AI ‘judge’ can predict court verdicts with 79 per cent accuracy
telegraph.co.uk
telegraph.co.uk
You can't optimize for overall accuracy, otherwise you'd have classifiers that always just predict guilty (or not sick, or not fradulent, and so on) because they'd achieve 95%+ accuracy.
In the end the probability cutoffs you choose are more often than not based on business context and costs associated with various actions. For example - if we're trying to predict whether a customer is going to leave or not, I'd choose a different threshold based on whether we're mass e-mailing people, or if we're having customer service reps call people, since the underlying costs for each action are so different.
In the first case it's fine to cast a wide net, and e-mail lots of people who we're less certain about, in the second we'd optimize to target people we're most sure are going to leave, since it costs so much more to reach out to them.
That's terrible accuracy. Imagine this algorithm would then be refitted to replace the judge (or augment judges in a way that judges mostly rely on the tool to make a decision, which is already starting to happen [1]) and automatically put people in prison - with only 80% accuracy. Does that number still seem high now?
[1] - https://www.propublica.org/article/machine-bias-risk-assessm...
However, do we have any reason to believe that a collection of random, qualified judges would be able to predict the outcome of a court verdict with 80% accuracy? I would be surprised if judicial decisions are not stochastic.
Also, we should consider a substitution effect. According to a cursory internet search, 'As of 2012, the pay for federal district judges was $174,000 per year'. If this system is substantially less expensive to purchase and maintain, then there is an argument to be made that it is more efficient, and provides better value, even with a perpetually reduced accuracy.
I think a random selection of second-year law students could probably beat 80% (by about 19 percentage points) predicting the outcome of a court decision from the facts and law sections of the same decision.
In the long run, such a model would likely be confounded by the parties, who'd model and adapt their arguments to drive the model to a favorable result or settle cases where the model provides a strong signal to which they can't adapt. (Which is not to suggest that this wouldn't be a very socially valuable result.)
Cases predicted to win might be taken on contingency, for a fraction of the judgment award. Cases predicted to lose might be pay-as-you-go for each appearance or filing.
I've heard of it at Google and Bing for popular searches. I've seen it firsthand at a company applying classifiers to twitter posts. Microsoft took how long to go from "80%" voice recognition to their stated 96% (with generous assertions about the environment and language)? Apple/Google/et.al. still rely on humans to act as validators for the training of thir personal assistants.
I have trouble seeing for-profit companies ever truly getting higher than the 80%; it's just too expensive when cheap labor also exists.
"High" is a relative term. At one point, a chess program beating any human was an impressive feat. Today, any human beating a chess program is an impressive feat.
There is definitely a possibility of bias in the characterization of the facts. The authors state that for their appellate court, the facts section is uncontested by the parties and derived from the lower court's findings. Even so, it is not clear whether the court's reversal rate is high or low. If low, it could be that the lower court's presentation of the facts was equally varnished.
With some judges, it is possible to predict the outcome of the case from the very first sentence laying out the facts and circumstances. In a famous dissenting judgement by Lord Denning, in which he attempts to save a cricket field, the judge begins as follows:[2]
In summertime village cricket is the delight of everyone.
Nearly every village has its own cricket field where the
young men play and the old men watch. In the village of
Lintz in County Durham they have their own ground, where
they have played these last 70 years. They tend it well.
The wicket area is well rolled and mown. The outfield is
kept short. It has a good club house for the players and
seats for the onlookers. The village team play there on
Saturdays and Sundays. They belong to a league, competing
with the neighbouring villages. On other evenings after
work they practise while the light lasts. Yet now after
these 70 years a judge of the High Court has ordered that
they must not play there any more.
Even with this critique, I think that this research is an excellent first step. It would be great to use pleadings, which are available with effort from some appellate courts. It would likely be necessary to OCR some PDF files.I doubt that this textual approach would work for the more technical parts of the law. Many cases are in areas with very few preceding cases from which to train an n-gram based algorithm. While the author's approach worked for certain human rights cases, it would likely fail for cases turning on an obscure tax provision.
[1] Original paper: https://peerj.com/articles/cs-93/
[2] Miller v Jackson [1977] QB 966, see: https://en.wikipedia.org/wiki/Miller_v_Jackson
I think the authors list down the "Accuracy" but I would be really interested to see "Precision, Recall and F1" scores for this classification task.
"The team found that judgements by the European Court of Human Rights are often based on non-legal facts rather than directly legal arguments, suggesting that judges are often swayed by moral considerations father than simply sticking strictly to the legal framework."
This is not like classifying pictures between obvious classes, but something more complex