AI & Criminal Justice Reform

1937 Ronald Coase: transaction costs are a central determinant of how economic activity is organized.

1997 Ronald Gilson: Imperfect markets give rise to intermediaries to lift the wedge between parties. “Lawyers are transaction cost engineers.”

2015 Nicole Shanahan (at Stanford CodeX): Technology supplements lawyers as transaction cost engineers. Technology is the ultimate transaction cost economizer.

Origins: I wanted to understand what my job as a lawyer was

What the article actually says is this:

When we shift focus from thinking about legal technology in terms of a lawyer’s

efficiency, to viewing these advancements within the context of socioeconomic

organization, we can begin to realize its true significance.

Borrowing from transaction cost theory, there should be 3 core tenets of legal technology:

1. Optimizing for the exchange of information.

2. Setting consistent expectations between parties.

3. Mitigating risks.

Our job as modern legal technologists is to build software that mimics the cognitive processes of lawyers. We expect that we can produce faster, cheaper and more accurate legal work products.

In the context of Criminal Justice

“Predictive Policing”

Prosecutor Discretion Tools

FOR THE FIRST TIME EVERTHIS IS ALL TECHNICALLY FEASIBLE

SO, WHAT DO YOU NEED TO UNDERSTAND ABOUT CRIMINAL JUSTICE AI?

General AI

MachineLearning

Logic/Rules Automation

General AI

MachineLearning

Logic/Rules Automation

DATADATA

DATA

DATA

DATA

DATA

DATADATA

DATA

DATA

“Bad” AI

MachineBias

Harmful Automation

UNDER WEIGHTED DATA

OVER WEIGHTED DATA

DATA WITHOUT POLICY

OBJECTIVES

NOISYDATABAD

DATA

MISSING DATA

Beliefs: A hungry person is allowed to steal.I have never felt sad about things in my life.

Life Status:How often to do you feel bored?How often do you have barely enough money to get by?How often have you moved in the past 12 months?How old were you when your parents separated?Have you ever been suspended or expelled from school?

If your score was high/positive in these categories, you were are more likely to be predicted to reoffend.

However, black defendants who don’t reoffend are predicted to be riskier than white defendants who don’t reoffend – this where the algorithm breaks down.

This is because attributes that predict reoffending vary by race.

General AI

MachineLearning

Logic/Rules Automation

Computa-tionalLogic (1) the representation of facts and

regulations as formal logic and

(2) the use of mechanical reasoning techniques to derive consequences of the facts and laws so represented.

Computa-tionalLaw

General AI

MachineLearning

Logic/Rules Automation

Super-vised

Learning

Unsuper-vised

Learning Training Data Hand-Labels “These e-mails

exemplify willful in-fringement”

Clustering “these e-mails have similar expres-sions of willfulness”

Dimensionality Reduction

General AI

MachineLearning

Logic/Rules Automation

Super-vised

Learning

Unsuper-vised

Learning(Deep) Neural Networks

Super-vised

LearningUnsuper-

vised Learning

30 Million Positions from previously played Go matches used as training data

It then began to play itself, creat-ing more data for “reinforcement” learning.

General AI

MachineLearning

Logic/Rules Automation

Painful and Slow

General AI

MachineLearning

Logic/Rules Automation

SUDDEN

The Future of Computational Criminal Justice

Making a bad system is easier and more likely than making a good system.

Making a good system requires us to incorporate “policy controls” on each and every algorithm. Think “computational policy”

Policy is more important today than it has ever been because of the reach of modern day computing.