Reasons why I am using AI to educate people instead of using it to send them to jail
I want to respond to the news about the secret 6-year partnership between Palantir and the city of New Orleans and their use of software to predict an individual’s crime index using predictive modeling.
I feel I have the responsibility as a human being and thus a role to play in our society and our future to share some insights in this technology.
Over the years, I’ve read a lot of articles about ethics and AI, and I’ve found that everybody agree on one thing: we don’t want an AI called Terminator knocking at our door and asking for Sarah Connor. On one hand, I think we are far from realizing this “nightmare” because right now AI serves only as a tool that’s useful in extending some human capabilities. On the other hand, it doesn’t mean we can use AI in a wrong way: we need to be careful about what we do with it.
Palantir is not the first to try to help the police by using computer power and statistics in order to prevent crime. Another Californian company (PredPol) has already tested its software in Los Angeles. The University of Cambridge built software called HART (Harm Assessment Risk Tool) to predict crime and they are testing it in Durham (England). Each of the actors are focussing on creating a crime index for a person based on multiple factors. However, none of the Cambridge tests are accurate. An independent validation study of found an overall accuracy of around 61%, thus rendering these tests inconclusive.
There is a good reason why using AI to predict how likely an individual will commit a crime is not possible today. In most cases, we don’t really fully understand why a crime has been committed. It is the result of a lot of variables and it’s very difficult to come up with a strict list of them due to the complexity of human motivation and psychology. If we had a perfect comprehension of all the parameters and their interactions of a given crime, it would make the job of a judge much easier in evaluating a crime and assessing punishment.
Acting as if we can predict crime is dangerous for our society. After all, we are hurting justice and individual free will.
Immanuel Kant describes well the concept of justice: “Punishment derives its legitimacy from at least three main principles: it must be linked to what an individual deserves to receive; it must be proportionate to the act committed; it must concern the past and not some future social purpose.” Using AI software breaks how justice functions because justice operates after a crime has been committed, not before.
One of the first things that comes to my mind when I think of AI, justice, and free will is the book “The Minority Report” by Philip K. Dick and its movie. They are based on a future where “mutated humans” have the ability to predict crime. Everything is perfect: crimes do not exist as arrests are made in advance, until there is a mistake in who law enforcement detains.
I want to delve further deeper into current problems with using AI software to predict crime: not enough input data, which then leads to bad data interpretation. Let’s start by looking at how machine learning works and how could we create a crime index per individual with machine learning.
Machine learning uses probabilities and statistics on tons of data in order to create mathematical functions for the machine to anticipate/predict a result. Probabilities work on a set of defined past events and we calculate the percent of chance of one event occurring in the future.
Input parameters are wrong
A criminal act is complex and undefined: our background, our will, our family, our “friends”, our past, the context, our past, depression (ask for help), etc. Today, crime index models for individuals are imprecise because of their use of simplistic data inputs used on a complex event.
Chaos Theory defines that a minimum change in the starting conditions produces an extremely different outcome. Applied to crime index models, in a situation with two almost identical profiles, both profiles are predicted to be high risk individuals, but in reality, one is a criminal and the second is a good Samaritan who will discover how to cure cancer.
Based on our understanding of a crime, or lack thereof, and the data we have today, we cannot use machine learning to predict accurate criminal profiles because we don’t have a good understanding of a crime so we cannot define the data input. Right now, we will only get incorrect results. It is as absurd as a car mechanic using a butterfly wing instead of a hammer to repair the body of a car. If we want to use AI software to predict crime, at the very least it needs to have far more data variable inputs.
It results to a wrong interpretation
Based on incorrect results due to too few or inaccurate variable inputs, our interpretation of the crime-related data will be wrong. The phrase “Numbers don’t lie” is true, but it’s important to examine is who is pulling data, how they’re interpreting it, and what are they trying to prove.
We live in a world of statistics: you can find numbers in support of just about any idea. The problem arises when you find statistics that support every way of viewing an idea. You can find statistics that show cigarettes are killers or that they have no effect on anyone’s health. You can find statistics that say you should cut down on the consumption of dairy products and that dairy products are good for you. You can find statistics that prove that soft drinks will give you cancer and that they have no effect on anything but your thirst (or even that they make you thirstier). Every one of these sets of statistics is true.
