<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Catherina Xu on Medium]]></title>
        <description><![CDATA[Stories by Catherina Xu on Medium]]></description>
        <link>https://medium.com/@catherinaxu?source=rss-d8eb32f35c71------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*xXGxx5I6Ml0qUK4jnqYDTQ.jpeg</url>
            <title>Stories by Catherina Xu on Medium</title>
            <link>https://medium.com/@catherinaxu?source=rss-d8eb32f35c71------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 03:19:34 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@catherinaxu/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Part IV: Breaking Down Data Analytics and AI in Criminal Justice]]></title>
            <link>https://medium.com/recidiviz/part-iv-breaking-down-data-analytics-and-ai-in-criminal-justice-942169a4f11b?source=rss-d8eb32f35c71------2</link>
            <guid isPermaLink="false">https://medium.com/p/942169a4f11b</guid>
            <category><![CDATA[criminal-justice]]></category>
            <category><![CDATA[data-analytics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Catherina Xu]]></dc:creator>
            <pubDate>Wed, 09 Mar 2022 07:37:08 GMT</pubDate>
            <atom:updated>2022-03-09T18:42:56.771Z</atom:updated>
            <content:encoded><![CDATA[<p>Written by Catherina Xu, Julia Dressel, and C. Andrew Warren</p><p><em>This is the final part in a four-part series Recidiviz has written to document the use of technology in the criminal justice ecosystem and share specific recommendations for practitioners in tech procurement, integration, and deployment (see </em><a href="https://medium.com/recidiviz/part-i-a-mapping-of-tech-in-the-criminal-justice-system-bce3c5f37fd7"><em>Part I</em></a><em>, </em><a href="https://medium.com/recidiviz/part-ii-why-good-data-and-feedback-loops-in-criminal-justice-are-hard-to-come-by-b07e7e2ee3d7"><em>Part II</em></a>, <a href="https://medium.com/recidiviz/part-iii-tips-for-procuring-and-adopting-criminal-justice-technology-db6e2f481941"><em>Part III</em></a><em>). As always, we welcome thoughts and feedback.</em></p><p>In recent years, artificial intelligence (AI) and data analytics tools (DAT) have exploded in popularity. And everyone wants in: vendors are taking advantage of this hype by “AI washing” and applying the AI label (sometimes indiscriminately) to their offerings; organizations are eager to see how DATs can help with their work.</p><p>The criminal justice tech space is no exception. As we explored in a previous <a href="https://docs.google.com/document/d/1vmZTL7lp7aGrLWMxKVfqc1_7BB9kI97jkC2_Y4xaecY/edit">post</a>, technology is increasingly being used across all parts of the justice system. Many new tools are AI-enabled, including algorithmic risk assessments, facial recognition, and gunshot detection.</p><p>But these tools aren’t perfect. Even the most sophisticated companies in the consumer AI space have grappled with algorithms that disproportionately impact <a href="https://www.reuters.com/article/us-amazon-com-jobs-automation-insight/amazon-scraps-secret-ai-recruiting-tool-that-showed-bias-against-women-idUSKCN1MK08G">women</a> or <a href="https://blog.twitter.com/engineering/en_us/topics/insights/2021/sharing-learnings-about-our-image-cropping-algorithm">people of color</a>. Most companies <a href="https://ai.facebook.com/blog/facebooks-five-pillars-of-responsible-ai/">now have “Responsible AI” teams</a> dedicated to studying and mitigating these biases. In the criminal justice space, where the stakes can be much higher, DATs may cause significant harm by perpetuating historical bias and violating privacy. One thing is clear — not all DATs are good, and criminal justice practitioners need to very carefully evaluate the types of technologies they are looking to adopt.</p><p>This posts aims to help practitioners make better procurement decisions by:</p><ul><li>Explaining the differences between common types of DATs, including those that use AI</li><li>Applying these definitions to the criminal justice context</li><li>Exploring the different types of data involved in different technologies and corresponding pitfalls</li><li>Providing practical and actionable advice to practitioners that can be implemented during the technology procurement and deployment process</li></ul><p>Why are we writing about this? First, our team has a lot of experience in data analysis, statistics, machine learning, and AI. We want to share what we’ve learned. Second, we’d like to show where our tools fit into this landscape, and the precautions we’ve found effective so far in working at the intersection of tech and criminal justice.</p><h3>Predictive analytics, deep learning, machine learning, reinforcement learning, and more….</h3><p>Before we dive into criminal justice specifics, let’s parse some of the common technical terms we encounter in this space. A few caveats to keep in mind throughout the rest of this section: First, these definitions are constantly evolving as technology changes; what constituted “machine learning” 30 years ago is very different from what it is today. This guide captures the present-day understanding of these terms. Second, data analytics is an expansive field — these definitions aren’t meant to be comprehensive, but we hope they will at least help to untangle the complex web that is data-related terminology, methodology, and use cases.</p><p>To better illustrate these concepts, imagine you are the owner of an orchard that produces Red Delicious apples, and your goal this year (and every year) is to harvest as many ripe apples as possible.</p><p><strong>Artificial Intelligence (AI) — Robot, please go out and pick the apples!</strong></p><p>AI is a broad term used to describe a computer system able to perform tasks that ordinarily require human intelligence. Machine learning (ML) is often, but not always, used to power these AI systems. A robot that picks apples on its own is an <em>AI</em> — picking fruit efficiently surely requires a certain level of intelligence. The robot uses ML, which we’ll explain next, to do specific tasks like recognize apples on a tree or grasp each individual fruit.</p><p><strong>Machine learning (ML) — Is this an apple?</strong></p><p>ML is a type of <em>method</em> that trains computer systems to identify patterns from data and make decisions without human help. We can use ML to train our apple-picking robot to recognize apples from other fruits, tree branches, and owls. The data we use to teach the robot might look something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*GLG06ytKUpD7iCck" /></figure><p>With enough of these examples, our robot will eventually learn the distinguishing features of apples (like roundness and redness) and be able to classify new images with relatively high accuracy.</p><p><em>Deep learning</em> and <em>reinforcement learning</em> are specific types of ML techniques.</p><p><strong>Predictive Analytics — How many apples can we expect to harvest next month?</strong></p><p>Predictive analytics is a specific category of data analysis that uses historical data and trends to make predictions about the future. This analysis can involve ML methodologies. For example, we can train our system with harvest records from previous years, which include information like weather conditions, pest patterns, and total apple output, to help it learn what conditions lead to what quantities of apple output.</p><p><strong>Descriptive Analytics — How many apples did our orchard produce this year? How does this compare to last year?</strong></p><p>Descriptive analytics is the “examination of data or content to answer the question ‘What happened?’”, for an event that’s already occurred. It may also be used to answer deeper questions like “Why did our orchard only produce half the number of apples it usually does?” Did a new type of pest emerge? Did the irrigation system break down? Or were our apples stolen during the night? Although data is critical to this analysis, ML is not usually used.</p><h3>What do these technologies look like in criminal justice?</h3><p>Data tools in criminal justice roughly fall into three different buckets:</p><ul><li><strong>Automated Identification Tools, </strong>which use a combination of AI and ML methods</li><li><strong>Predictive Forecasting Tools</strong>, which use predictive analytics</li><li><strong>Tools that Help us Understand the System, </strong>which use descriptive analytics</li></ul><p>Each type of tool is useful in its own way, but also carries risks specific to its inner technical workings. In this section, we’ll explore both risks and potential mitigation strategies associated with each of the three different tool types.</p><p>1.<strong> Automated Identification Tools</strong></p><p>Automated identification tools use ML to help with complex classification problems. These tools are used to automate tasks of varying difficulty, from call transcription to fingerprint matching, and help humans extract insights from large amounts of data.</p><p><strong>How can I spot these tools?</strong></p><p>If you’re using a piece of technology to answer these kinds of questions, then it’s an automated identification tool:</p><ul><li>Does this fingerprint match a print in our database?</li><li>What words were just said in this courtroom?