Artificial intelligence in law enforcement should strengthen officers’ learning and judgment, not replace the thinking required to do their jobs. Agencies can gain efficiency from using AI while reducing risk by requiring users to verify sources, challenge automated conclusions, and maintain the professional skills needed to recognize errors. Leaders, supervisors, and trainers should treat AI as a tool for deeper inquiry rather than a shortcut around learning.
In prior articles, I have addressed issues related to the use of artificial intelligence in public safety. My articles addressed tangible topics such as leadership and ethics, policy and legal implications, and risk management. But the more I learn about AI and absorb all the ubiquitous information about it and its impacts, I cannot help but wonder how the proliferation of AI will impact us as human beings.
Since I was feeling philosophical, I thought it would be fitting to tap into the writings of the prominent philosopher, Plato. Plato frequently used allegories to transform abstract philosophical questions into vivid stories that required his audience to think, interpret, and search for deeper truth.
One of those, written more than 2,000 years ago, seems as if it were written for us today.
In Book VII of The Republic, Plato asks us to imagine a group of prisoners chained inside a cave since childhood. They can look only at a wall. Behind them is a fire, and people carry objects that cast shadows on the wall. The prisoners have never seen the objects, the fire, or the outside world. To them, the shadows are reality.
Eventually, one prisoner is released. Turning toward the fire is painful, and the light outside initially blinds him. Eventually, he sees his former world was only a collection of shadows. When he returns to tell the others, they do not welcome the news. They prefer the “reality” they already know.
Plato could not have imagined social media, generative artificial intelligence, or a police officer receiving an automated recommendation from a facial recognition system. But he understood the underlying human problem: It’s easy to mistake a representation of reality for reality itself. It’s also easy to become so comfortable with those representations that we resist the effort required to examine what produced them.
That brings us to an observation attributed to entrepreneur Mark Cuban. In a 2026 post and later during a public discussion about AI, Cuban said there are generally two types of people who use large language models:
- Those who use AI to learn everything.
- Those who use AI so they don’t have to learn anything.
For law enforcement, that distinction is not merely philosophical. AI is being used in law enforcement in both authorized and unauthorized manners. While generative AI and other types of AI-assisted specialized functions and platforms are proliferating rapidly, law enforcement and society still have time to determine our own intellectual destiny. We must ask the question: Do we want to use AI as our reality — the shadows — so we don’t have to learn anything? Or do we want to use AI to unchain our infinite ability to learn more about our own reality? The former is easy. The latter is more difficult but can and will be infinitely more rewarding.
“AI should help us develop more knowledgeable officers, not merely more efficient ones.”
A Wall Filled With New Shadows
In a May 2026 Washington Post opinion piece (paywall), Kathleen Parker described a digital world populated by AI-generated material, bots, manufactured social media accounts, and advertising disguised as authentic opinion. Her concern was the growing difficulty of knowing what is real, who created what we see, and why it was placed in front of us.
The people carrying objects in front of the fire may now be algorithms selecting what keeps us engaged, marketers manufacturing popularity, bots creating the illusion of consensus, or people clipping a few seconds from a longer encounter. Increasingly, the shadow may be produced by AI.
Law enforcement has dealt with incomplete representations for a long time. A witness sees only one part of an event. A camera records only what appears within its frame. A radio transmission compresses a complicated encounter into a few hurried words. A social media clip may begin after the conduct that prompted an officer’s response and end before the outcome is known.
AI adds another layer with all the functions that AI-based or assisted platforms can provide. Why write a report yourself when an AI tool can do it for you? Why spend time researching and creating lesson plans when GenAI can do it for you?
The problem begins when a summary or an unverified interpretation becomes a substitute for the source. The shadow is no longer recognized as a shadow — it is interpreted as the object itself.
I can attest that AI has tremendous value. I am astounded by the doors of knowledge this technology has opened to me. In Mark Cuban’s terms, I am using it to learn everything I can, and it has reinvigorated me. But at the same time, I realize how easy it would be for officers to become overreliant on AI summaries and interpretations. Generative AI can easily lure us into thinking we understand everything we need to about a subject. We must push back against this and cultivate officers with the curiosity and skill to use AI to its full potential.