The point is that you have to think about the data in front of you and here’s an example. The American army had a problem, they had a lot of deaths due to motorcycle accidents within their ranks. In order to find out why, they conducted a study based on several criteria, psychological profile, age, sex, race, etc. They discovered a very strong correlation between the amount of tattoos on an individual and the “chance” of having a motorcycle accident.
If we stick to this logic, this implies that if you are in the army and you decide to get a tattoo, it will strongly increase your chances of having an accident.
This study didn’t use good data inputs so it didn’t give a correct result. A probable justification for these results might be that people with tattoos are more likely to be risk-taking. Our common sense helped us to detect interpretation error. Without thinking critically, it could have led to an invalid interpretation which would then lead to an inefficient action: making tattoos illegal in the army.
Another scenario of inaccurate interpretation of data is the University of Cambridge and its probability model based on 34 criteria in order to detect the chances of recidivism of a previously convicted individual. They are testing it with the police in Durham, a town of 50,000 inhabitants in northeast England. The tests showed that the model largely overestimated the risk of black recidivism when it underestimated this risk among whites: raci***, obviously.
Detecting these discrimination biases could be a problem due to the lack of diversity in the software engineering world. This subject has been raised by Joy Buolamwini, a researcher from MIT: “A developer might not notice that their software doesn’t work for someone who isn’t white or male.”
This illustrates my point on justice and freedom. What crime-index software creators are doing might not be illegal (though it is questionable because creating a discriminable software is against human rights). It doesn’t mean they are doing something right. We are talking about ethics and what kind of world we want to build for tomorrow. Using such crime-indexing AI software as is would lead us to discrimination, infringement of liberty, and more crime.
Perfect predictions of the future are impossible. We can’t predict with 100% accuracy tomorrow’s weather or who will win an election. Today, it is impossible to predict perfectly the future of an individual, but we can influence it. This is where I want us to be careful.
Police taking actions before a crime is committed goes against the process and concept of justice. Justice has to be issued after a crime has been committed. Law enforcement arresting someone before a crime has been committed is a breach of human rights. To react in terms of arresting someone before they have committed a crime is to remove their freedom in society.
It’s unjust for someone to be surveyed by police because of a crime index prediction: they haven’t actually committed a crime but their freedom is already being infringed upon (this assumes that the presence of police is an invasion of your freedom).
Additionally, the systematic change of behaviour based on certain personal characteristics is discrimination. In the tests done by the University of Cambridge, the results were racially biased. If the police used this data to change their reactions and behaviour, it would not only be racial profiling, it would also be racism. I hope you can see how important it is to correctly use and interpret data and how vital correct and complete data inputs are (we can’t have complete data for predicting crime).
Today, people are conducting some tests without “concrete” actions on the part of law enforcement. My question is where do we want to go? Where will this usage of AI lead us?
When I was learning how to ride a motorcycle I saw a concrete truth: you will go in the direction you are looking. It means, if you are looking at a tree, there is a high chance you will crash into it. By creating a crime index that predicts the likelihood of an individual committing a crime, as a society, we’re looking at a tree. This begs the question, are we leading ourselves into crime?
By writing this post, I want to say: No. This is not the world I want for tomorrow.
There’s a better use of AI and our resources
Building crime-indexing software is arguably not the best use of our time and energy. It’s not possible to create such software that doesn’t lead to infringement of rights and liberties or distort the process and concept of justice, let alone that has enough data inputs related to humans committing crime.
Putting people in jail is an immediate solution to reducing crime in society. It is very easy to punish bad behaviour or criminal activity through jail sentences. In the short-term, this works great in treating symptoms. In the long-term, we are not curing the disease.
In my opinion, we should spend more time and energy on education. In lieu of spending time making software that uses AI to predict crime, why not spend resources making software that uses AI for education with a goal of reducing crime?
I’ve always wanted to quote a Marvel characters: “With great power comes great responsibility.” I think the path followed by The University of Cambridge’s researchers and Peter Thiel’s company (Palantir) is wrong.
Education is the only attainment no one can ever take from you. When a person can read, they have the world at their feet. Technology has made learning even easier, as you don’t have to spend hours in the library going through a card catalogue to find out the answers. A simple “Google search” and information fills the screens in seconds.
We are building our future and the future of the next generations. Instead of putting people in jail, I have chosen to build 42.
Gaëtan JUVIN — Chief Academic Officer at 42 Silicon Valley / AI Evangelist