</li><li>Was that sound outside a gunshot or fireworks?</li></ul><p>Traditionally, these tasks are what a team of analysts are paid to do — now, computers can look at millions of images or records in just a few minutes. This is possible because the availability of data, as well as the ability to process that data, has significantly increased.</p><p><strong>What are some common examples in the CJ space?</strong></p><p>The most common types of data processed by these tools are <strong>images</strong> (e.g. face detection in bodycam redaction, facial recognition in prison surveillance, license plate recognition), <strong>voice</strong> (e.g. prison call monitoring, court transcription, audio analytics for emergency calls), <strong>sounds</strong> (e.g. gunshot detection), and <strong>text </strong>(e.g. entity matching between different types of records).</p><p><strong>What are the risks associated with these tools?</strong></p><p>Recall from our orchard example that in order to train a robot to recognize apples, we had to show it example images of apples and non-apples. Similarly, to train a face detection tool using ML, the system is shown many images of faces and non-faces, so that the tool is able to learn the distinguishing features of a human face when it encounters one in bodycam footage; for gunshot detection, the distinguishing features between a real gunshot and fireworks, and so on.</p><p>Now, imagine that you’re training your robot to recognize different types of apples. What happens if you only have one or two images of green apples on hand? Not surprisingly, your robot probably won’t do too well when asked to classify a green apple. Unfortunately, this problem is pervasive in data that is used to train criminal justice tools. Popular open source image datasets that companies use to train facial recognition models <a href="https://venturebeat.com/2021/03/16/imagenet-creators-find-blurring-faces-for-privacy-has-a-minimal-impact-on-accuracy/">have been shown to overrepresent</a> males between 18–40 and underrepresent people with dark skin. Various other commercial classifiers <a href="http://gendershades.org/overview.html">tend to perform worse</a> on women and those with darker skin tones. And sometimes, even without obvious data issues, the underlying problem ML is trying to solve is just really hard — a new analysis in Chicago showed that police officers who respond to ShotSpotter, a gunshot detection tool, subsequently come back without finding a problem <a href="https://chicago.suntimes.com/2021/5/4/22417660/shotspotter-analysis-macarthur-justice-center-chicago-police-chicago-gun-violence-editorial">86% of the time</a>.</p><p><strong>How do we mitigate these risks?</strong></p><p><em>Ask a lot of questions about the data. </em>Are there known issues with the types of data that were used to train the system (e.g. image datasets that don’t adequately represent faces with dark skin)? What, if any, steps have been taken to ensure that the data is high quality, fresh, and representative across protected characteristics? Data that is not appropriately representative can lead to systems that don’t work as intended or actively perpetuate harm.</p><p><em>Ask whether the tool actually works.</em> What is its proven accuracy? What is its false positive rate, and false negative? Just because a piece of technology sounds impressive doesn’t mean that it actually works. Understanding the limitations of a tool are crucial in mitigating the risks of its use.</p><p>2. <strong>Predictive Forecasting Tools</strong></p><p>Predictive forecasting tools apply ML and other statistical methods to data from the past in order to make predictions about what will happen in the future. These tools are used to guide decision making in various stages of the criminal justice system.</p><p><strong>How can I spot these tools?</strong></p><p>If you’re using a piece of technology to answer these kinds of questions, then it’s a predictive forecasting tool:</p><ul><li>What is the expected population of the county jail next month?</li><li>What is the likelihood that this person will show up to their trial if they are released from jail?</li><li>How would the decriminalization of marijuana impact the size of the prison population over the next 10 years?</li></ul><p><strong>What are some common examples in the CJ space?</strong></p><p>Predictive forecasting tools are used to make predictions about an individual, a community, or an entire system. <strong>Individual-level predictive tools</strong> include risk assessment instruments like <a href="https://en.wikipedia.org/wiki/COMPAS_(software)">COMPAS</a>, <a href="https://doc.mo.gov/justice-reinvestment-initiative/oras">ORAS</a>, and the <a href="https://advancingpretrial.org/psa/factors/">PSA</a>. These instruments rely on trends from historical criminal records to predict the future behavior of an individual, e.g. whether they are likely to commit another crime or not show up to court. <strong>Forecasting at the</strong> <strong>community-level </strong>includes predictive policing tools like <a href="https://www.predpol.com/">PredPol</a>, which analyzes historical arrest data to determine where criminal activity is most likely to occur in a city throughout the day. <strong>System-level predictive analytics </strong>include models of how a system has changed over time in order to predict how it will change in the future. For example, prison facility population models rely on historical prison admission and release trends to forecast the growth of the prison population in the coming years.</p><p><strong>What are the risks associated with these tools?</strong></p><p>These tools are usually trained on what we call “<strong>tabular data</strong>,” a fancy way of saying information that is stored in a table. Each row represents a data point (e.g,. a formerly incarcerated person) and the columns represent different facts about that individual (e.g. age, race, sex, nature of crime committed, whether they recidivated, etc). In our orchard example, this would be equivalent to the harvest records that not only document the number of apples harvested in a given year, but also various other features like soil acidity and rainfall. The system will then use that data to learn which features contribute to a successful harvest.</p><p>But criminal justice records are far from perfect — <a href="https://www.nyulawreview.org/wp-content/uploads/2019/04/NYULawReview-94-Richardson-Schultz-Crawford.pdf">research has shown</a> that these records are produced during “documented periods of flawed, racially biased, and sometimes unlawful practices and policies.” For instance, the ACLU of Illinois found that a significant number of stop and frisks conducted by the Chicago Police Department were unlawful, and Black residents were <a href="https://www.aclu-il.org/sites/default/files/wp-content/uploads/2015/03/ACLU_StopandFrisk_6.pdf">disproportionately subjected to these unlawful stops</a>. Records from these encounters constitute data that can play a role in (wrongly) teaching a computer system that crime is more likely to happen in Black neighborhoods, or that a Black person is more likely to commit a crime, just because they are Black. For example, <a href="https://gizmodo.com/crime-prediction-software-promised-to-be-free-of-biases-1848138977">researchers have found</a> that PredPol over-recommends policing Black and Latino neighborhoods and “mostly avoid[s] Whiter neighborhoods” due to the crime report data it is trained on.</p><p>The use of individual and community-level criminal prediction tools can have severe and disparate consequences. Using biased or flawed data to inform new decisions about where to send police forces, or which individuals should be barred from parole, can further incriminate vulnerable communities — which in turn creates more records that are used to inform decisions down the line. These self-reinforcing systems replicate and deepen the existing disparities that we see in the current criminal justice system.</p><p>There are also risks in forecasting the impact of changes on an entire population. A policy modeling tool that doesn’t look at projected outcomes across race or other important factors can lead to the promotion of policies that at first seem promising, but end up increasing adverse outcomes for marginalized groups. Any issues in data quality or freshness can lead to decisions made on inaccurate information and loss of ability to track the impact of new policies or practices.</p><p><strong>How do we mitigate these risks?</strong></p><p><em>Interrogate the underlying data used to make predictions. </em>Is this a domain in which the data is likely to reflect historical biases? (The answer is likely yes). Do any of the inputs correlate highly to protected class characteristics? Is the data used in the system likely to be a good indicator of what the tool is trying to do? (For example, recorded arrests could be a flawed input for predictive use cases — an arrest may be deemed unconstitutional, or might not actually lead to a conviction).