Using AI as the means to “learn everything“ opens the door to questions you never knew to ask. Using AI so you don’t have to learn anything will leave those questions forever unasked and unanswered.
Is Convenience Creating Our “Chains”?
A 2025 review published in Brain Sciences examined research involving excessive screen time, digital addiction, doomscrolling, fragmented attention, and the consumption of low-quality online content, particularly among adolescents and young adults.
The literature associates excessive consumption of fragmented and emotionally stimulating digital content with cognitive overload, mental fatigue, disrupted attention, memory difficulties, and psychological distress. The review also discusses how digital platforms reward continued engagement and accustom users to rapid, emotionally satisfying content.
The study did not examine police officers. Still, the underlying concern deserves attention in a profession where people already work under stress, fatigue, interruption, and information overload.
Officers’ social media feeds may include reels depicting attacks on police and controversial uses of force — designed and curated to provoke anger or fear. Repetition may shape expectations and create the impression that rare but dramatic events are common simply because the algorithm keeps displaying them. Availability bias occurs when people draw conclusions and opinions on the information most available to them. This is a real concern since frequency and repetition often win out over accurate information that is not as prominent. The impact of availability bias is multiplied by the effect of social media algorithms, which pay attention to a user’s preferred interests and continue serving up similar content in an effort to get more engagement.
Watching short videos does not make an officer incapable of sound judgment. But habits of attention matter. If we become accustomed to receiving complicated subjects in short, emotionally charged packages, it becomes harder to work through ambiguity. Reading the complete case, reviewing the entire recording, or examining contrary information requires effort.
AI can reduce that effort, which is both its appeal and its risk.
“A Human Checked It” Is Not Enough
Most discussions about responsible AI use eventually arrive at the same reassuring phrase: Keep a human in the loop. It sounds responsible. By itself, though, it means very little.
A human can be present without exercising independent judgment. If that person lacks the knowledge to evaluate an AI recommendation, routinely accepts it, or no longer remembers how to perform the underlying task, review becomes ceremonial. Reading an AI answer and agreeing with it is not verification.
A 2026 study published in Scientific Reports helps illustrate the problem. Researchers asked 295 participants to classify 80 faces as real or AI-generated. Each participant received guidance supposedly produced either by a human or by AI. The guidance was correct only half of the time.
The results were not a simple story of people unquestioningly trusting machines. Participants generally followed correct guidance more often than incorrect guidance, and the AI group was not significantly less accurate than the group receiving human guidance. That is encouraging.
Other findings should give us pause. Participants who reported always using the guidance were less accurate than those who used it selectively or not at all. Among those receiving AI guidance, more positive attitudes toward AI were associated with a reduced ability to distinguish real faces from synthetic ones.
This was a controlled, low-stakes task, not police decision-making. Even so, it shows AI support depends partly on the person receiving it and how that person uses it.
That is why a policy requiring human review is hopelessly inadequate unless the agency defines what the human must know and do. Meaningful oversight requires a person who understands the task, knows the system’s limitations, can independently evaluate the important facts, has enough time to do so, and possesses the authority to reject the output.
A person cannot meaningfully check an answer they are not qualified to evaluate. If no one in the process can detect the error, it is the “human in the loop” who will later be asked why the machine was trusted. And “the AI did it” will not be a valid defense.
When the Tool Begins to Replace the User
This issue is particularly important when AI performs work that also develops, or should develop, professional competence.
Report writing is an obvious example. It requires an officer to organize observations, identify gaps, distinguish facts from assumptions, and explain the actions taken. AI may help improve a report, but if it performs too much of that work, the officer can receive a polished narrative without engaging in the thinking that should precede it. Cautionary examples can be found almost daily now.
