Week 3 of 4, The Negatives and Challenges of AI in Law Enforcement

🚨 The Hard Truth: When AI in Policing Goes Wrong

Week 2 was about what AI can do when it works. This week is about what happens when it fails, and who pays the price.

I am not here to tell you AI has no place in law enforcement. The evidence from last week shows it does. But technology is never neutral. It reflects the data it is trained on, the priorities of those who deploy it, and the biases of the society that builds it. When those foundations are flawed, the consequences can be devastating.

Let me walk you through it.

👤 Wrongful Arrests: Real People, Real Harm

The most immediate and undeniable cost of flawed facial recognition technology is wrongful arrest. As of early 2025, at least eight people have been publicly documented as victims of arrests driven by faulty AI-powered facial recognition matches. Nearly all of them were Black. These are not hypothetical scenarios. These are real people whose lives were upended by an algorithm.

Robert Williams, Detroit, 2020.

Robert Williams was arrested at his home in front of his wife and daughters for a larceny he did not commit. The sole piece of evidence linking him to the crime was a facial recognition match between a grainy surveillance photo and his driver's license photo. The match was wrong.

Williams was detained for 30 hours. The police showed him the surveillance image and asked, "Is this you?" He looked at the blurry photo of a stranger and replied, "No, that is not me. You think all Black men look alike?"

In June 2024, Williams reached a historic settlement with the city of Detroit that included monetary compensation and forced the city to adopt the nation's strongest municipal restrictions on police use of facial recognition. The new policy prohibits arrests based solely on a facial recognition match. It took a lawsuit and national outcry to establish what should have been obvious from the start: no one should be arrested based on an algorithm alone.

Porcha Woodruff, Detroit, 2023.

In February 2023, Porcha Woodruff, eight months pregnant, was arrested at her home in front of her two children. She was accused of robbery and carjacking. She was handcuffed, taken to jail, and held for 11 hours. During her detention, she experienced contractions from the stress.

The basis for her arrest? A facial recognition match from surveillance footage. The match was wrong. Woodruff is believed to be the first woman wrongfully arrested due to this technology. Her case is part of ongoing litigation against Detroit.

Nijeer Parks, New Jersey, 2019.

Nijeer Parks was arrested for shoplifting and attempting to hit a police officer with a car. The evidence? A facial recognition system flagged him as a "possible hit" by comparing a blurry photo on a fake driver's license left at the crime scene to a police database.

Parks was 30 miles away when the crime occurred. He spent 10 days in jail. He faced charges for nearly a year before the case was dismissed. The Innocence Project, which has documented his case and others, warns that law enforcement is placing "almost blind faith in AI," using it as a primary investigative tool rather than a supplementary lead. This creates tunnel vision, where investigators ignore contradictory evidence and lock onto the machine-identified suspect.

These stories are not outliers. They are warning signs.

⚖️ Algorithmic Bias: The Numbers Do Not Lie

The wrongful arrests are not random bad luck. They are the predictable result of a well-documented technical flaw: facial recognition algorithms are significantly less accurate when identifying people of color, women, children, and the elderly.

The National Institute of Standards and Technology (NIST) conducted one of the most comprehensive evaluations of facial recognition systems in 2019, testing 189 algorithms from 99 developers. The results were unambiguous.

False Positives by Race.

False positives occur when the system incorrectly matches two different people. In law enforcement, this is catastrophic. It can lead to the arrest of an innocent person. NIST found that for many algorithms, false positive rates were 10 to 100 times higher for Black and Asian faces compared to white faces.

In one-to-many matching (the kind used to search a database for a suspect), false positives were highest for American Indians, with significantly elevated rates also seen for African American and Asian populations.

False Positives by Gender and Age.

Women were consistently misidentified at higher rates than men. Children and elderly individuals were also misidentified more frequently than middle-aged adults.

Why This Happens.

The bias is not a deliberate feature. It is a product of unrepresentative training data. Most facial recognition systems are trained primarily on datasets dominated by images of white men. When the training set does not reflect the diversity of the real world, the algorithm's performance degrades for underrepresented groups. As the ACLU puts it, this leads to "biased, suspicionless surveillance" that disproportionately targets already over-policed communities.

