Teaching OpenAI to assess risk, with CopBot!

submited by
Style Pass
2023-03-19 16:00:03

In case you’ve been living under a rock and missed all the recent excitement around ChatGPT, it works really, really well now… Like, “oh god that’s actual wizardry” well. The fact it would probably breeze through a Voight-Kampff test should probably worry me a bit, but more importantly, what cool stuff can we do with it?

Given using the tool in development is actually quite affordable, I thought I’d build a few prototypes for public safety use cases, and see how it performs… and after a few hours of work, “CopBot” is alive!

Reading, evaluating and prioritising risk is a core policing skill: from investigating crimes with piles of witness statements in dingy offices, to responding to life threatening incidents incidents at 2am on a rainy street, the key thread is you’ve got loads of risk, and you need to decide what to do first, taking into account decades worth of policing legislation and policy, and making sure your decision is justifiable when it inevitably goes wrong.

So can an AI learn to “speak police” and make vaguely convicing risk assessments? I scraped all the policing guidance I could find, fed it to an OpenAI powered model, and asked it to predict risk for missing people in an explainable way…and it kind of works! You can try out the prototype here.

Leave a Comment