How AI is Revolutionizing Criminal Justice

• By Mehaan Ahuja

How AI is Revolutionising Criminal Justice

Artificial intelligence (AI) is making a big splash in criminal justice and is set to become even more influential. In 2023, AI’s impact on the system is huge, affecting everything from crime monitoring and prevention to the judicial and correctional systems. Its influence is everywhere—traffic safety systems, crime forecasts, criminal pattern recognition, and more. Even if you never directly interact with the system, AI in criminal justice affects you in some way.

AI in Action

Smarter Surveillance

Law enforcement is using AI to enhance video and image analysis. Today's tech goes beyond just identifying people and objects; it can detect complex accident and crime scenes in real-time and post-event. Recent advancements even allow for improving the detection of faces captured with poor image quality, at odd angles, or partially obscured. By degrading clear images to mimic lower-quality ones, AI can use these representations to make identifications.

Advanced DNA Analysis

Forensic DNA evidence has revolutionised criminal justice since the late 1980s, helping solve cold cases and exonerate wrongfully convicted individuals. AI now enables labs to process low-level, degraded, or previously unusable DNA evidence. It can detect tiny amounts of DNA and extract usable samples from old evidence. AI algorithms can also decipher large, complex data sets to separate and identify individual DNA profiles, even from mixed samples.

Gunshot Detection

AI algorithms are used to analyse gunshot patterns, differentiate muzzle blasts, determine timings, assign shots to specific firearms, and estimate identifying probabilities. Sensors installed in buildings and streetlights capture the timing and sound of gunshots, helping pinpoint the shooter’s location and allowing police to respond quickly without needing a call.

Audio analysis of gunshots is based on the observation that the content and quality of gunshot recordings are influenced by firearm and ammunition type, the scene geometry, and the recording device used. Advanced computational techniques can exploit these facts to answer investigative questions. For example, in much the same way that signal processing for human speaker recognition can help reach conclusions regarding the gender, age, identity, or national origin of the speaker, it appears possible that similar methods may be able to answer firearm-specific questions from audio data.

To date, these advanced computational methods have not been applied to the audio analysis of gunshots. As more crimes are captured on audio recordings, more examiners will be asked to answer firearm-based questions. The proposed aims include a significantly more mathematically rigorous approach than has been previously performed. We will use a fine-grained mathematical representation of the frequency spectrum and analyze this representation with a series of advanced machine learning techniques for clustering and pattern recognition. Through the proposed aims, we will assemble a large set of recorded gunshots, make these recordings available to the research community, and develop advanced algorithms to provide investigative insight. We will develop algorithms to detect gunshots, differentiate muzzle blasts from shock waves, determine shot-to-shot timings, determine the number of firearms present, assign specific shots to firearms, and estimate probabilities of class and caliber. These aims represent a series of R&D steps towards the creation and validation of computational tools for the audio analysis of gunshots.

The proposed work will develop analytic techniques, grounded in mathematical science and able to provide quantified answers to the audio analysis of gunshots. This should benefit law enforcement and their ability to utilize this important new source of evidence. With the increased prevalence of smartphones and body cameras, the work should impact investigators country-wide.

Predicting and Preventing Crime

One of AI's biggest impacts on law enforcement is shifting from reactive to proactive crime prevention. Decades of crime data and new technologies enable agencies to spot patterns in financial records, geospatial imagery, surveillance footage, social media, public records, news feeds, and more. This understanding helps allocate resources better, significantly reducing crime by adjusting when, where, and how to deploy personnel and tools.

Fingerprint Analysis

Fingerprint analysis is critical to the success of the nation's criminal justice system. In fact, fingerprints left at a crime scene — referred to as latent prints — are the most common type of forensic science evidence and have been used in criminal investigations for more than 100 years.

The examination of fingerprint evidence consists of a series of steps involving the comparison of latent print to a known print (called an 'exemplar'). During this step-by-step matching process, latent print examiners must reach correct conclusions; they are also expected to produce records of the examination and, in some cases, present their conclusions — and the reasoning behind them — in court.