NIST noted that the most accurate, top-tier algorithms showed minimal demographic differences. But many law enforcement agencies are not using top-tier systems. They are using whatever vendor won the contract, often with no independent testing or validation. Until every deployed algorithm can prove equitable accuracy across all demographics, the risk of misidentification remains a fundamental, structural problem.

📹 Surveillance Overreach: The Panopticon Is Here

AI has made mass surveillance technically feasible and financially affordable. Facial recognition, automatic license plate readers (ALPRs), cell-site simulators, and social media monitoring software allow law enforcement to track, identify, and analyze vast numbers of people with minimal effort. The scale of this surveillance was unimaginable a generation ago. Today, it is routine.

Automatic License Plate Readers (ALPRs).

Many cities are blanketed with networks of cameras that capture images of every passing vehicle, recording the license plate number, location, date, and time. This data can be stored for years, allowing police to construct detailed timelines of a person's movements. Where you go to worship. Where you seek medical care. Who you visit. What political meetings you attend. All of it logged, searchable, and analyzable.

In 2020, the Electronic Frontier Foundation (EFF) won a lawsuit to gain access to ALPR data in order to study how law enforcement uses this massive repository of location information. The findings were troubling: the surveillance was indiscriminate, sweeping up data on millions of people who were not suspected of any crime.

Clearview AI: The Facial Recognition Database Built Without Consent.

Clearview AI scraped billions of images from public websites and social media platforms to build a massive facial recognition database. The company then sold access to law enforcement agencies, enabling them to identify individuals from a photo. This effectively turns every public camera and every tagged photo online into a potential surveillance device.

This was done without the knowledge or consent of the people whose images were harvested. You may have never heard of Clearview AI, but if you have ever posted a photo online or been tagged in one, there is a good chance your face is in their database.

Cell-Site Simulators (Stingrays).

These devices mimic cell phone towers, tricking all nearby phones into connecting to them. This allows police to collect identifying information and location data from hundreds or thousands of phones in a given area, all without the knowledge of the phone owners. It is a dragnet search that sweeps up data on innocent people in the hope of finding one suspect.

Civil liberties groups, including the ACLU and EFF, argue that this level of pervasive monitoring is fundamentally incompatible with a free society. It creates an environment where people cannot move, communicate, or associate without the possibility of being tracked and cataloged by the state.

🔒 Civil Liberties Under Threat: The Chilling Effect

The expansion of AI-powered surveillance directly threatens constitutional rights, particularly the First and Fourth Amendments.

The First Amendment: Freedom of Speech and Assembly.

When people know they might be watched, they change their behavior. If you believe that police are using facial recognition to identify everyone at a political protest, or social media monitoring to track activists, you may choose not to attend. You may choose not to speak. This is the "chilling effect," and it is a direct threat to the freedoms of expression and association that are foundational to democracy.

The knowledge that your face, your movements, and your associations are being logged and analyzed by an opaque AI system can deter lawful, protected political activity. This is not hypothetical. Studies and advocacy reports document that surveillance technologies are disproportionately deployed in communities of color and used to monitor activists and protest movements.

The Fourth Amendment: Protection Against Unreasonable Searches.

The Fourth Amendment requires that searches be reasonable, typically supported by a warrant based on probable cause. But many AI surveillance tools operate in a legal gray area. Using ALPRs to conduct long-term, warrantless tracking of a person's movements, or deploying Stingrays to indiscriminately collect data from every phone in an area, raises serious constitutional questions. Are these searches? Do they require warrants? The law has not caught up to the technology.

Community Control Over Police Surveillance (CCOPS).

In response to these threats, the ACLU and EFF launched the Community Control Over Police Surveillance campaign in 2016. The campaign advocates for local ordinances requiring police departments to disclose the surveillance technologies they use and to obtain approval from elected officials before acquiring or deploying them. This empowers communities to decide if and how they want to be policed.