</p><p><em>Ask for context-specific validations of the tool in your jurisdiction. </em>Predictive tools are often not built or tested in the jurisdiction for which an agency is looking to procure technology, despite vendors’ claims that their tools are “nationally validated.” What the tool learns from data in one place may not be useful when applied to another (risk factors are likely to be a bit different in rural New Mexico compared to New York City). Agencies should ask vendors to do validations for a specific location and use case, and revalidate whenever there are significant changes in local law, composition of the community, or other factors. Important questions to ask: What are key differences between our jurisdiction and where the tool was tested or previously deployed? How might those differences impact performance? How does the tool actually perform on data from our jurisdiction? Have we talked to other jurisdictions who use the tool, to understand common failure scenarios or edge cases they’ve found?</p><p><em>Recognize that “fixing the data” is not always a solution</em><strong>. </strong>Data is not always “fixable.” For example, even if crime data was 100% accurate and certified to be obtained from legal practices that resulted in conviction, that data still inherently reflects historical biases in policing. Similarly, excluding race information from the data will not automatically make a tool less biased; often, other information fields like zip code and income may serve as a proxy for race. There isn’t really a way to “remove” these biases, and in many cases it may make sense to stop using these technologies altogether (as <a href="https://www.latimes.com/california/story/2020-04-21/lapd-ends-predictive-policing-program">some agencies have already done</a>).</p><p>3. <strong>Tools that Describe the System</strong></p><p>There are a handful of other data tools that fall into this third bucket:<strong> </strong>ones that describe the system. These are also known as <strong>descriptive statistics tools</strong>, and don’t rely on any ML techniques.</p><p><strong>How can I spot these tools?</strong></p><p>If you’re using a piece of technology to answer these kinds of questions, then it’s a descriptive statistic tool:</p><ul><li>How many people were admitted to prison last year because they had their parole revoked?</li><li>How many people currently on probation in this county are overdue for being discharged?</li><li>Which counties have had the highest failure to appear rates over the last 10 years?</li></ul><p><strong>What are some common examples in the CJ space?</strong></p><p>Descriptive analytics tools in criminal justice include dashboards that describe the historical trends of a given system, e.g. how a state’s county jail population has changed over the last 10 years, or provide a snapshot of what’s happening in the system today, e.g. the number of people currently on probation in each county in the state. These tools don’t always present aggregated data; for instance, case management tools tell a parole officer who is on their caseload, what the contact requirements are for each person, and when each person is scheduled to be discharged.</p><p><strong>What are the risks associated with these tools?</strong></p><p>Descriptive statistics can only be as accurate as their underlying data. As with any of these tools, inaccurate data collection directly impacts the usability of the tool. If a correctional officer tends to misclassify new admissions from the court as revocations from probation, then the descriptive statistics that report these numbers will be inaccurate.</p><p>Even with accurate data, there is a risk that the data presented by these tools will be interpreted to harmful ends. Poorly designed data visualizations can be easy to misinterpret or taken out of context. There is a risk that someone will make a decision based on data they are seeing in these kinds of tools, even though their understanding of the data visualization is mistaken.</p><p><strong>How do we mitigate these risks?</strong></p><p><em>Build disaggregated views into dashboards.</em><strong> </strong>It is possible for important disparities in outcomes among racial and other minorities to be hidden from view in data visualizations showing aggregate trends. For instance, a chart may show that the rate of probation success has been rising steadily over the last 5 years even though the completion rate has actually been <em>declining</em> for Native Americans during that same time period. Tracking important outcomes for marginalized groups is important to understanding if a given problem is getting worse for communities that are the most vulnerable.</p><h3>An aside: How does Recidiviz fit into the picture?</h3><p>Recidiviz focuses on building <strong>system-level predictive forecasting</strong> and tools that <strong>describe the system</strong>.</p><p>Our system-level forecasting tools include <a href="https://www.recidiviz.org/policy">policy impact modeling</a>. We build probabilistic data models that simulate people moving through the criminal justice system to forecast the projected impact of passing a given policy. We project five-year impacts on cost, population changes, and life years saved by a person being out of the system. We have also built system-level forecasting capabilities in our <a href="https://www.recidiviz.org/leadership-tools">leadership tools</a>, which help agencies anticipate future changes in their system’s population sizes.</p><p>The descriptive statistics tools that we build include the public-facing dashboards showing the number of people <a href="https://dashboard.cor.pa.gov/us-pa/narratives/parole/2">currently on parole in Pennsylvania</a> and that highlight how the <a href="https://dashboard.docr.nd.gov/us-nd/narratives/prison/3">prison population in North Dakota</a> plummeted when COVID hit. The data visualizations in our internal leadership tools help state agencies better understand trends in their systems. We also build tools that surface data at the individual-level; our tool for parole and probation officers identifies which individuals on an officer’s caseload are overdue for discharge based on the person’s sentence and the state’s discharge policies, which helps ensure that no one stays on supervision past their sentence’s expiration date.</p><h3>Parting thoughts about AI and data analytics in criminal justice</h3><p>In a previous post, we published a <a href="https://docs.google.com/document/d/1dFy9J9k8sUi8dOcplN0mIvz42et7wF73wtVpr3UMXZY/edit#">Criminal Justice Tech Playbook</a> that aims to help practitioners procure and deploy tech tools more effectively. Tools that use AI and data analytics bring their own set of benefits, caveats, and risks. In this section, we will expand upon the playbook with considerations specifically tailored to the use of these tools.</p><p>1. <strong>Know that AI, and more complex AI, is not always better</strong></p><p>Terms like “deep neural networks’’ are popping up to describe new products with high price tags. But just because a technology sounds fancy doesn’t mean that it is actually better than something simple. In fact, a <a href="https://advances-sciencemag-org.stanford.idm.oclc.org/content/4/1/eaao5580">research study showed</a> that “despite [the COMPAS automated risk assessment’s] collection of 137 features, the same accuracy can be achieved with a simple linear classifier with only two features.” The latter tool is not only far easier to use, but is also far more explainable.</p><p>2.<strong> Never use high-risk technologies on their own</strong></p><p>Tools that can lead to significant negative consequences for an individual (e.g. a facial matching tool used to make an arrest, or risk assessments that determine eligibility for parole) should never be fully trusted for its decisions. Instead, they should be treated as decision aids for people making the final decision, similar to high-risk tools in the medical field (e.g. cancer screeners).</p><p>3. <strong>Don’t underestimate the downsides of scale</strong></p><p>Scale is what enables technology to be effective — by ingesting more data, extracting more trends, and making more predictions in a minute than thousands of humans can do in a week, data analysis tools can dramatically speed up operations and derive insights that otherwise would be missed entirely.</p><p>But scale comes at a cost, and two of the most significant costs are <em>privacy </em>and <em>transparency</em>. Just because automated license plate readers and facial recognition robots can be placed everywhere doesn’t mean that they <em>should </em>be, or <em>can</em> be constitutionally. Though machines are now able to process and understand data in ways that help them perform tasks better, these ways are sometimes so technical and complex that they can no longer be explained — sometimes even by the people who developed them. These are all costs that should be factored into a decision to use any data analytics tool.</p><p>Lastly, we’ll emphasize that these considerations stand <em>in addition to </em>those described in the Playbook — analyzing the system’s performance across different groups, tracking impact and feedback loops, and weighing cost/benefit to decide whether to use the tool at all are still (if not even more) important to do with data analytics tools.</p><p>As the AI landscape changes and criminal justice technologies involve, we hope to continuously update this piece. If you have thoughts or feedback, please let us know!