchers so much that they persuaded science publisher Springer to pull the paper that described the system.
I mean what could possibly go wrong if you take an AI that’s set up to do what its programmers want it to do, and feed it with decades worth of legal data from a justice system that’s biased against people of colour, the poor and minorities? Especially in the United States where many states have the death penalty.
Would cops check if the rules for the AI were sane or just attempt to meet their key performance indicators, especially in tense situations?
That’s the thing really: if we trust the machines and don’t check on what we feed them, a “Robojudge Dredd” of sorts could become reality.
If you haven’t come across the Judge Dredd comic books, it’s a sharp British satire in which the main character is a “law enforcement judicial officer” who can arrest, convict, sentence and execute criminals.
Like the Idiocracy cult comedy that depicts a future in which the human population has become morbidly stupid with a political system attuned accordingly, Judge Dredd was never meant to be a manual for the future even if we now have the technology to make the ideas in it way worse than imagined in the past.