Steve Goldsmith (4:17)

That's a lot of stakeholders in that last answer. Thinking back when you were doing these local government interviews, one or two stand out in your mind as somebody making a comment and you thinking, well, I really have to set up this justice tracker to watch out for things like this. Or did somebody give you an example of a really good way they were using AI that motivated you? One or two maybe aha moments.

Dr. Andrea Headley (4:46)

So I would say two different aha moments, one on an example and then another one on more of like, the lack of knowledge. Right. And so starting with the latter one, actually, one of the things we realized very quickly is that both city government employees as well as communities just wanted to know what was happening. Like, cities wanted to be able to say, like, oh, well, what's this other city doing in this department? Or what have you, or comparable things that we might either want to do or want to avoid. And that finding that information was very hard for them. And then community members also just wanted to know. So I would say, like, one of the aha moments was really on this gap of like, there is not awareness that could then spur to innovation or safeguarding and things like that. And then I would say the other aha moment in terms of actually an example was around talking to someone who was actually formerly incarcerated and who was recently released and was talking about how they would have loved an opportunity to have AI help them navigate the reentry process, and then asked whether there were any tools like that. And we were like, oh, wow. Well, you know, we should know the answer to that. Right? Like, it's actually a very simple question of whether there's tools or whether there could be opportunities to create tools. Right. And so I would say that then also sparked this, like, well, let's go find out. Right? Are there tools? Are there companies who are offering tools? Are there cities trying to innovate and pilot certain tools like that to help people and then became the tracker?

Steve Goldsmith (6:22)

That's terrific. I like those stories. I don't think we told you before you agreed to come on the podcast, but. So I've been active in criminal justice for like a century or so. I was a DA then. I was a mayor. Obviously, as mayor, you're responsible more or less for police department. As deputy mayor of New York, I've been thinking for some number of years that if we could do a better job of using what we're now calling AI, of using AI to examine all of the written documents that law enforcement officers put out or all of the dispatches they use, or all of the language they use, that we could dramatically improve training in order to mitigate bias. There's been a lot of attention over the last eight years on algorithmic unfairness and inequity. But I guess my question is, how do you think about using these tools to identify and correct and mitigate those attitudes as a positive influence of AI?

Dr. Andrea Headley (7:33)

That's a great question, and I feel like it's the million dollar question that I'd be rich if I knew the answer to. But I can tell you what I'm thinking and kind of where we are right now. And so one of the things we have been thinking, particularly to your point, because there has been a lot of talk on just like, how potential uses of AI and enhanced technologies will only exacerbate biases and inequities embedded in the system. We have been trying to identify cases that are either trying to reduce inequities. Right. Or actively promote more positive behaviors. And those cases, I will say, are far and few between right now, but it doesn't mean that the opportunities for them are not there. And so one example that has been used in San Francisco actually has been a AI tool that redacts race from certain algorithms for sentencing decisions and in court procedures. And the idea behind that was because of all the racial disparities that we. We have seen with differences in terms of sentencing outcomes and risk scores and. And all of the court outcomes that follow. And so there was this race redaction tool that was developed actually by some researchers and a partnership at the time, at least they were at Stanford to do that. And it was deployed all across, actually the state of California in the end, again, with the goal of mitigating racial bias and reducing biases in the downstream effects. And so I think that is one example of trying to use the tool in a way that actually tries to intervene on points that we know There are already disparities that exist in terms of law enforcement in particular. There are also examples that are being put forth of like, okay, can we analyze language in real time? Can we identify potentially, you know, biased language from camera footage, for instance, and then send, you know, reminders or notes or identify people that might need additional training, etc. To your point. Right. That then hopefully either reduces that behavior altogether or at least a little bit. Right. And moving the needle forward. And so I think those are definitely some areas and opportunities. There's some talks that are very high level, I would say that some researchers that I know are having now with vendors trying to see how they can leverage AI identification, particularly for identifying positive examples of officers, for instance, who did a really good job de. Escalating, did a really good job in engaging with a community that may have been a, you know, different race or different socioeconomic status or what have you in resolving the conflict in a way that avoided harm. Right. That wasn't necessarily a harsh use of force, for instance, and then uplifting those as training examples of like this is what could be right as opposed to just having the viral videos that go off around when things go wrong.