</p><p>For further reading:</p><ul><li><a href="http://aimyths.org/">aimyths.org</a> — a resource that aims to debunk the most common and harmful misconceptions about AI</li><li>AI Now Institute’s <a href="https://medium.com/@AINowInstitute/algorithmic-impact-assessments-toward-accountable-automation-in-public-agencies-bd9856e6fdde">Algorithmic Impact Assessments</a> — a framework that describes how public agencies can increase accountability</li><li><a href="https://cjtec.org/files/5f5f93ea20ded">AI in the Criminal Justice System</a> — a set of briefs authored by the The National Institute of Justice that looks at AI technology used in different parts of the justice system</li></ul><p><em>We would like to thank </em><a href="https://www.linkedin.com/in/clementine-jacoby-50578483/"><em>Clementine Jacoby</em></a><em> and </em><a href="https://www.linkedin.com/in/nikhil-bhargava/"><em>Nikhil Bhargava</em></a><em> for their helpful comments and contributions.</em></p><p><em>Recidiviz is proud to be a </em><a href="https://glginsights.com/glg-social-impact/"><em>GLG Social Impact Fellow</em></a><em>. We would like to thank GLG for, among other things, connecting us to some of the experts who provided valuable insights showcased in this post.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=942169a4f11b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/recidiviz/part-iv-breaking-down-data-analytics-and-ai-in-criminal-justice-942169a4f11b">Part IV: Breaking Down Data Analytics and AI in Criminal Justice</a> was originally published in <a href="https://medium.com/recidiviz">Recidiviz</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Part III: Tips for procuring and adopting criminal justice technology]]></title>
            <link>https://medium.com/recidiviz/part-iii-tips-for-procuring-and-adopting-criminal-justice-technology-db6e2f481941?source=rss-d8eb32f35c71------2</link>
            <guid isPermaLink="false">https://medium.com/p/db6e2f481941</guid>
            <category><![CDATA[technology-procurement]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[criminal-justice]]></category>
            <dc:creator><![CDATA[Catherina Xu]]></dc:creator>
            <pubDate>Tue, 01 Feb 2022 07:27:24 GMT</pubDate>
            <atom:updated>2022-02-03T19:37:23.198Z</atom:updated>
            <content:encoded><![CDATA[<p>Written by Catherina Xu and C. Andrew Warren</p><p><em>This is the third part in a four-part series Recidiviz has written to document the use of technology in the criminal justice ecosystem and share specific recommendations for practitioners in tech procurement, integration, and deployment (see </em><a href="https://medium.com/recidiviz/part-i-a-mapping-of-tech-in-the-criminal-justice-system-bce3c5f37fd7"><em>Part I</em></a><em>, </em><a href="https://medium.com/recidiviz/part-ii-why-good-data-and-feedback-loops-in-criminal-justice-are-hard-to-come-by-b07e7e2ee3d7"><em>Part II</em></a><em>). As always, we welcome thoughts and feedback.</em></p><p>The criminal justice system is <a href="https://medium.com/recidiviz/part-i-a-mapping-of-tech-in-the-criminal-justice-system-bce3c5f37fd7">awash with new tech tools</a> and <a href="https://medium.com/recidiviz/part-ii-why-good-data-and-feedback-loops-in-criminal-justice-are-hard-to-come-by-b07e7e2ee3d7">inherently a difficult place to establish data-driven feedback loops</a>. Practitioners looking to use technology must wade through a dizzying array of options to determine what to use, when to upgrade, and how to implement changes and track impact. Drawing on both what we’ve learned from building Recidiviz and successful practices from the private sector and government, we created this guide to help criminal justice practitioners procure and deploy technology effectively.</p><h3><strong>1. Start with the problem and the people, not the solution</strong></h3><p>Before looking at technology solutions, we should clearly articulate a) the need we’re trying to address, b) the people who would use the solution, and c) how it could fit into their workflow.</p><p>For example, imagine our goal is to increase the efficiency of re-entry case managers to reduce deferred parole hearings. It might be tempting to jump straight to looking at checklist or project management software, or designing something in-house. But this skips to the solution, before we really understand the problem. If the main issue holding back re-entry readiness is that vocational and rehabilitation is often required for parole, but only available in certain facilities, there are a lot of options to explore to fix the issue, and many don’t require new tools at all — e.g., expanding programming availability across facilities might help more than a to-do list tool would have. Another common issue is understanding the problem, but not how our approach would fit into existing workflows: we might rush to procure a great to-do list app, only to find that case managers only have time to check their to-do’s during lunch time, when away from their computer, and the app we procured isn’t built for smartphones. Knowing the right place to intersect with their workflow might have helped us find the right solution.</p><p>One area to be especially wary is with shiny new technology, like <a href="https://www.ibm.com/topics/what-is-blockchain">blockchain</a> or <a href="https://www.ibm.com/cloud/learn/deep-learning">deep learning</a>, where exciting buzzwords can obscure that the tool doesn’t actually address the problem at hand. Thinking about the problem, and the user, rather than starting the solution can help weed out costly missteps and lead to more adoption within the agency. Without adoption, shininess doesn’t matter much. As an added bonus, this process will also lead to a better RFP that treats user-experience as a first-order priority.</p><p><strong>Checklist</strong></p><ul><li>Answer five key questions up-front:<br>◦ <strong>Who is the key user of this tool?</strong><br>◦ <strong>What is the need, as they would articulate it? </strong>(It’s really easy to build tools that we think people need, but if it’s not what they think they need, you’re not going to see much uptake)<br>◦ <strong>What’s the core problem they’re trying to solve? </strong>(If the user says they need a to-do list tool — why? What’s the real stumbling block for the organization that they’re trying to fix with it? Try to separate what they want from what they’re trying to solve.)<br>◦ <strong>In what types of environments will the tool be used?</strong> (Consider the types of platforms, e.g. mobile or desktop, that might be appropriate)<br>◦ <strong>What would wild success look like with this tool </strong>(or, to put it in RFP language: what is the “desired impact”)? Is there more than one way to get there?</li><li>Interview users to prioritize pain points and figure out where a new tool would be most effective in their workflow</li><li>Talk to peer agencies to see if they’ve had similar needs and how well different solutions have worked for them. (Trust numbers above anecdotes — did program matching go up, and by how much? Do they have usage metrics to show how often staff use the tool?)</li><li>Articulate product requirements clearly and specifically in the RFP</li><li>Whenever possible, ask vendors for a trial period <em>prior</em> to signing a contract to validate that the tool addresses these requirements</li></ul><h3>2. <strong>Think about the long-term</strong></h3><p>Custom-built solutions may seem like a good way to meet a specific or immediate need, but they can lead to negative consequences in the long term. They also make it less likely that the new system will ever be able to communicate with other parts of the justice system. When products and data are integrated <em>across</em> agencies, rather than in individual silos, we can look at improvements and impact more holistically.</p><p>We should ask questions like: Does the new bodycam software output videos in a format that is well-supported by the police department’s database, or with other video software? Can other organizations, e.g. correctional facilities, easily read or access information from the probation office’s new record management system?</p><p>Choosing a custom solution also makes it more difficult to switch vendors if product requirements change down the line, leading to lock-in. When only one vendor in the world knows how your software works, you’ll only ever be able to work with that one vendor. Before committing to any tool, and especially a highly custom one, the following considerations may help to improve long-term efficacy:</p><p><strong>Checklist</strong></p><ul><li>Determine how interoperable the inputs and outputs of the tool are with other systems, both inside and outside the agency</li><li>Ensure adequate training for personnel to incorporate new tools into their workflow</li><li>Understand who owns the data collected or generated by the tool, and ensure it’s the agency</li><li>Draft shorter-term contracts that make incremental and measurable improvements to the system, rather than one big overhaul that will be difficult to change as needs change. Assume the work will take twice as long as promised.