AI promises multiple benefits, so we are currently testing in what fields we can apply it. But sometimes we might have to slow down the implementation a bit.
One of the fields we are testing AI is in court. We already had an AI program that, based on vast amounts of data sets, would give a verdict. This verdict was compared to the verdict of the actual judge, and in 79% of cases the AI had the same verdict (Johnston, 2016). Even in some American courts, AI is already used to help decide if, and how long, someone should be jailed (Judges Now Using Artificial Intelligence to Rule on Prisoners , 2018). Again, the computer analyses thousands of previous cases and will base its verdict on measurements it has learned from previous cases.
The problem is that the AI is only basing his verdict by analysing previous cases. But where is the human aspect if we let a machine base its verdict on information of previous cases?
The mere purpose of a court system is to prevent someone from making the same mistake again. Actually, if the defendant shows remorse, a judge is tended to reduce the sentence. This reduction is because the judge can interpret the remorse as a sign the defendant will not make the same mistake again (van Doorn, 2013). Even though AI is making improvements on interpreting language, emotion and image recognition, those are still the fields a computer has the most problems with to interpret (Brynjolfsson & McAfee, 2017). Let that be exactly what a judge in a courtroom uses to assess if the defendant shows signs of remorse.
An even bigger problem with AI in court is how it bases a probability score about a defendant based on analysing previous cases (Judges Now Using Artificial Intelligence to Rule on Prisoners , 2018). This measuring has already been tried when Bayesian statistics was used in court. I explicitly write was used, since for example the English Court of Appeal banned the use of probability measurements like Bayesian statistics or the Sherlock Holmes doctrine. The problem with measuring with statistics is that the argument with the highest probability will be used as explanation, just because other arguments have a less high probability (Spiegelhalter, 2013). By this reasoning an unlikely explanation might be used as the leading explanation, because statistics say so. While we stopped with using those probability statistics in court we now introduce AI, which does the same in a more sophisticated manner.
Since AI gives in the most cases the same verdict as a judge AI will have a good use in court in the future. But until the moment that AI can evaluate the defendant as good as a judge, and that we found a way around the probability problem, we must leave the final verdict up to a human.
Brynjolfsson, E., & McAfee, A. (2017). The Business of Artificial Intelligence. Boston: Harvard Business School Publishing Corporation.
Johnston, C. (2016, 10 24). Artificial intelligence ‘judge’ developed by UCL computer scientists. Retrieved from The Guardian: https://www.theguardian.com/technology/2016/oct/24/artificial-intelligence-judge-university-college-london-computer-scientists
Judges Now Using Artificial Intelligence to Rule on Prisoners . (2018, 02 07). Retrieved from Learning English: https://learningenglish.voanews.com/a/ai-used-by-judges-to-rule-on-prisoners/4236134.html
Spiegelhalter, D. (2013, 02 25). Court of Appeal bans Bayesian probability (and Sherlock Holmes) . Retrieved from Understanding Uncertainty: https://understandinguncertainty.org/court-appeal-bans-bayesian-probability-and-sherlock-holmes
van Doorn, B. (2013, 08 15). Spijt betuigen in de rechtbank: ‘Als dader kan je het beter maar wel doen’ . Retrieved from Omroep Brabant: https://www.omroepbrabant.nl/nieuws/163061/Spijt-betuigen-in-de-rechtbank-Als-dader-kan-je-het-beter-maar-wel-doen
Interesting article, Sander! Perhaps it’s best for all of us to not let artificial intelligence take care of our justice system for now. Although, artificial intelligence suffers from biases just as we humans do. Isn’t it normal for artificial intelligence to make the same mistakes that we humans make? The best solution could be in a collaboration between human judges and artificial intelligence. Let us call it supervised use of artificial intelligence. The artificial intelligence provides sentencing based on previous cases and sentences, whereas the human judge corrects this sentence based on the remorse that the defendant is showing. Let me know what you think!
Hello Sanders, thank you for your article, the use of AI in the court is also a topic that really interest me.
You pinpoint the one major flaw that AI indeed deals with in its decision-making process: it cannot make use of intuition to take into account sincere emotions from the defendant or notice an unlikely explanation that needs to be reconsidered and make the case re-evaluated.
Yet, humans are under the influence of biases that are not as obvious as the lack of intuition in AI, but as frightening : David Khaneman recalls in his book academic experiments that show that judges can make poor decision-making, having dramatic consequences on the defendant. The disconcerting fact is that it is not because of lack of expertise that judges make mistakes, but just because they are under the influence of cognitive biases that not even the best judge can avoid. It has been found that prisoners have 65% of chances of being parolled if the judges just had lunch, next to almost 0% two hours later, when the judges are hungry (attentional bias). Khaneman also recalls the disturbing anchoring bias experiment. This bias appears when one makes an illogical association between what he/she has observed in the past and let that association rules his/her actions. In the experiment, judges before going to court somehow see a dice being rolled. It was discovered that judges would give a sentence that is the number that was the outcome of the dice rolling, regardless of the facts of the case!
Perhaps we do notice these biases since we are also passive victims of them in our everyday life but they have as dramatic and dangerous consequences as the lack of intuition of AI. The best solution would be to combine the skills of both humans and AI in court, with AI playing the more suitable role of advisor (with the judge being more qualified to give critical feedback), especially when it has been a long day for the judge and can rely on AI to not give in to cognitive biases.
References:
Zoe Corbyn. 2011. Hungry judges dispense rough justice. [ONLINE] Available at: https://www.nature.com/news/2011/110411/full/news.2011.227.html. [Accessed 19 October 2018].
Khaneman, D., 2011. Thinking Fast and Slow. 1st ed. United States: Farrar, Straus and Giroux.
Hi Sander,
Thanks for the interesting read. The case you described is clearly an illustration of the main limitation concerning AI implementation. AI systems are based on prior statistical truths and contain biases that are (un)intentionally incorporated in the training material. An example related to your topic is Northpointe’s COMPAS (Correctional Offender Management Profiling for Alternative Sancations)1. The algorithm often predicted white defendants to at a lower risk of recidivism. Within two years, analysis indicated that white defendants who recidivated were nearly as twice as likely to be misclassified as a lower risk compared to back defendants. Other examples of flawed algorithms are the AI used in Nikon cameras and word embedding, an AI used to process large amount of language data2. AI in Nikon cameras tends to interpreted Asian as ‘always blinking’ and word embedding systematically characterized European American names as pleasant, whereas African American names as unpleasant.
Viewing these examples, we can consider AI as a replication of previous work. I argue that the challenge is not embedded in the technology, as the AI becomes biased because of the data input. If the data or training material is of a quality that does not conform our desires, we should expect that the output will be the same. A biased system withdrawing advice from a biased AI that is trained with biased data, will only lead to biased results – further strengthening the status quo. I do not think that slowing down implementation in combination with leaving the final verdict to humans (for the court system) is part of the solution in acquiring the desired result. It will yet slow down the creation of more biased data, as biased output data will create loops in which current output will become input for future outputs.
I thus argue that the solution resides in obtaining and creating unbiased and neutral (as far as possible) training material as input for AI functionality.
1 https://www.propublica.org/article/how-we-analyzed-the-compas-recidivism-algorithm
2 https://www.nature.com/articles/d41586-018-05707-8