As of 2024, at least 26 jurisdictions covering nearly 18 million people have adopted CCOPS laws. San Francisco became the first city in the nation to ban government use of facial recognition in 2019. These are meaningful victories, but they remain piecemeal. Most communities still have no say over the surveillance tools deployed in their neighborhoods.

🔍 Lack of Transparency and Accountability: The Black Box Problem

A pervasive lack of transparency shrouds the use of AI in law enforcement, making accountability nearly impossible.

Black Box Algorithms.

Many AI systems, particularly complex deep learning models, are opaque. The internal logic that leads to a particular conclusion is not easily explainable, even to the engineers who built the system. When a police department relies on such a system to identify a suspect or flag a neighborhood as a "hot spot," there is often no way to audit the decision for fairness or accuracy. This is the "black box" problem.

Proprietary Secrecy.

The problem is compounded by the fact that most AI tools are developed by private vendors who claim the inner workings of their algorithms are trade secrets. This prevents independent researchers, defense attorneys, and the public from examining the software for bias or flaws. Communities are left in the dark about how decisions affecting their safety and liberty are being made.

Departments Hiding Their Use of AI.

Law enforcement agencies have a documented history of failing to disclose their use of AI tools. An audit of the Los Angeles Police Department's predictive policing program found "significant inconsistencies" in data entry that led to biased predictions, and a lack of formal procedures for running the program. A Chicago inspector general report found the city's "heat list" program was not only inaccurate but also over-relied on arrest records, effectively punishing people for unproven acts.

This combination of technical opacity, corporate secrecy, and governmental non-disclosure makes it nearly impossible to hold anyone accountable when an AI system makes a mistake. Without transparency, there can be no oversight. Without oversight, the risks of bias, error, and abuse multiply unchecked.

🔮 Predictive Policing: Automating and Scaling Bias

Predictive policing systems use historical crime data to forecast where and when crime is likely to occur. In theory, this allows for more efficient resource deployment. In practice, critics argue these tools do not predict crime. They predict policing. And in doing so, they create a dangerous feedback loop that reinforces systemic bias.

"Garbage In, Garbage Out."

Predictive policing algorithms are trained on historical data collected by police departments. But this data is not a neutral record of criminal activity. It is a record of where police have chosen to deploy resources and whom they have chosen to arrest. Because of historical and ongoing biases, policing has disproportionately targeted minority and low-income neighborhoods. When an algorithm is trained on this biased data, it learns to associate these communities with criminality.

The Feedback Loop.

The algorithm then directs police to increase their presence in the neighborhoods it has flagged. This heightened presence leads to more arrests, often for minor, discretionary offenses that would go unnoticed in wealthier, whiter neighborhoods. These new arrests are fed back into the system as fresh data, "validating" the algorithm's original prediction. The system creates a self-fulfilling prophecy, continually justifying over-policing in marginalized communities.

Chicago's Strategic Subject List: A Case Study in Failure.

The Chicago Police Department's "heat list" assigned risk scores to individuals, predicting their likelihood of being involved in a shooting. An investigation found the program to be highly inaccurate and racially biased. Being placed on the list, often based on factors like having been arrested (regardless of conviction), made it more likely that an individual would have negative interactions with police. This reinforced their risk score, creating a cycle that was difficult to escape. The program was terminated due to its ineffectiveness and discriminatory impact.

Predictive policing does not eliminate human bias. It automates and scales it, giving a veneer of scientific objectivity to practices that perpetuate historical inequalities.

🤝 Erosion of Community Trust: The Ultimate Cost

The cumulative effect of biased algorithms, wrongful arrests, mass surveillance, and lack of transparency is a severe erosion of trust between communities and law enforcement.

When residents, particularly those in marginalized communities, feel they are being unfairly targeted, constantly watched, and treated as suspects by an opaque technological system, police legitimacy suffers. The perception that you are living in a perpetual, digital lineup creates an adversarial relationship with police, undermining efforts to build the cooperative partnerships that are essential for effective public safety.

Procedural Justice.