</li><li>Incorporate maintenance and update requirements (especially for security updates) into the RFP</li><li>Incorporate documentation requirements into the RFP (getting monthly exports to your data warehouse isn’t helpful if your IT staff don’t have a data dictionary from the vendor to help your team understand them)</li></ul><h3>3. <strong>Choose vendors that can explain their tools in plain language</strong></h3><p>It is often a red flag if a tech vendor can’t produce a clear answer to how a tool works or how an algorithm weighs specific factors in a predictive decision. This is unfortunately still common — many companies have <a href="https://cdt.org/wp-content/uploads/2017/08/2017-07-31-Trade-Secret-Algorithms-as-Barriers-to-Social-Justice.pdf">historically resisted disclosure</a>, calling the workings of their tools protected trade secrets.</p><p>Take the example of probabilistic genotyping software, which is used by investigators to predict the probability that a genetic sample came from a specific individual. Practitioners should ask: In what types of cases will the software likely fail? What kind of mistakes does it make (e.g., can it fail in both directions — saying the samples came from the same person in cases where they didn’t, <em>and</em> other times saying they didn’t come from the same person when they actually did)? What are the consequences of these mistakes (e.g. leading to punitive outcomes or opportunity denial)? How often do mistakes happen? If the tool’s output is challenged, do I (or the vendor) have enough information to determine and explain what went wrong?</p><p><strong>Checklist</strong></p><ul><li>Ask vendors to describe how their tools work in a language that is understandable and not overly technical. (Your agency may need to provide a similar explanation to the public someday.)</li><li>Bring the most technical member of your team along to ask all the questions they can think of, then have them write up a report internally explaining how the system works</li><li>Always assess machine learning tools by how explainable their outputs are</li><li>Ensure that staff who will be using the tool have a solid understanding of its limitations</li></ul><h3>4. <strong>Analyze for impact on underrepresented and historically marginalized people</strong></h3><p>Technology has the potential to make the criminal justice system more fair and efficient. However, <a href="https://www.brookings.edu/research/understanding-risk-assessment-instruments-in-criminal-justice/">a growing body of research</a> has shown that certain types of technology can be discriminatory. Luckily, we can assess this up-front and use what we learn to inform procurement decision-making.</p><p>All tools, especially those that use machine learning, should be assessed for performance (i.e. accuracy) across protected characteristics (i.e. race, gender, age, sexual orientation, and socioeconomic status). A good example is the <a href="https://www.nist.gov/news-events/news/2019/12/nist-study-evaluates-effects-race-age-sex-face-recognition-software">set of analyses</a> the <a href="https://www.nist.gov/">National Institute of Standards and Technology</a> conducted for facial recognition software, evaluating 189 algorithms from 99 different developers for bias across race, age, and sex. It’s a good sign if a tool has already been evaluated by another agency or academic group that has no financial ties and a history of objectivity.</p><p>For tools that benefit justice-involved individuals, bias can come in the form of denial of opportunity — for example, socioeconomically disadvantaged groups who don’t own smartphones or live in areas with reliable cellular service won’t be able to reap the benefits of court reminder apps. In those cases, we need to think of additional ways to reach these audiences. For example, this might include mail or postcard campaigns, e-mail as a backup to text messaging, or using contacts with parole and probation officers.</p><p><strong>Checklist</strong></p><ul><li>Research the different risks associated with the specific technology being procured</li><li>Ask vendors to provide information about their tool’s performance across different protected characteristics such as race, gender, age, sexual orientation, and socioeconomic status</li><li>Ask vendors to explain any disparities and how they will be mitigated</li><li>Check for unbiased, third-party analysis of the tools’s impact on protected classes. If there is none, check with other agencies who use the tool for any internal assessments or analyses they’ve tried to run.</li><li>Monitor the impact of the tool across these groups after deployment</li></ul><h3>5. <strong>Create a data-driven rollout plan</strong></h3><p>Before launch, we can develop a data-driven rollout plan to reduce the risk of failure, track impact, and inform future iterations. Rollouts can be tricky in the criminal justice context — traditional tech experimentation frameworks, like A/B testing, can become ethically questionable when people’s lives, freedoms, and futures are at stake.</p><p>At Recidiviz, we use pre-determined “backstop” metrics during our launches to determine when changes should be rolled back (essentially the opposite of success metrics). For example, if a new feature for our <a href="https://www.recidiviz.org/line-staff-tools">line staff tools</a> were to disproportionately lead to fewer positive outcomes for Black individuals on supervision, our metrics would show that as it started to happen, giving us the chance to stop the feature roll-out and make changes before trying again.</p><p><strong>Checklist</strong></p><ul><li>Align on a set of specific metrics to track before launch, including both indicators of success and criteria for rollback; this doesn’t need to be a heavyweight process and it’ll pay off when you know whether the rollout worked. These metrics should be measurable and quantifiable, not just anecdotal.</li><li>Success metrics should cover both adoption (are staff using the new tool?) in addition to impact (is it solving the problem we procured it to solve?). “Every police department has some piece of tech on their shelves that they don’t use anymore,” is something we’ve heard over and over again while talking to departments across the country.</li><li>Start with a small set of initial users and features</li><li>Proactively check in with users to understand reasons for drop-off and points of frustration</li></ul><h3>6. <strong>Weigh the benefits against the costs</strong></h3><p>Ultimately, every tool will have benefits and drawbacks. Surveillance technologies that use facial recognition can help law enforcement quickly wade through footage and identify potential suspects, but can also violate the privacy and security of those being monitored. If these costs can’t be mitigated, or if the tangible benefits can’t be realized, then it may be time to <a href="https://www.latimes.com/local/lanow/la-me-laser-lapd-crime-data-program-20190412-story.html">consider another type of tool, or none at all</a>.</p><p><strong>Checklist</strong></p><ul><li>Systematically track successes and failures attributable to the deployment of a new tool</li><li>Check how the system is performing compared to what was in place before</li><li>Regularly check in with stakeholders and users to determine if funding for the tool should continue</li></ul><h3>We want your feedback</h3><p>This is an evolving piece. We’d love to hear your feedback and suggestions on the Playbook, especially if you have experience working in this space. Reach out to us at <a href="https://www.recidiviz.org/contact">this link</a>.</p><p><em>We would like to thank </em><a href="https://www.linkedin.com/in/clementine-jacoby-50578483/"><em>Clementine Jacoby</em></a><em>, </em><a href="https://www.linkedin.com/in/samantha-harvell-359b9814/"><em>Samantha Harvell</em></a><em>, and </em><a href="https://www.linkedin.com/in/lauren-haynes/"><em>Lauren Haynes</em></a><em> for their helpful comments and contributions.</em></p><p><em>Recidiviz is proud to be a </em><a href="https://glginsights.com/glg-social-impact/"><em>GLG Social Impact Fellow</em></a><em>. We would like to thank GLG for, among other things, connecting us to some of the experts who provided valuable insights showcased in this post.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db6e2f481941" width="1" height="1" alt=""><hr><p><a href="https://medium.com/recidiviz/part-iii-tips-for-procuring-and-adopting-criminal-justice-technology-db6e2f481941">Part III: Tips for procuring and adopting criminal justice technology</a> was originally published in <a href="https://medium.com/recidiviz">Recidiviz</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Part II: Why good data and feedback loops in criminal justice are hard to come by]]></title>
            <link>https://medium.com/recidiviz/part-ii-why-good-data-and-feedback-loops-in-criminal-justice-are-hard-to-come-by-b07e7e2ee3d7?source=rss-d8eb32f35c71------2</link>
            <guid isPermaLink="false">https://medium.com/p/b07e7e2ee3d7</guid>
            <category><![CDATA[feedback-loop]]></category>
            <category><![CDATA[government-technology]]></category>
            <category><![CDATA[criminal-justice-reform]]></category>
            <category><![CDATA[data-analytics]]></category>
            <dc:creator><![CDATA[Catherina Xu]]></dc:creator>
