Learning AI #22

Should AI be used to predict crime?

We posed this question to Claude and Gemini. The AIs agreed that using AI to predict a crime is a bad idea.

Gemini produced the following table to help us understand the issues.

So, crimes can be predicted based on crunching a lot of data, but those data include historical biases. If more crimes were committed in a zip code, the algorithm will train itself on that data and tell us a crime is more likely to happen there versus the neighboring town.

Same goes for the demographics of an individual. If a person is part of a demographic group that has a high likelihood of criminal behavior, that person will be “scored” as a probable criminal.

Claude did a nice job of nailing the real issue:

The entire structure of Western criminal law is built on the idea that you are innocent until proven guilty, meaning until you do something. Predictive AI inverts this. It says: based on who you are, where you live, and what people like you have done, you are more likely to commit a crime. That's not justice. That's a profile.

In Minority Report, the great science-fiction novel and later a Tom Cruise movie, the “pre-cogs” were able to truly see the crimes before they happened so Cruise could rush in and save the day.

Using today’s AI models in crime prediction will not save the day, but ruin the lives of many innocent people.

Things I think about

The ocean floor contains more than one million volcanoes, more the rest of the Earth combined.

Free gift for my subscribers

Here is the link for a free audiobook chapter from my new book, No Adult Left Behind. AI is Learning. Are You?

There is no login or anything like that, just click and listen. The chapter is all about how to use prompts with AI. A good starting point for anyone who is serious about learning AI. (The audiobook is narrated by my AI-cloned voice.)

Response from Claude

Should police be allowed to use AI to predict who will commit a crime before they do it?

No — and the reason isn't that the technology doesn't work. It's that even when it "works," it's doing something we shouldn't want done.

Predictive policing AI is trained on historical arrest data. But arrest data doesn't reflect who commits crimes — it reflects who gets arrested. Those are different populations. Police have historically over-policed Black and low-income neighborhoods, so the training data reflects that bias, and the AI learns to flag the same people, which leads to more arrests in those neighborhoods, which reinforces the model. It's a feedback loop dressed up in math.

The deeper problem is the presumption of innocence. The entire structure of Western criminal law is built on the idea that you are innocent until proven guilty — meaning until you do something. Predictive AI inverts this. It says: based on who you are, where you live, and what people like you have done, you are more likely to commit a crime. That's not justice. That's a profile.

The Philip K. Dick version of this — Minority Report — at least had the decency to make the predictions certain. Real predictive AI doesn't even offer that. The RAND Corporation studied one of the most widely used tools (PredPol) and found it performed barely better than chance.

So you'd be stripping people of their presumption of innocence, supercharging existing bias, and building a surveillance apparatus — all for a tool that's barely more accurate than a coin flip.

The answer is no.

Response from Gemini

The question of using AI for predictive policing—specifically to predict who will commit a crime before it happens—is one of the most intense debates in modern law enforcement and technology ethics.

It forces us to balance the potential of technology to prevent harm against core constitutional rights and human biases. While the idea sounds like science fiction, the technology exists and is being deployed in various forms, sparking massive controversy.

Here is a breakdown of the core arguments on both sides, followed by how these systems actually work.

The Core Debate

Arguments for Allowing Predictive AI (The Proponents)

Proponents argue that predictive AI is simply a highly advanced version of traditional crime analysis.

• Proactive Resource Allocation: Law enforcement has limited resources. If an algorithm can identify high-risk areas or individuals with high accuracy, police can deploy patrols to deter crime before it happens, potentially saving lives and reducing property damage.

• Objective Data vs. Human Gut Instinct: Supporters suggest that a well-designed algorithm, relying strictly on objective data (like crime frequency, timing, and location types), can be less biased than a human officer's subjective "hunch" or personal prejudices.

• Crime Prevention: The primary goal of law enforcement is to prevent victimization. If data can intervene—perhaps by triggering social services or community outreach to high-risk individuals—it moves policing from reactive to preventative.

Arguments Against Allowing Predictive AI (The Critics)

Critics, including civil liberties advocates and data scientists, argue that the technology poses severe risks to justice and civil rights.

• The "Feedback Loop" of Biased Data: AI is trained on historical data. If past policing practices disproportionately targeted specific minority neighborhoods, the AI will learn that those areas are "high risk." It sends more police there, who make more arrests, creating a self-fulfilling prophecy that reinforces systemic bias.

• Erosion of "Presumption of Innocence": Labeling a person as a "future criminal" fundamentally undermines the legal bedrock of innocent until proven guilty. Acting on a prediction violates civil liberties by treating potential future actions as present justification for surveillance or stops.

• Lack of Transparency ("Black Box" Problem): Many predictive tools are proprietary software owned by private companies. Neither the public, defense attorneys, nor sometimes even the police know exactly what variables the AI uses to make its predictions, making it impossible to challenge in court.

How It Works: Two Types of Prediction

When people talk about predictive policing, they usually mean one of two distinct methodologies:

Ultimately, the consensus among technologists and legal experts is shifting toward extreme caution: while data can help optimize where a squad car sits to deter theft, using AI to judge the future guilt of a human being crosses a dangerous ethical and constitutional line.

**********