Steve Goldsmith (10:23)

Yeah. So take all the body cam language, run it through a generative AI filter and say, let's give a recognition to the officer who did the best job of defusing a situation. Or what can we learn about the techniques of the 10 officers who had the most productivity in the fewest confrontations or a whole set of questions. Right. That might inform how we trained and promoted and disciplined and congratulated.

Dr. Andrea Headley (10:50)

Yeah, that's exactly right. And I'm also curious. I mean, one of the things that I've been thinking a little bit about that, you know, maybe there's a company that wants to try to create this is like taking some of what we know, even from behavioral science and nudges around how we change people's behaviors or at least remind them of things. And could there be, for instance, an AI assistant for a police officer that's actually a proactive reminder while they're in the car that's saying like, hey, Remember, you shouldn't be doing 1, 2 and 3. Right. What you should be doing is this type of language. You should be de. Escalating. You shouldn't be using, you know, harsh language or whatever. The case might be more as a like kind of prompt for positive engagement. A prompt in real time right before someone is going to interact. Right. As a reminder. And I think on one, right. Some people might say, well, that doesn't solve the innate systemic or structural problems in the first place. But I think opportunities for changed behavior that reduce harm in real time are really important still. And figuring out where technology can leverage some of the things that we know about how individuals make decisions under constraints and under stressful times, and when there are time pressures or, you know, a lot of information. Right. Like how we can use AI to potentially nudge people to promote the positive behavior is important.

Steve Goldsmith (12:12)

So what sorts of things are you tracking with your tracker?

Dr. Andrea Headley (12:16)

Yeah, we're tracking all types of things. So what we essentially ended up doing was we identified the top 100 cities. We looked across courts, corrections and law enforcement to see what we can find from publicly available information. But in a systematic search process where we have specific keywords for every city that we're searching for to try to see how they're using AI in a way that is obviously has to be documented. So it's not the total world of it, but it's a snapshot of what's publicly available. And so after we identify the cities and the three domains that we're focusing on, we then categorized all of the different use cases that we found into a taxonomy of six different categories, ranging from technology that is more surveilling, technology that is detecting, such as like a weapon or an object, technology that is at the front end of the system, which we are kind of defining as interacting or engaging with the public. So being that interface between the system and the public, technology that's at the back end of the system. And so technology that's streamlining processes, but not necessarily, you know, identifying an object in real time in the community or talking to the public. And then technologies that are more kind of forensic analysis, analyzing DNA and things like that. And so we had these different categories and there's lots of different use cases that we've seen that fall into them, from things like facial recognition technology, obviously is one of the most popular or gun detection technology to more rare cases that are jury chatbots or Miami Dade county deployed autonomous police vehicle. And so those ones that are a little bit more rare, but also we've

Steve Goldsmith (14:05)

Been tracking, and you referenced earlier, using AI to help a parolee find resources in the community. Or maybe you would be using AI to help the parole officer find the best way to support the parolee. How are you looking at or what great uses have you seen of AI on kind of that side of the justice ladder?

Dr. Andrea Headley (14:31)

So I would say, unfortunately, we actually haven't seen it yet. In talking to both people who are engaged in the system and employees, it's been a very clear opportunity identified of like wanting that. But we haven't actually found a city thus far that has actually deployed avoid anything in that realm. Right now, I would say in terms of this like interface between the public and the system, the things that we have found more often than not have been the chat bots that are helping either a juror navigate the system or a victim chatbot. So there was a victim chatbot that was deployed particularly for burglaries, I want to say, where people could then ask questions and things like that. But we haven't necessarily seen one yet that is focusing on, for instance, reentry on the kind of latter end of the system.

Steve Goldsmith (15:23)

Our listeners are more on kind of the civil side of government. CIOs and data officers, analysts, people work for the mayor, but they obviously have an interest in justice generally. How should they, the CIOs think about preparing data that would lend itself to the insights you've been talking about? Because there's some of it's justice data, but some of it's housing data or where the jobs are available. So what's the role of the CIO in making the data as accessible for justice type AI interventions?

