AI and machine learning will improve policing


Artificial intelligence and machine learning are hot topic terms that are sometimes used interchangeably. Artificial intelligence (AI) is the theory and development of computer systems able to perform tasks that normally require human intelligence. Machine learning is the field of computer science and application of AI that gives computers the ability to learn without being explicitly programmed through access to data. 

Google’s DeepMind lab project ‘AlphaZero’ recently learned chess in four hours. It’s being labeled a ‘superhuman performance’ because not only did AlphaZero learn chess, it absorbed all data on the entire history of chess and then figured out how to beat anyone or anything in the game in one-sixth of a day.

This type of technological advancement not only changes the way chess will be played, but it creates endless possibilities for technology that can be applied to other aspects of everyday human life.


In policing there is a huge amount of ‘big data’ floating around. There are data sets relating to daily policing tasks like UCR/NIBRS, 911/CAD, RMS, sex offender, accidents, traffic, homelessness, recidivism, drug use, cold cases, homicides, etc. And there is  data from a multitude of agencies - city, county, regional, state, federal. And data from different parts of a city like housing, finance, demographics, poverty levels, medical care, etc.

Humans, and police officers, are overwhelmed with the amount of information available, making it hard to analyze data and respond to front of the line police work.

By unlocking data and making it consumable by machines, artificial intelligence and machine learning will develop new tools and help to fight crime.


Artificial intelligence and machine learning can help cities and police agencies do more with less. AI can be better at data-analysis than humans and can drive actionable intelligence from the analysis. Machine learning will learn from previous computations to produce reliable, repeatable decisions and come up with rational solutions.

Take this scenario for example:
An enormous amount of cross jurisdictional data (like, crime, 911 calls, topography, demographics, etc) on Baltimore and Milwaukee is fed to a computer.

The computer will take the data and learn about each city. Shapes around a specific part of Baltimore and a specific part of Milwaukee with be drawn by the computer on a map, both shapes including similar topography. From these shapes, the computer will analyze data from both cities and determine things like:
  • When and how many officers to send.
  • What types of crimes will occur and probability of when they could occur.
  • What types of police interactions work best in which neighborhoods.
  • What other resources need to be attributed to each neighborhood (homelessness, housing, mental health, etc)
  • If and what environmental factors are affecting the crime rate.
  • Solve cold cases.
With Moore’s law (the overall processing power for computers will doubling every two years) and the cost of computing decreasing, the opportunities to create better communities with computer intelligence are endless and quite possibly beyond what can be imagined by the human brain. Better policing. Safer cities. More resources being deployed efficiently and to the people who need them most. We just have to make it possible for computers to lead the way, in a responsible manner.


Focus needs to be on data. Big data. Share data and share more of it making it more easily accessible not only within a police agency and city, but across jurisdictions, across state lines, and with the public.

The more data, the ‘deeper’ the learning. NIBRS/UCR is good, calls for service ‘raw’ data is better. From our experience, sharing ‘raw’ data, like 911 calls, is very uncomfortable for police agencies because of the fear that since it’s not ‘official’ data they submit to the state and FBI, it will make them look bad.

It’s true, 911 data has its fallibilities. A call for gunshots may turn out to be fireworks, or high crime areas see fewer calls for service because residents are so used to seeing petty thefts and car break ins they don’t think to report them to the police.

Despite all its problems, 911 data is so much more robust and useful. When we talk to professionals who want to analyze data, most of the time they request 911 data and do not even look at UCR/NIBRS.

The great thing about AI and machine learning is the capability to consume an array of large data sets. It can distinguish between 911, RMS, and NIBRS/UCR data, then draw conclusions without failures of the human mind.

The biggest hurdle is getting set up. Once the data is captured, and technology is in place, computers can take it the rest of the way.

Finally, agencies need to make the process as transparent as possible. There always lies an existing possibility that AI and machine learning can be wrong. For example, the bail and sentencing algorithm was found to be biased against blacks. To prevent this from happening, the kinds of data being consumed, the methodology, and algorithms need to be shared publicly. This creates oversight and accountability to check for bias and incorrect conclusions to be drawn. And, it allows the police and public to trust ‘outsourcing’ work to a computer. It is possible that one day AI will audit AI.


Currently, machine learning and AI are being used by police agencies across the world. Below are just a few examples.
INTELLIGENCE-LED POLICING and CRIME ANALYSIS - The European Union’s VALCRI scans millions data sets from a wide range of mixed-format sources -police records, interviews, pictures, videos - to identify connections that it thinks are relevant. It then creatively analyzes the data, displays its findings with easy-to-digest visualizations, and comes up with possible explanations of crimes.
A concept used by agencies around the world which uses mathematical, predictive, and analytical techniques to identify potential criminal activity. LAPD has utilized PredPol to lower their crime rate. IBM’s Crime Information Warehouse (CIW) marries the concepts of crime analytics and predictive policing.
IDENTIFYING SERIAL KILLERS - Thomas Hargrove created an algorithm identifying crime patterns that point to if there is a serial killer present in a city.
COMPUTER AIDED DISPATCH - Police have already been using the help of computers for years. Computer Aided Dispatch - also known as CAD, calls for service, or 911 calls - provides displays and tools so that the emergency dispatcher has an opportunity to handle calls for service as efficiently as possible.
FINDING BETTER SOLUTIONS TO COMMUNITY PROBLEMS - Recently opioid crisis data was analyzed to find out what systems were and were not helping fight the US opioid epidemic.
BAIL/SENTENCING - In the UK the Harm Assessment Risk Tool (HART) classifies individuals based on a low, medium, or high risk of committing a future offense. The US uses proprietary algorithms to predict future behavior by defendants and incarcerated person in order to set bail, determine sentences, and even contribute to determinations about guilt or innocence. These algorithms are proprietary meaning there is no transparency associated with methodology leaving them controversial in nature.

For SpotCrime, using better technology to surface and analyze data seems like the correct path to policing. We hope these trends help encourage police departments to be more transparent.

We don’t expect a magical panacea. Right now quality police are, and will still be, needed on the streets doing the real work.

And, we need to be aware of the the pitfalls (like bias conclusions and any data privacy concerns) of pursuing this technology. We hope when police departments are buying new technology they are doing it with a view of how the systems will serve them and surface the data in the easiest ways possible. Often we see poor procurement where the new database system hog ties the department’s data and prevents analysis.

Computers, and computer systems, should be purchased to serve the departments and ultimately the public - not the vendor selling the systems - and newer systems should be less complex that the ones they are replacing. 


Popular posts from this blog

SpotCrime Weekly Reads: AI, police conduct, transparency

SpotCrime Weekly Reads: Transparency, gun violence, crime data

SpotCrime Weekly Reads: Violent crime, AI tech, transparency