            <pubDate>Tue, 19 Oct 2021 00:05:47 GMT</pubDate>
            <atom:updated>2022-08-08T20:24:44.401Z</atom:updated>
            <content:encoded><![CDATA[<p>Written by Catherina Xu and C. Andrew Warren</p><p><em>This is the second part in a four-part series Recidiviz has written to summarize some of the biggest issues we see with the use of technology in the criminal justice space and to create a “playbook” with specific recommendations that practitioners can utilize in tech procurement, integration, and deployment (see Part I </em><a href="https://medium.com/recidiviz/part-i-a-mapping-of-tech-in-the-criminal-justice-system-bce3c5f37fd7"><em>here</em></a><em>). As always, we welcome thoughts and feedback.</em></p><p>Imagine you’ve just built and deployed a new software tool for consumers — say, better spell-check for email. Building and launching that tool is only half the battle — the big test is what happens after. How much time are users spending with the tool? How often does it crash, or do they seem to get stuck and walk away? How satisfied are they with the usability of the tool? What do they need that it can’t do yet (for instance, checking grammar)? And most importantly, to what extent is the tool actually impacting their goal: to send messages faster and with fewer mistakes? Based on the answers to these questions and more, your team will spend a lot of time on the next stage: iterating and making incremental changes, all the while ensuring that the quality of the product stays high throughout the process.</p><p>Most products are not great when they start out — as the saying goes for any app or website, “The first day is the worst the product will ever be — it only gets better from there.” But in the criminal justice tech space, where getting something wrong <a href="https://medium.com/recidiviz/part-i-a-mapping-of-tech-in-the-criminal-justice-system-bce3c5f37fd7">can cause a lot of harm</a>, whatever goes out the door on the first day — right or wrong, good or a pain in the neck — is pretty much the same as it will ever be. The feedback loops we expect and benefit from in our private lives are often elusive. Why?</p><p>In this post, we’ll walk through a few contributing factors we’ve noticed from our experience working with governments from the tech side. The challenges essentially boils down to three things:</p><ol><li>An antiquated government procurement process</li><li>Fragmentation of users and agencies across government and the justice system</li><li>A lack of standardized data and metrics</li></ol><h3>1. Procurement: Noble goals, unintended consequences</h3><p>The modern procurement process looks something like this: an organization that wants to procure new technology assesses its needs and develops a comprehensive list of criteria, which is turned into a Request for Proposals (RFP). This document is published, and vendors submit bids for what they can do and how much it will cost the government. The government agency collects the bids, scores them according to their criteria, then selects the highest-scoring, lowest-bidding vendor for the contract.</p><p>It’s not surprising that procurement doesn’t seem tailor-fit for information technology tools. The practice in the U.S. can be <a href="https://washingtonpapers.org/resources/articles/supply-problems-plagued-the-continental-army-from-the-start/">traced all the way back to supply purchases for the Continental Army</a> in the Revolutionary War, which included ammunition, blankets, and “a large number of hogs…to be purchased, slaughtered and salted.” In fact, the oldest procurement law still on the books was <a href="https://www.governmentcontractors.org/aboutdetail.aspx?tn=NAGC-History">created just a few years later</a>, in 1808. The core principles behind these rules have more or less stayed constant:</p><ul><li>Promote competition so government can get the best price</li><li>Keep public trust — minimize risk of project failure and public fiasco</li><li>Make it difficult for corruption and unfair practices to take place</li></ul><p>All of these seem reasonable. As citizens, we don’t want our tax dollars going to waste, we don’t want to risk infrastructure projects on unproven providers, and we certainly don’t need a repeat of <a href="https://en.wikipedia.org/wiki/Tammany_Hall#Tweed_regime">Tammany Hall</a>. These are good ideas, so where do things break down?</p><h4>Size feeds complexity, and vice versa</h4><p>First, take a look back at those core principles — every one seems to push governments toward taking large contracts with well-established vendors. To ensure open and fair competition, the procurement process itself is intentionally cumbersome and long — it can take over a year to go through the bidding process and finalize a contract. Thus, governments prefer to procure in large bundles (i.e. all-in-one solutions including database management, frontend engineering, design, etc) rather than in separate pieces. The resulting contract size, combined with the government’s inherent risk-aversion, gives <a href="https://www.quora.com/What-does-the-phrase-Nobody-ever-got-fired-for-choosing-IBM-mean">large vendors with long histories</a> an undeniable edge in the process.</p><p>The size of these projects, in turn, pushes vendors toward delivering a technical system all at once after years of work, instead of making incremental changes to continually validate functionality, ensure usefulness, and expose early bugs — the exact opposite of agile development. This is like taking a journey of a million steps with a blindfold on — if you’re even a few degrees off, you’ll end up far away from where you wanted to be. And of course the world doesn’t hold still: even as workflows, requirements, security standards, and technology itself inevitably change over time, vendors are often stuck delivering what was initially agreed-upon in the contract five or ten years prior. “The length of the process is really terrible,” said a former official we talked to at a state Department of Corrections. “We can get locked into a 3–5 year contract even when we know something isn’t working.”</p><h4>The cost of switching is high</h4><p>Furthermore, winning vendors have every incentive to make it as hard as possible for the government to switch providers, even once the contract ends. The vendor’s hope is that the software becomes so complex and custom-built that any additions also have to be custom-built, creating an increasingly unmaintainable software system that only one vendor in the world can work with. Tools riddled with custom data formats, intellectual property restrictions, and other types of lock-in are not only bad for the agency and the public; nowadays, they also require more effort to build than tools that are open and standards-compliant. The fact that the vendor’s time and attention is spent on figuring out lock-in, as opposed to essential functions for the tools, is a sadly common occurrence. This helps to explain why there are so many legacy government <a href="https://www.gao.gov/products/gao-19-471">IT systems that are 30–50+ years old</a> — it’s simply too difficult and costly to switch, even when the system is clearly outdated and struggling to meet modern demands.</p><h4>What comes after isn’t top of mind</h4><p>Lastly, because contractors primarily compete on price, and because the federal grants that are often used for big capital improvements are only meant for new investments, important yet intangible factors like ease of use or maintenance often fall to the wayside. With little budget or incentive specifically attached to staff forward-looking efforts, it becomes exceedingly difficult to do much beyond the initial delivery of the product. While private sector buyers increasingly consider the total cost of ownership (TCO) of adopting new technologies, governments on the whole have yet to adopt this trend.</p><h4>There is a tech knowledge gap</h4><p>Another reason that government (and especially criminal justice) tech seems stagnant is that in some places, the agency is trying to spec their own technology without much experience doing so. If they’re not <a href="https://www.goodreads.com/quotes/15297-if-i-had-asked-people-what-they-wanted-they-would">asking for the right things</a>, they are unlikely to get them through the hyper-rigid procurement process. In others, it’s because the vendor ecosystem is more predatory, <a href="https://fcw.com/articles/2014/09/16/who-owns-gov-data.aspx">taking advantage of public servants</a> who shouldn’t have to understand the ins and outs of niche technology. In other words: there are a lot of challenges here.¹</p><h3>2. Fragmentation: When the justice system isn’t one system</h3><p>The criminal justice “system” is actually a collection of institutions from separate branches and levels of government. “There isn’t a single individual, group, or entity responsible or accountable for all stages of the system, or even a line of communication between many of the stakeholders,” one former Secretary of Justice notes. As a result, there’s no global accountability for the system’s performance, or often even a way to know when a change (to policy, practice, or tools) in one part of the system has had a positive or negative effect in the rest. In other words, there’s no feedback loop between the inputs to the justice system and its outcomes.