Procedural justice holds that fairness and transparency in processes are as important as outcomes. The secretive adoption of biased and error-prone surveillance technology is the antithesis of procedural justice. The wrongful arrests of Robert Williams, Porcha Woodruff, and Nijeer Parks are not just individual tragedies. They are public demonstrations of a system that can fail catastrophically and unjustly. Each case further convinces communities that the justice system is not designed to protect them.

Displacing Relationship-Based Policing.

Over-reliance on technology can also displace more effective strategies. Resources spent on expensive, unproven AI systems could be invested in community engagement, de-escalation training, mental health services, and other initiatives proven to build trust and reduce crime. When police are seen as an occupying force armed with an all-seeing technological arsenal rather than as members of the community, trust becomes nearly impossible to establish or maintain.

Public safety depends on public trust. By deploying AI without sufficient safeguards, transparency, and community input, law enforcement risks winning short-term technological battles while losing the long-term war for the public's confidence and cooperation.

🚧 The Regulatory Vacuum: No Rules, No Accountability

Despite the rapid proliferation of AI in policing and its profound impact on civil liberties, a comprehensive legal and regulatory framework is conspicuously absent in the United States. Law enforcement agencies are largely left to self-regulate, operating with minimal standardized oversight, inadequate training, and few enforceable rules.

No Federal Framework.

There is no overarching federal law regulating how police use AI. President Biden's "Blueprint for an AI Bill of Rights" (2022) offered guiding principles, but it is non-binding. It does not carry the force of law. In the absence of federal leadership, a patchwork of state and local laws has emerged. Most jurisdictions have no specific regulations on AI in policing.

Inadequate Training and Auditing.

The wrongful arrest cases in Detroit revealed that the police department adopted facial recognition technology without establishing adequate quality control measures or providing sufficient training on its limitations. Officers were not properly instructed on the high potential for error or the inherent biases of the system. There are no national standards for independent auditing and testing of these systems before they are deployed.

Piecemeal Progress.

Some progress is being made. California enacted SB 524 in 2025, requiring law enforcement to disclose when AI is used to assist in writing police reports. The settlement in the Robert Williams case forced Detroit to adopt new restrictions on facial recognition. Advocacy groups continue to push for moratoriums, like the Facial Recognition and Biometric Technology Moratorium Act. The NAACP and others are calling for a ban on using biased historical data in predictive algorithms and the establishment of independent public oversight bodies.

But these efforts remain scattered and incomplete. Until binding, nationwide regulations are enacted that mandate transparency, accountability, independent oversight, and equitable accuracy, the use of AI in law enforcement will continue to be a high-risk experiment with civil liberties and public trust as the stakes.

The Bottom Line

Technology is not neutral. It reflects our choices.

The risks I have outlined in this newsletter are real. They are documented. They are ongoing. Facial recognition systems are misidentifying people at alarming and racially disparate rates. Wrongful arrests are happening. Mass surveillance is expanding. Predictive policing is reinforcing the very biases it claims to eliminate. And all of this is happening with minimal transparency, oversight, or public input.

But these risks are not insurmountable. They are the product of choices: choices about what data to use, what safeguards to build, what level of accuracy to demand, and what role the community should have in decisions about how they are policed. We can make different choices.

The question is not whether AI has a role in law enforcement. The question is: on what terms? With what safeguards? Under whose oversight? And who gets to decide?

📅 Next week, in Week 4, we close out this series with a look at recent advances, the path forward, and why AI, no matter how sophisticated, cannot replace human officers. If you have stayed with me through the positives and the negatives, you will not want to miss the conclusion. Stay tuned.

💬 What stood out to you most in this week's newsletter? Which risk concerns you the most? Drop a comment below or share this with someone in the field. If you are new here, hit Subscribe so you do not miss the final installment.