Dr. Andrea Headley (16:01)

I would say it's central, honestly. And so in doing some of the interviews, we actually talked to people outside of justice, right, in these kind of more centralized offices in particular, because of the interconnectedness of it, and learned a ton. But I would say in particular, right, having systems within government that talk to each other and that can be connected, particularly for ways that are more preventative, I feel like is a real opportunity. And so there is one city that I remember viewing that essentially had a AI resource that essentially identified people who were most in need of housing assistance and housing services and wraparound services, and then provided those services to them as a means of avoiding. Avoiding the justice system, actually. And so the article that was written up about it was in conversation with criminal justice. But instead of identifying people who are potentially at risk of committing crime, they said, how can we identify these people that have some of the same risk factors? But then provide them with a different service. Right. And so to me, the role of the CIO in that is making sure that the data inputs are there across different organizations. Right. Within a local government to be able to facilitate the types of actual preventative wraparound public service delivery that we hope that we could be providing to the community writ large. Especially when we know that oftentimes, particularly when people are calling 911. Right. They're not always calling about criminal justice issues. They're calling about issues that are about lots of other things and they don't necessarily have somewhere to call or to go to. Right. 311 is here and that depending on the city, there are some resources that could be called. But I think there's a clear role for some of the data infrastructure to be set up in a way that allow organizations within government to be talking and leaning on each other for that.

Steve Goldsmith (17:54)

Yes. We had a project we wrote about a couple years ago. I think it was in San Diego where they were looking at what does the police department use as its performance measures and how do those performance measures relate to the way the public evaluates safety. Right. And they weren't necessarily related because the data reflected the public sense of safety, as was really kind of how safe the streets felt and looked, cleanliness and homelessness and other activities, and the police department was looking at crimes. So there was just kind of a mismatch. So using AI to understand what's important to the community and how to correct it in terms of responsiveness, I guess I don't know if you'd call that a justice tracker, but we've been thinking about in terms of responsiveness.

Dr. Andrea Headley (18:44)

I mean, I love that. I think that that's exactly right. Especially because at a broader level. Right. A lot of the things that community cares about when we think about safety writ large. Right. If we're thinking about what it means to have a community that you live in that is safe and you're thriving and there's well being that's flourishing and things like that. I think the indicators that you are bringing up are exactly the right ones. And there's a lot of literature actually also within kind of the crime prevention space that talks about things like crime prevention through environmental design, the built environment and how street lights and certain greenery and certain provisions actually make for safer communities. Right. And so there are ways in which we can think about how can AI maybe identify what in communities maybe needs further improvement, for instance, and leverage. Kind of this responsivity to community concerns is really important. One other thing, Steve, that I wanted to bring up though, with regards to the kind of role of the cio, I would also say one of the things we've seen in looking at cities is that there are certain cities that have really been leading in having these AI inventories, and they're not that many yet, but San Jose in particular kind of led this initiative to have a publicly available AI inventory across their government. And that was spearheaded by the CIO's office. Whereas like New York and San Francisco had a legislative mandate right around like trying to have these publicly available inventories. In San Jose, it was the, the CIO office that was like, we should be putting this out there for our public as a means of transparency so that everybody knows what's happening. We can uplift best practice, we can disseminate knowledge, we can enhance public service by making sure people are informed. Right. And so I would also say that the role of CIO offices, for instance, can be potentially spearheading some of that innovation in terms of making AI uses transparent as well.

Steve Goldsmith (20:42)

This is so exciting. I've been working on criminal justice data systems information for four decades. I'm that old, literally four decades. And the opportunities now to constructively intervene to help people. But whether you're, as you said before, training people in the system to act better, predicting outcomes, identifying resources, there are so many things that could be accomplished. So for our listeners here, what should they do over the next 24 months that would utilize these tools to unlock the most capacity to make sure that justice is done, that people are helped? What would you nudge our listeners to do?