</p><p>Say Prison A has implemented a new education program. To decide whether they should keep the program, Prison A wants to find out whether recidivism rates have decreased since its introduction. Unfortunately, this isn’t possible without reliable and timely access to arrest records from the police, and revocation data from the parole system. And arrests would only be a weak, early indicator of recidivism: if Prison A wanted to know the number of convictions that actually resulted from those arrests, they would need access to court records, too. That’s not even in the same branch of government.</p><p>Much of the organizational and technical infrastructure that would be needed to bridge these gaps doesn’t exist today, impeding the system’s ability to make informed and feedback-driven decisions. How did this happen? As the Secretary puts it: “Information sharing is simply not a priority, resulting in the system we see now — one riddled with silos that rarely come down.” Without an understanding of what’s working and what’s not, it’s hard to see the system making any great steps forward in the near future.</p><h3>3. Data: Both a problem and a potential solution</h3><p>Even if it <em>were </em>easy to get data from any institution we wanted, we would have a hard time making sense of it. Standard reporting formats, like the <a href="https://developers.google.com/transit/gtfs">General Transit Feed Specification (GTFS)</a> for the transportation industry, don’t yet exist for the criminal justice system. Any data that’s shared would have to be cleaned, reformatted, validated, and linked across silos before it could be used, requiring an amount of time and expertise that is often not available. This is part of why federal surveys of information like <a href="https://www.fbi.gov/services/cjis/ucr">arrests</a> and jail populations are voluntary, only sample some jurisdictions, and can take years to publish after being collected.</p><p>Not surprisingly, a set of standardized metrics that should be calculated doesn’t exist, either. As a result, the definition of recidivism becomes slippery and confused — if an individual commits a technical violation while on parole, that’s recidivism to some jurisdictions, while re-arrest is in others, and only a return to prison counts in still more. The fact that the answer varies from place to place makes it very hard to make meaningful comparisons or projections — information that could be critical in informing a policy decision.</p><h3>Parting thoughts</h3><p>Consider the spell-check tool example from the start of this post. If logging is set up correctly, feedback is almost instantaneous — with a simple query, we can see the number of users, the average amount of time it takes to compose an email, the average number of misspellings per email, etc. Now imagine that logging is outsourced to 10 individual companies who track different metrics but don’t talk to each other, and that each email takes <em>years </em>to go from draft, to sent, to received. That’s how the criminal justice system can look, coming from most other fields. In a space where outcomes are the opposite of instantaneous (e.g. revocation can happen years into parole) and so much is at stake, it becomes even more important for feedback loops to be as tight as possible to drive better decision-making.</p><p>Though it isn’t realistic to overhaul the system overnight, there are ways to push for change within the system and build feedback loops despite these structural limitations — our next post will outline some of what we’ve learned operating in this space.</p><p><em>We would like to thank </em><a href="https://www.linkedin.com/in/clementine-jacoby-50578483/"><em>Clementine Jacoby</em></a><em>, </em><a href="https://www.linkedin.com/in/alexkasavin/"><em>Alex Kasavin</em></a><em>, </em><a href="https://www.linkedin.com/in/samantha-harvell-359b9814/"><em>Samantha Harvell</em></a><em>, and </em><a href="https://www.linkedin.com/in/saunders-jess/"><em>Jessica Saunders</em></a><em> for their helpful comments and contributions.</em></p><p><em>Recidiviz is proud to be a </em><a href="https://glginsights.com/glg-social-impact/"><em>GLG Social Impact Fellow</em></a><em>. We would like to thank GLG for, among other things, connecting us to some of the experts who provided valuable insights showcased in this post.</em></p><h3>Footnotes</h3><p>¹ If you think the government should be able to spec its own tools and products, and agencies shouldn’t be easy victims for predatory vendors, you’re not wrong — but you’ll need to help get the public comfortable with government agencies paying wages that allow them to compete with the private tech sector for talent. Organizations like the U.S. Digital Service and 18F are helping to pave the way on this front.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b07e7e2ee3d7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/recidiviz/part-ii-why-good-data-and-feedback-loops-in-criminal-justice-are-hard-to-come-by-b07e7e2ee3d7">Part II: Why good data and feedback loops in criminal justice are hard to come by</a> was originally published in <a href="https://medium.com/recidiviz">Recidiviz</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Part I: A mapping of tech in the criminal justice system]]></title>
            <link>https://medium.com/recidiviz/part-i-a-mapping-of-tech-in-the-criminal-justice-system-bce3c5f37fd7?source=rss-d8eb32f35c71------2</link>
            <guid isPermaLink="false">https://medium.com/p/bce3c5f37fd7</guid>
            <category><![CDATA[algorithmic-bias]]></category>
            <category><![CDATA[data-analytics]]></category>
            <category><![CDATA[criminal-justice]]></category>
            <category><![CDATA[data-transparency]]></category>
            <dc:creator><![CDATA[Catherina Xu]]></dc:creator>
            <pubDate>Mon, 04 Oct 2021 21:40:31 GMT</pubDate>
            <atom:updated>2021-10-14T16:02:01.966Z</atom:updated>
            <content:encoded><![CDATA[<p>Written by Catherina Xu and C. Andrew Warren</p><p><em>This is the first part in a four-part series Recidiviz has written to summarize some of the biggest issues we see with the use of technology in the criminal justice space and to create a “playbook” with specific recommendations that practitioners can utilize in tech procurement, integration, and deployment. As always, we welcome thoughts and feedback.</em></p><p>In 2008, the Los Angeles Police Department (LAPD) began working on a project that was full of promise — to predict crime using data and analytics. Various agencies at the Department of Justice jumped on board, as did academics at UCLA. Within a few years, the resulting software, PredPol (now Geolitica) became a leading vendor of predictive policing technology in the U.S.</p><p>But the gleam of this new technology soon faded. Several studies proved that the use of historical data in these systems may have led to discriminatory policing and reinforced racial inequalities. In 2019, LAPD’s <a href="https://www.latimes.com/local/lanow/la-me-lapd-precision-policing-data-20190703-story.html">own internal audit</a> concluded that there was insufficient data to determine if PredPol actually achieved its purpose — to reduce crime. By 2020, the LAPD <a href="https://www.latimes.com/california/story/2020-04-21/lapd-ends-predictive-policing-program">had stopped using PredPol entirely </a>— and the Fourth Circuit Court of Appeals had <a href="https://epic.org/2020/07/federal-appeals-court-sounds-a.html">delivered a blow to the constitutionality of these technologies</a> altogether.</p><p>Despite the blow, enthusiasm continues to grow for incorporating technology into the criminal justice ecosystem. And rightfully so, as new technology has helped to <a href="https://uptrust.co/impact/">decrease failure to appear rates in court</a>, <a href="https://www.codeforamerica.org/programs/criminal-legal-system/clear-my-record/">automate criminal record clearance</a>, and improve <a href="https://global-uploads.webflow.com/5e7ff048d75a9b3c5df52463/5ffa6f714ad6fa2f7afeceb4_IL_3strikes_010621.pdf">policy decisions through modeling</a>. But at the end of the day, tech is a hammer — helpful when used in the right way, good at creating a mess when it’s not.</p><p>The majority of the Recidiviz team comes from tech: we are a team of data scientists, software engineers, product managers, and designers. And while we’ve seen tech as a game-changer, many have also seen the downside. In this post, we’ll explore where things can go wrong in criminal justice work when technology is inaccurate, biased, or used improperly. We’ll cover why it pays to be cautious in applying new tools to old problems, and how to take the community, the justice-involved, and the agencies that serve both into consideration when looking at the evolving tech landscape.</p><h3>Tech pitfalls in the justice system</h3><p>Technology is increasingly used at every stage of the justice system: to detect gunshots; to dispatch squad cars to far-flung neighborhoods; to decide whether defendants will be eligible for community service or diversion programs; and to monitor the location of parole and probationers to make sure they don’t miss curfew. These decisions have the potential to improve the efficacy of the justice system, but can also cause harm to already-vulnerable communities.