In July 2026, Alcorn State University history professor Jason Gibson said 32 of 35 students across two summer classes submitted midterm responses containing nonsensical references to Madagascar. Gibson had planted a hidden instruction in the assignment prompt that would be captured if the entire prompt were copied into an AI system. One of the resulting grades was later changed on appeal because of a display issue, and the university has not publicly audited the incident. Even with those qualifications, the episode offers a striking warning. The fact that the students used AI was not the most troubling part. Instead, it was that so many apparently submitted its words without reading, questioning, or understanding them.
The same concern applies to supervisors who review an AI-generated body-camera summary instead of the recording, trainers who accept an AI-generated legal lesson without reading the case and investigators who treat an algorithmic facial identification as a conclusion rather than an investigative lead.
Skill loss can happen quietly and quickly. The system works often enough that checking it begins to feel unnecessary. The user moves from active decision-maker to passive monitor. When the technology finally makes a serious mistake, the person assigned to catch it may no longer have the knowledge, attention, or confidence needed to intervene.
This is not a new problem. Aviation and other high-risk fields have studied automation bias and the loss of situational awareness for decades. AI increases the concern because a large language model predicts a plausible response from patterns. It does not experience the event, understand the consequences of its answer, or bear responsibility for what happens next. Yet it can produce language that sounds reasoned, confident, and complete. A wrong answer can look every bit as professional as a correct one.
Using AI to Leave the Cave
The answer is not to ban AI or frighten officers away from it. That approach would sacrifice AI’s considerable benefits and probably fail anyway. Cuban’s first category of user offers the better path.
An officer can use AI to elaborate on an unfamiliar concept and then follow the explanation to the source. A supervisor can ask it to identify weaknesses or generate alternative explanations. A trainer can use it to develop scenarios and anticipate questions. An investigator can use it to organize information while preserving the distinction between a lead and evidence.
Used this way, AI does not eliminate thought. It provokes more of it.
The difference can be found in the questions the user asks. The passive user asks, “What is the answer?” The active learner asks, “Why might this be the answer? What evidence supports it? What information is missing? What would make it wrong? Where can I verify it?”
Before relying on AI output in any consequential law enforcement task, the user should be able to answer several questions:
- Do I understand the issue well enough to recognize a questionable answer?
- Can I identify and examine the primary sources or underlying evidence?
- Have I considered facts or explanations that contradict the output?
- Can I explain, in my own words, why I accepted or rejected the recommendation?
- Am I prepared to take responsibility for the decision without blaming the tool?
These questions impose useful friction. Efficiency matters, but friction is not always negative — even if it adds time to the task. In high-consequence work, the effort required to verify, explain, and defend a decision is part of the safety system.
We Should Want Officers Who Keep Learning
Plato’s allegory contains a discouraging ending. The prisoner who escapes returns to help the others, but they resist him. They are comfortable with the shadows and suspicious of anyone who challenges their perceptions.
Law enforcement leaders should work toward a different ending.
Timely legal analysis on law enforcement-related cases: SUBSCRIBE NOW!
We should want officers who are willing to turn around, question what they are being shown, and follow the first prisoner out of the cave. That requires more than an annual warning about AI hallucinations or a policy directing employees to verify AI output. Agencies must create a culture in which curiosity is expected, primary sources are accessible, and people are given the time and training necessary to understand the tools they use.
Supervisors should ask officers how they reached their conclusions, not merely verify whether the final product looks acceptable. Trainers should use AI to create learning opportunities, not replace the learning process. Policies should distinguish between AI assistance and the surrender of professional judgment. Agencies must preserve foundational skills because officers cannot supervise processes they no longer understand.
AI did not create Plato’s cave. Human beings have always mistaken appearances for reality and preferred convenient explanations to difficult truths. AI does, however, allow us to create more shadows, make them more convincing, and deliver them directly to every officer’s hand.
It can also help us ask better questions, find information faster, and understand subjects that once seemed beyond our reach. The same tool can deepen the cave or help us find the way out.
The goal should not be officers who can produce the fastest answer. It should be officers who understand why an answer may be right, know what could make it wrong, and remain willing to do the work necessary to tell the difference. AI should help us develop more knowledgeable
- Webinars