Dr. Andrea Headley (21:28)

So I would nudge the listeners out of one, at a very basic level to talk to the people who are going to either be most involved in implementing AI, right, or using it in real time in different parts of government and, or most impacted in the public. And so one of the things that we have realized and kind of launching the tracker, talking to people about it, but also doing our interviews with people, is that the opportunities that exist already exist within the people who are closest to the issues, right? And so they know the pain points in the system, they know what might make their job easier, they know what might make their transitioning through, you know, or navigating a system easier. They might not have the solution, right? And so that's where we bring in vendors and, and maybe collaborate and co design, have co designed design sessions with vendors, government, community, right. Where we're trying to really maximize the potentials that are here. But also what I would love to see if it's not already being done, is taking a strategic management approach where we're doing like pre mortems. So in the business world, they do all these pre mortems, which are essentially trying to predict the implementation that is about to fail and then working backwards and saying, well, why is this going to fail? And then designing preventative measures in hopes that you can prevent the failure from happening in the first place. But it's done at the very beginning of something, as opposed to waiting to failure, as opposed to waiting till the worst case scenario happens. And now we're trying to adjust for it on the back end. And so I would really love if we can take that more proactive approach and strategic approach that is potentially identifying any risks that are there any worst case scenarios, failures that could happen when we're deciding about any specific use case, and then working backwards to make those informed decisions of whether it should be implemented, how it should be implemented, what safeguards might need to be made, what types of evaluations might need to be done that's ongoing, and how they then continuously learn. And then the last thing I would say is really thinking about the human in the loop, particularly because I know, Steve, one of the things you said, even if the listeners are not necessarily justice focused, one of the things that I find personally so important about understanding the criminal justice system is that it impacts so many other policy areas and it has implications for trust in government writ large and people's notions of citizenship and how they engage with government. And because of it being a high risk context. Right. I think if we get it right in a high risk context, then we can get it right in a lot of other areas. But if we get it wrong, then the amounts of legislation, the amount of restrictions, the amount of things that will really have to be put in place that's going to affect government writ large, not just criminal justice, are really something to, I think you be mindful of. And we have the opportunity, I would say, to be mindful now. But the one example that came to mind actually that is to me a story about implementation failure is the case that happened where a gun detection system in a school in Baltimore misidentified a Dorito bag. It was in a pocket of a young man and thought it was a gun. And essentially police were sent, police were armed, they pointed the gun at the young boy. They. He ended up being arrested. They found out later it was a Dorito bag, but all his friends were there, right? And when the story unfolds and like there's different articles of talking to the family, talking to the youth, talking to the administrators. Pretty much the administrators and the vendors say the system did what it was supposed to do. It identified a potential object that could have been a gun. And there's debate around, like, misidentification, a Dorito bag and things like that. But putting that aside for a second. Second. But really what they then say is after that, it's up to the humans in the system to figure out what the best course of action is and to make sure the course of action corrects potentially for a misidentification. Right. Like, those are human things that we need to be thinking about. Human implementation decisions that can hopefully take serious the potential concern while also safeguarding a potential misidentified issue and being mindful of the different factors that are at play and hopefully that prevents, you know, the youth ends up saying in the end he thought he was going to die. Like, that's, you know, basically for. For lack of better words, how can we avoid an outcome like that while also taking safety very seriously? Right.

Steve Goldsmith (26:04)

Well, it's a good example. I suspect with any challenging example, you could have a situation where AI makes the human in the loop worse or AI makes the human in the loop better. Right. Depends on the human and the AI. And so what we're hoping to do today is just raise up attention to the potential and the challenges. And it's a great example of the good you're doing with your Evidence for Justice Lab and your Justice AI Tracker. So, Andrea Headley, thank you so much. We'll encourage people to look up your website at Georgetown and see how they can take advantage of the Justice AI Tracker. Thanks for your time.

Dr. Andrea Headley (26:45)

Thank you. If you liked this podcast, please Visit us at datasmartcities.org and find us on itunes, Spotify, or wherever you get your podcasts. This podcast was hosted by Stephen Goldsmith and produced by me, Betsy Gardner. Thanks for listening.