</p><p>In the following table, we highlight some real-world examples of how these technologies have led to negative consequences:</p><ul><li><strong>Historical bias: </strong>Bias is reinforced when past or current data is used to inform decisions for the future. For example, facial recognition systems are often trained on datasets that don’t adequately represent women or darker skin tones, making <a href="http://gendershades.org/overview.html">misidentifications more likely for people in these groups</a>. These inaccuracies, according to the ACLU, can lead to “missed flights, lengthy interrogations, watch list placements, tense police encounters, false arrests, or worse.”</li><li><strong>Lack of transparency and explainability: </strong>Criminal justice tech can deny defendants the ability to challenge results from technical systems like risk assessments, which are often used to determine bail amounts and influence sentences. These tools must be tested and held accountable broadly. <a href="https://epic.org/ai/criminal-justice/index.html">Companies often cite trade secret or IP defenses</a> to avoid publishing comprehensive reports about the inner workings of their tools, e.g., the specific weight they give to each factor in making a decision.</li><li><strong>Privacy risks: </strong>When any type of personal data is collected or stored, risks loom especially large in the criminal justice context where sensitive information can be used to incriminate or surveil. According to a <a href="https://bjs.ojp.gov/content/pub/pdf/pauchi.pdf">study conducted by the Bureau of Justice Statistics</a>, nearly 1 in 8 Americans feel that they have been victims of an improper invasion of privacy by a law enforcement agency. While criminal proceedings are public information, a lack of controls around personal data of individuals in the system have often led to extortion (and resulted in <a href="https://theappeal.org/pennsylvania-county-owes-67-million-after-man-finds-arrest-records-on-mugshots-com/">litigation against agencies</a>) and <a href="https://www.themarshallproject.org/2018/05/14/the-man-with-the-stolen-name">fraud</a>.</li><li><strong>Financial hardship: </strong>Usage of technology in the justice system is not free. When costs can’t be absorbed by the state budget, they sometimes are passed along to the individuals in the system — an approach that can <a href="https://www.brennancenter.org/our-work/research-reports/steep-costs-criminal-justice-fees-and-fines#footnote1_4iq2an0">seriously undermine</a> the potential for successful re-entry into the community. Costs in Florida, for example, can <a href="https://www.vera.org/downloads/publications/the-high-price-of-using-justice-fines-and-fees-florida.pdf">top $1,000 for a misdemeanor and $5,000 for a felony</a>, excluding any additional fees incurred for use of services like phone calls or electronic ankle monitors. Costs like these prevent people who’ve completed their sentence from rebuilding their lives.</li></ul><p>These challenges can become uniquely troubling in the context of criminal justice. In the consumer context, even the most egregious privacy-abusing app can be uninstalled by its users; by contrast, the criminal justice system is compulsory for the people under its care. What’s more, those of us who help select or deploy new criminal justice technology rarely have direct experience with the justice system ourselves, due to CJIS regulations on who can work with this kind of data — making it less likely that the tech ecosystem working in this area can fully empathize with the people it impacts.</p><p>In the table below, we highlight some of the ways these risks intersect with new uses of technology in the criminal justice system. It’s not exhaustive, and most risks compound (e.g., predictive policing systems can be biased due to their training data, which directly reinforces historical bias and indirectly undermines the accuracy of the tools).</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fairtable.com%2Fembed%2Fshr7YQOidHQqF7BDl&amp;display_name=Airtable&amp;url=https%3A%2F%2Fairtable.com%2Fshr7YQOidHQqF7BDl&amp;image=https%3A%2F%2Fstatic.airtable.com%2Fimages%2Foembed%2Fairtable.png&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=airtable" width="800" height="533" frameborder="0" scrolling="no"><a href="https://medium.com/media/92c4a8b2bccca2bd77bfbef631912314/href">https://medium.com/media/92c4a8b2bccca2bd77bfbef631912314/href</a></iframe><h3>Some key takeaways</h3><p><strong>It’s apparent from the chart above that not all harms are equal</strong> — an unnecessary arrest is not the same as a leaked email address. When evaluating new technology, the following considerations can be helpful in putting the pitfalls in context.</p><ul><li><strong>Frequency: </strong>How often will the tool be used? How often does it make a mistake?</li><li><strong>Severity</strong>: What are the consequences of a tool being wrong? Are they different for false positives vs. false negatives? Can the tool lead to unnecessary arrest or incarceration, or other loss of freedoms? What safeguards are in place, and what recourse does an individual have if the tool is wrong?</li><li><strong>Impacted audience: </strong>Does the tool have the potential to adversely affect people who aren’t in the justice system, e.g., predictive policing technologies? The broader community has different constitutional protections from those who are incarcerated. Additionally, creating unnecessary first-time interactions with the system may carry more risk due to the criminogenic effect.</li><li><strong>Active harm versus opportunity denial: </strong>Can the tool cause active harm toward an individual or group? Or does it deny benefits to certain populations, such as people who don’t own smartphones?</li></ul><p><strong>Another consideration is the negative impact on criminal justice agencies themselves.</strong> Agencies have faced diminished public trust, lawsuits, and significant new <a href="https://www.nola.gov/nopd/nopd-consent-decree/">compliance regimes</a> in response to the adoption of tools that unintentionally create biased outcomes or lack of transparency. When agencies can’t explain what led to a certain result, trust in both the tool and the process can erode. Clear documentation and processes for data access are crucial — a good example is the recent push for <a href="https://www.calpublicagencylaboremploymentblog.com/public-safety-issues/governor-signs-sb-1421-and-ab-748-dramatically-increasing-public-access-to-peace-officer-personnel-records-2/">free and timely release of body camera footage</a> in California, which gained enough public support to pass as law in 2019.</p><p><strong>Finally,</strong> <strong>it’s easy for technology to create reservoirs where biases can propagate.</strong> Tools that rely on historical data are used in every step of the criminal justice system, from predictive policing and surveillance enabled by facial recognition, to sentencing and release-date calculations.</p><p>Carefully weighing community harms against community benefits can help agencies to decide when technology is the right tool for the job, and when it’s just adding risk. When technology does make sense to help solve a problem, evaluating potential edge cases and establishing safeguards to protect against these risks can be a small investment that saves lives, time, and resources for the agency and the community in the long-run.</p><h3>What can we do?</h3><p>We believe there is a useful, even crucial, role for certain types of technologies to play in the criminal justice system. But indiscriminate use of technology can also cause harm to people and weaken public trust in the system. This isn’t new — polygraph machines, social media-based <a href="https://www.engadget.com/lapd-ban-third-party-facial-recognition-clearview-ai-112526446.html">surveillance</a>, and other ill-fitting technology have come and gone before. What is new is the increased enthusiasm around information technology, and the very real limitations of it for some applications.</p><p>Over the next few blog posts, we’ll dig into a few of the structural issues that have made technology hit-and-miss in this domain, scrutinize the types of artificial intelligence used across different criminal justice technologies, and synthesize a “playbook” that we hope will be helpful to practitioners navigating this increasingly complex space.</p><p><em>We would like to thank Clementine Jacoby, John Tilley, Samantha Harvell, Mayuka Sarukkai, and Jason Tashea for their helpful comments and contributions.</em></p><p><em>Recidiviz is proud to be a GLG Social Impact Fellow. We would like to thank GLG for, among other things, connecting us to some of the experts who provided valuable insights showcased in this post.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bce3c5f37fd7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/recidiviz/part-i-a-mapping-of-tech-in-the-criminal-justice-system-bce3c5f37fd7">Part I: A mapping of tech in the criminal justice system</a> was originally published in <a href="https://medium.com/recidiviz">Recidiviz</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>