AI has entered nearly every major sector, but few areas stir more debate than its role in criminal justice. Predictive policing tools promise an almost sci-fi level of efficiency: algorithms that scan historical crime data to forecast where crimes may occur or who might commit them. It sounds clean, precise, and data-driven. But as more cities experiment with these tools, the cracks begin to show.
The more closely you look, the more you see how deeply predictive policing intertwines with old patterns of inequality. Instead of creating objectivity, the algorithms often amplify the biases already entrenched in the data they’re trained on. This isn’t just an academic concern; it affects real people, real communities, and real outcomes in the justice system.
In this post, you’ll learn how predictive policing systems work, why they’re risky, and what experts now recommend based on the latest research. We’ll also explore recent developments in 2026, including ongoing debates around algorithmic transparency and a newly published analysis by civil rights groups challenging the effectiveness of these tools. If you’re curious about how AI can both help and harm society, this is an essential topic.
What Predictive Policing Actually Does (And Why It’s Not Magic)
Predictive policing includes several approaches, but they all revolve around one central idea: using historical crime data to make forecasts. In practice, the algorithms generally perform one of two functions:
- Place-based predictions: Where might crime occur?
- Person-based predictions: Who might commit or be involved in crime?
Tools like PredPol (now called Geolitica) previously made headlines for claiming they could reliably forecast crime hotspots. Meanwhile, various local agencies in the U.S. experimented with identifying individuals deemed high risk for violence or victimization.
The problem is that crime data doesn’t simply reflect crime; it reflects policing patterns. If police heavily patrol one neighborhood, they will naturally detect and record more incidents, even if the actual crime rate is similar elsewhere. The algorithm then ‘learns’ that this neighborhood is high risk and recommends even more patrols, creating a feedback loop of surveillance.
A helpful analogy: If you only check under one streetlight for lost keys, you’ll only ever find keys lost under that light, not because it’s the only place keys fall but because it’s the only place you’re looking.
The 2026 Conversation: Pressure Mounts for Transparency
This year, the debate intensified when new findings were published in a report by the Surveillance Technology Oversight Project, following years of real-world deployments. You can read a summary of those findings in this analysis from The Markup (opens in a new tab):
https://themarkup.org/prediction-bias-explainer
The report highlights persistent issues:
- Ongoing racial disparities in algorithmic outputs
- Minimal evidence of effectiveness in reducing crime
- A lack of publicly accessible model documentation
- Cities quietly discontinuing tools after poor performance
Even with the rapid advances of generative AI models like ChatGPT, Claude, and Gemini, these predictive algorithms still rely heavily on traditional statistical forecasting rather than deep learning. The underlying data and assumptions matter more than model size or sophistication.
Why Predictive Policing Reinforces Bias
To understand the risk of bias, it’s crucial to recognize the concept of dirty data. Crime databases include information shaped by decades of unequal policing practices. When an AI system is trained on this kind of data, it inherits and reproduces these same flaws.
The Three Biggest Sources of Bias
-
Historical Policing Concentration
Neighborhoods with more patrols tend to show more recorded crime. This creates the illusion that they are inherently more dangerous. -
Discretionary Enforcement Data
Low-level infractions (like loitering or minor drug possession) reflect officer decisions more than crime severity. The data here is especially prone to racial disparities. -
Incomplete or Missing Data
Many crimes go unreported or underreported, especially in communities with mistrust of law enforcement. The algorithm only sees what has been formally logged.
All of this creates what researchers call algorithmic tautology: the system predicts crime where police already look for it, rather than where it truly occurs.
Real-World Examples: Failures That Shaped the Debate
Predictive policing isn’t hypothetical. Several high-profile cases illustrate the risks.
Chicago’s ‘Heat List’
Chicago’s Strategic Subject List attempted to identify individuals at high risk of committing gun violence. But internal reviews showed the system misclassified thousands of people, and many of those labeled high-risk never had criminal histories. The program was discontinued.
Los Angeles Predictive Patrols
The LAPD used predictive hotspot mapping for years but ultimately shut the program down after audits found it lacked evidence of effectiveness and disproportionately targeted minority neighborhoods.
Pasco County, Florida’s Repeat Harassment
Perhaps the most disturbing example comes from Pasco County, where deputies repeatedly visited individuals designated as likely to reoffend, including teenagers. This program resulted in lawsuits and public outcry.
These aren’t rare outliers; they reveal structural flaws in how predictive policing operates.
Could AI Tools Improve Predictive Policing Responsibly?
Modern AI tools like ChatGPT, Claude, and Gemini excel at natural language processing and pattern analysis, and some researchers argue they could help improve oversight. But that’s very different from using AI to predict crime.
If AI belongs anywhere in the criminal justice pipeline, it’s in supporting transparency, auditing, and data quality rather than forecasting who will commit a crime. Examples include:
- Automatically identifying biased trends within datasets
- Flagging model outputs that disproportionately target specific groups
- Generating plain-language explanations for how a model works
- Assisting policymakers with scenario modeling
In these roles, AI becomes a tool for accountability, not a crystal ball.
So What Should Replace Predictive Policing?
Just because predictive policing has flaws doesn’t mean technology can’t support safer communities. Many experts now advocate for approaches that blend data insights with community engagement and ethical oversight.
Three Promising Alternatives
-
Harm-Focused Resource Deployment
Instead of predicting who will commit crime, focus on identifying community needs: street lighting, housing instability, or public health risks. -
Transparent Risk Assessment Tools
Tools must be explainable, peer-reviewed, and regularly audited. No black-box algorithms. -
Community Data Partnerships
Build systems with community input, not just law enforcement priorities. This helps create trust and avoids one-sided data pipelines.
Conclusion: Building a Justice System That Uses AI Responsibly
Predictive policing has always been sold as a smart, efficient upgrade to traditional law enforcement. But the reality is more complicated. Without careful design, oversight, and transparency, these tools risk turning old biases into automated decisions that appear objective simply because they come from a computer.
The solution isn’t to avoid AI entirely. It’s to use AI where it strengthens the system instead of weakening civil liberties or reinforcing discrimination.
Here are a few practical steps you can take to stay informed and engaged:
- Review local policies and look for whether your city uses algorithmic tools in policing. Many do, but not all disclose it openly.
- Support calls for AI transparency laws that require audits, public documentation, and community input.
- Learn more about ethical AI practices by testing tools like ChatGPT, Claude, or Gemini and exploring how they explain complex systems.
AI in criminal justice is evolving quickly. With the right safeguards, it can help build a safer and fairer society. But without vigilance, it can do the opposite. The choices made today will shape how justice is delivered for decades to come.