#AI #LawEnforcement #CivilLiberties #Accountability #LinkedInNewsletter

Sources

1. Innocence Project: When Artificial Intelligence Gets It Wrong: https://innocenceproject.org/news/when-artificial-intelligence-gets-it-wrong/
2. Scientific American: Police Facial Recognition Technology Can't Tell Black People Apart: https://www.scientificamerican.com/article/police-facial-recognition-technology-cant-tell-black-people-apart/
3. ACLU Comment to USCCR re: Facial Recognition Technology (PDF): https://www.aclu.org/wp-content/uploads/2024/04/ACLU-Comment-to-USCCR-re-FRT-4.8.2024.pdf
4. Humanoid Liability: AI Facial Recognition Wrongful Arrests List: https://humanoidliability.com/resources/ai-facial-recognition-wrongful-arrests/
5. NIST: NIST Study Evaluates Effects of Race, Age, Sex on Face Recognition Software: https://www.nist.gov/news-events/news/2019/12/nist-study-evaluates-effects-race-age-sex-face-recognition-software
6. Scientific American: How NIST Tested Facial-Recognition Algorithms for Racial Bias: https://www.scientificamerican.com/article/how-nist-tested-facial-recognition-algorithms-for-racial-bias/
7. EFF: Victory! EFF Wins Access to License Plate Reader Data: https://www.eff.org/press/releases/victory-eff-wins-access-license-plate-reader-data-study-how-law-enforcement-uses
8. ACLU of Ohio: Police Surveillance and Civil Liberties: https://www.acluohio.org/news/police-surveillance-and-civil-liberties/
9. ACLU: Technology 101: A Primer on Law Enforcement Surveillance Tools (PDF): https://assets.aclu.org/live/uploads/publications/tc2-technology101-primer-v02.pdf
10. ACLU: The People, Not the Police, Should Decide on Surveillance Technologies: https://www.aclu.org/news/criminal-law-reform/the-people-not-the-police-should-decide-if-and-how-surveillance-technologies-are-used-in-their-communities
11. EFF: Join the Movement for Community Control Over Police Surveillance: https://www.eff.org/deeplinks/2016/09/join-movement-community-control-over-police-surveillance
12. ACLU: Guiding Principles for CCOPS Ordinances (PDF): https://www.aclu.org/sites/default/files/field_document/ccops_guiding_principles_-_final.pdf
13. Brennan Center for Justice: The Dangers of Unregulated AI in Policing: https://www.brennancenter.org/our-work/research-reports/dangers-unregulated-ai-policing
14. Knight First Amendment Institute: Transparency's AI Problem: https://knightcolumbia.org/content/transparencys-ai-problem
15. Effective Altruism Forum: Bias-in, Bias-out: Artificial Intelligence Reflects Real-World Prejudices: https://forum.effectivealtruism.org/posts/my67NQtGJJh49LKRN/bias-in-bias-out-artificial-intelligence-reflects-real
16. New York University Law Review: Criminal Justice Policy and the Fraught Future of Big Data Policing (PDF): https://nyulawreview.org/wp-content/uploads/2019/06/NYULawReview-94-3-ODonnell.pdf
17. arXiv: Towards Trustworthy AI for the Public Sector: A Case for Explainable AI in Predictive Policing: https://arxiv.org/html/2405.07715v1
18. Veritone: AI and Privacy: Balancing Technology and Compliance in Law Enforcement: https://www.veritone.com/blog/ai-and-privacy-balancing-technology-and-compliance-in-law-enforcement/
19. Gov1/Police1: IACP 2025 Quick-Take: Law Enforcement Urged to Proceed Cautiously with AI Procurement: https://www.gov1.com/products/technology/iacp-2025-quick-take-law-enforcement-urged-to-proceed-cautiously-with-ai-procurement
20. NAACP: Artificial Intelligence & Predictive Policing Issue Brief: https://naacp.org/resources/artificial-intelligence-predictive-policing-issue-brief
21. Senator Scott Wiener: Governor Signs Bill Requiring Law Enforcement to Disclose Use of AI in Police Reports: https://sd07.senate.ca.gov/news/governor-signs-bill-requiring-law-enforcement-disclose-use-ai-police-reports
22. EFF: VICTORY: California Requires Transparency for AI in Police Reports: https://www.eff.org/deeplinks/2025/10/victory-california-requires-transparency-ai-police-reports
23. ACLU: Community Control Over Police Surveillance: https://www.aclu.org/community-control-over-police-surveillance