In recent years, the accuracy of latent print identification has been the subject of increased scrutiny, including from the National Academy of Sciences in their 2009 report, "Strengthening Forensic Science in the United States: A Path Forward." To help address this issue, the National Institute of Justice and the U.S. National Institute of Standards and Technology (NIST) convened an expert working group to do a scientific assessment of the effects of human factors on forensic latent print analysis and to develop recommendations to reduce the risk of error.

The working group addressed issues ranging from the acquisition of impressions of friction ridge skin to courtroom testimony, and from laboratory design and equipment to emerging methods for associating latent prints with exemplars. In addition to a comprehensive discussion of how human factors relate to all aspects of latent print examinations — including communicating conclusions through reports and testimony — the report offers specific recommendations to improve the understanding and management of human-factors issues in fingerprint analysis.

Creating a Culture of Openness Among Examiners

The working group made a number of important recommendations that touch on everything from the time latent prints are submitted for examination to testimony in the courtroom. Among the recommendations is the need "to create a culture in which both management and staff understand that openness about errors is not necessarily a path to punitive sanctions but rather is part of an effective system to detect deviations from desired practices and incorrect judgments in latent print casework."

To help achieve such a culture, the panel recommended that management in forensic crime laboratories:

• Employ a system to identify and track errors and their causes.

• Establish policies and procedures for case review and conflict resolution, corrective action, and preventive measures.

One particularly helpful tool in the report is a flowchart that shows the Analysis, Comparison, Evaluation, and Verification (ACE-V) process for latent print examination that is currently used in the nation's forensic crime laboratories. This chart, developed by the working group, offers a detailed view of steps in the ACE-V process where human error risks should be minimized.

Research and Development Recommendations

The working group also identified a "critical need" for research into the interpretive process that is at the heart of the ACE-V process. "For example," they said, "only a handful of studies have assessed variation in the feature selection process. Generally, the findings have shown wide variation among examiners during the task of feature selection."

Other areas where research is recommended include:

• Determining the educational and cognitive abilities that should be prerequisites for training a latent print examiner.

• Developing methods that reduce the variation of the print features selected for comparisons.

• Developing a better understanding of the link between variations in feature selection and the examiner's ultimate determination.

• Addressing the utility/sufficiency of latent print analysis.

• Developing measures/metrics for latent print analysis — and using these to assess the reproducibility, reliability, and validity of various interpretive stages of the process.

• Identifying key factors related to variations in the performance of latent print examiners during the interpretation process.

• Creating large, anonymous databases of exemplars and latent prints to facilitate the validation of probabilistic models and other statistical research.

• Determining the most appropriate tests of visual function for friction ridge examiners.

In particular, the working group said that the federal government should support research programs to improve automated fingerprint identification systems. Such programs could, for example:

• Expand the algorithms used to match prints to account for the fact that the diagnostic value of minutiae depends on the region in which they are located.

• Make fingerprint and palm print databases interoperable among local, state, and federal automated identification systems.

• Increase the compatibility between automated identification systems and other latent print software tools, including digital enhancement programs, probability calculation programs, and automated quality assessment programs.

Fraud Detection with AI

AI in fraud detection uses algorithms to monitor data and stop fraud before it happens. It learns from historical data and adjusts its rules to detect new threats, something standard fraud software can't do. AI also reduces false positives by improving rule accuracy without impacting user experience. The best AI cybersecurity solutions are lightweight and don't affect website or app performance.

Common Types of Fraud AI Can Detect

• Card Fraud: AI can distinguish bots from people to block card fraud bots.

• Fake Account Creation: AI tracks variables to block bad bots without affecting user account creation.

• Account Takeover (ATO): AI traces bots to stop ATOs discreetly.

• Credential Stuffing: AI monitors traffic changes and login failure rates to stop credential stuffing attacks.

AI provides real-time fraud detection and protection, improving over time with data and sharing knowledge globally among its instances. It reduces false positives and allows employees to focus on business-forward projects.

Benefits:

• Real-time detection.

• Continuous improvement.

• Reduced reactive time.

Risks:

• Social fraud remains a risk.

• AI can be a "black box" and hard to understand.

• False positives, though minimal, can occur.

Beyond the Basics

AI has countless other applications in criminal justice, including:

• Traffic safety systems that catch violators and improve safety through automatic accident detection.

• Assisting medical examiners in determining causes and manner of death through radiological image analysis.