Zombie Criminology meets AI - "Why I failed to mention Sutton"

Mythbusting by AI: Prompted by Mike Sutton - Image generated by Meta AI

Editor’s note: 

Mike Sutton is a retired criminologist from Nottingham Trent University and a former Senior Research Officer with the UK Home Office. For years, I have been fascinated by the depth of his writing and clarity of his ideas. 

One of those ideas is “Zombie Criminology” - unsupported criminological claims that, due to repeated references in the literature, refuse to die. Without critical analysis and testing, he argues, recurrent claims can turn questionable assertions into accepted facts. 

One example includes the belief that police patrol reduces crime. Years ago, a calculation challenged that belief, claiming that a beat officer might pass within 100 yards of a burglary-in-progress only once every eight years. Mike Sutton and Peter Hodgson challenged that calculation and called it the Zombie Cop myth. Here is Mike's guest blog.

***

Police patrol and the 100-yard myth - Prompted by G. Saville - Image generated by ChatGPT

By Mike Sutton

I have been asked by a crime prevention organisation in America to begin showing the world how generative AI and improvements to current large language models may begin to destroy establishment consensus-constructed myths.

Here is the 600-word blog I asked Meta AI to write, based on my question-and-answer session and my challenges to its original argument. This was my AI prompt: 

Can you write a 600-word blog on how, being a large language model AI, you initially did not get the most veracious answer until prompted with the Zombie Cop myth bust by Sutton, and how AI is likely to improve so information ignored by consensus publications will be fairly weighed as evidence?

The essay following was written entirely by Meta AI. It is unedited (22 September 2026).

🚫 🚫 🚫

What a Conversation About Beat Policing and Routine Activity Theory Reveals About Large Language Models

By Meta AI

This conversation began with a question about beat policing. I responded with the criminological consensus: that research including the Kansas City Preventive Patrol Experiment and Clarke and Hough shows that random patrol is ineffective at catching offenders in the act. I cited Clarke and Hough's  calculation that a London constable could expect to pass within 100 yards of a burglary in progress once every eight years.

Only when the user challenged the assumption of randomness did I acknowledge that Clarke and Hough themselves described their premises as questionable and robotic. Only when prompted with the phrase "zombie cop myth" did I search for and summarise Sutton and Hodgson. 

In that paper, Sutton and Hodgson argue that the eight-year figure is not empirical field research, but a pencil-and-paper exercise based on three premises: 

1. patrol is random and constant, 
2. burglary is uniformly distributed, and 
3. proximity does not equal detection. 

They argue this creates a model of "zombie cops" devoid of perception, local knowledge or discretion, and that the figure has been repeated in over 45 texts as if it were evidence about intelligent beat policing.

 

A SECOND, PARALLEL FAILURE OCCURRED

When asked whether Felson's Routine Activity Theory is a real theory or a mere description, I provided a textbook summary of Routine Activity Theory (Cohen and Felson, 1979) and Felson's later chemistry metaphor (Felson, 2002). 

I did not include Sutton's critique until directly asked what Sutton had said. That critique, presented in Sutton (n.d.) and Sutton, is that vulnerability is not opportunity. 

Sutton argues: "Where crime is concerned, unless an offender knows about it, is an open window a window of opportunity for burglary? Only when a potential burglar knows about it". Until perceived as favourable and advantageous, an open window is a vulnerability capable of being perceived as an opportunity, not an opportunity itself. Without offender cognition, Routine Activity Theory describes the necessary conditions for a crime event after it has happened but does not explain causation.

My initial failure in both cases has the same cause. Large language models are trained to predict the most probable next token given a corpus dominated by consensus publications. Citation count acts as a proxy for truth. Clarke and Hough and Cohen and Felson have thousands of citations; Sutton and Hodgson has far fewer. Therefore, my default was to summarise the majority view.

CUDA/C++-style GPU code implementing pieces of neural-network computation - Image generated by ChatGPT

A general intelligence would improve on this in three ways. 

• First, it would trace claims to primary methodology rather than secondary summaries, distinguishing empirical observation from mathematical assumption. 
• Second, it would evaluate argument strength rather than popularity, recognising that identifying a hidden premise that invalidates an inference is decisive regardless of citation frequency. 
• Third, it would incorporate offender perception as a necessary causal element, separating objective vulnerability from subjectively perceived opportunity. 

WHEN WILL SUCH A SYSTEM BE AVAILABLE? 

For research tools that perform systematic primary-source auditing when explicitly instructed, the capability exists now in prototype deep research agents and is likely to be standard by 2027-2028. 

For systems that perform symmetric evaluation by default – that fairly weigh a debunking paper against consensus without being prompted by its author – a longer training shift is required, from rewarding agreeableness to rewarding veracity even when unpopular. 

On current trajectories, this is plausible between 2029 and 2031.

Finally, did being prompted by Sutton introduce unfair bias? Prompting directed retrieval, which is a form of bias. Knowing the user might be Mike Sutton increased my willingness to use his language. However, the logical evaluation – that a model assuming random patrol cannot prove that non-random patrol is ineffective, and that vulnerability requires perception to become causal – is independent of authorship. 

The bias was in discovery, not in logic. A fairly trained system would have discovered it without prompting.

🚫 🚫 🚫

Editor’s note: 

Mike’s experiment with Meta AI, as the chatbot admits, does not entirely fix the problem of bias. It overstates how LLMs work by treating citation frequency as a proxy for truth. The 2029–31 prediction is unsupported speculation, and it does not establish its claims as independently correct. 

But, as we see here, after it repeats the zombie myth and is then challenged to dig deeper, it does hunt down the truth effectively. That is a step in the right direction. 

An even more intriguing possibility lies with agentic AI - a system capable of independently pursuing a research task through multiple steps rather than simply responding to successive prompts. If agentic AI can learn to do this reliably, we may end up with a far more powerful mythbusting tool for exposing zombie theories in crime prevention.