Author: John Hancock

Reading my news feed this season, it seems like Artificial Intelligence (AI) is hitting a level of hype that may be unprecedented even for people like me who have spent a career in data analytics (and don’t even get me started on blockchain). 

A lot of the content published lately is talking about the impact that AI will have on various jobs.  When will autonomous cars replace Uber drivers? When will MacDonald’s workers be replaced with autonomous systems that already know whether we would like fries with that?  Interestingly, AI researchers think that the job “AI Researcher” is safe for many years.   

At Hubstream we spend a lot of time thinking about how to help investigators understand crime at scale and take action, and automating tasks where it makes sense.  For example, we’re adding AI-based features like the recommended triage actions to take on thousands of reports, or to automatically identify security features in pictures and videos of counterfeit products.  When will AI move beyond helping investigators to actually take over their jobs?  (Yes R. Daneel Olivaw, we’re looking at you.)

Well, let’s go beyond the hype and look at the state of the art for the use of AI in criminal investigation scenarios.  Here are some examples that we have seen.

Categorizing data

Humans are really good at categorizing data, but it can take a while.  For example, they intuitively know the difference between video footage showing a bank holdup and video footage showing a customer making an ordinary withdrawal.  Machine learning-based approaches are increasingly able to take a training set of data that is labeled by humans, and create a model that can apply human-level categorization of new data.  As investigators are dealing with larger sets of data, they can use these models to pre-categorize data and let them focus on the most relevant matches.

Processing language

Modern investigations often involve large volumes of text-based data, like chat logs or emails.  Natural Language Processing (NLP) is being applied in lots of innovative ways to zero in on the most important conversations.  For example, out of a few months of a suspect’s text messages, which texts are about meeting someone in person at a real-world location?  Or, given a set of texts from an individual who claims to be a certain age, what is the likely age range based on the language they use?

Recommending actions

Some crime types generate huge volumes of leads that investigators need to review.  This is especially true for criminal activity that uses social or ecommerce platforms eBay or Facebook, where it’s easy to find lots of targets.  Investigators need to conduct an initial review or triage or incoming leads to determine the right action to take, where actually starting an investigation is only one possible option.  Machine learning can be used to take historical data on the decisions that were made, and recommend an action for the new reports. This can help speed up review, and focus their attention on the high-risk leads that need quick action.

So, are investigator’s jobs safe or not?

In all of the above examples, we are just providing better tools to help investigators deal with the vastly increased volume of data that they have to deal with.  The basic characteristics of a great investigator are probably largely unchanged since the first day shift at the City of Glasgow police.

I’m absolutely positive that the day is coming soon when you could get killed by an autonomous vehicle.  But I’m equally sure that it will be a human that shows up to see whether the Tesla had it in for you.