top of page

Racial bias in AI: unpacking the consequences in criminal justice systems

Artificial Intelligence (AI) isn’t just about smart speakers and self-driving cars. It’s playing a big role in something a lot more serious: our criminal justice systems. AI is popping up fast in this sphere, from helping decide who gets parole to predicting where crimes might happen. But here's the catch: AI isn’t as unbiased as we'd like to think, especially when it comes to race.

The tricky world of racial bias in AI within the halls of justice is a topic that’s not just about lines of code and algorithms, but about real people and real lives.

Take this as a wake-up call: a study by the National Institute of Standards and Technology (NIST) revealed a pretty shocking fact. Those facial recognition tools that police love to use? They mess up way more often with people of color than with white folks (Grother et al., 2019). That’s not just a small glitch. It’s a big problem that could lead to some serious unfairness.

So, with a keen eye on criminal justice and a dash of tech-savvy curiosity, let's unravel this complex and important issue. We’re not just talking tech here; we’re talking about fairness, justice, and real human stories.

From records to predictions: the evolution of AI in criminal justice 

The journey of AI in criminal justice is a tale of technology evolving from basic record-keeping to advanced predictive analytics. In the 1970s, law enforcement agencies started digitizing records, a humble beginning that set the stage for today's sophisticated algorithms. These AI systems, such as COMPAS (Correctional Offender Management Profiling for Alternative Sanctions), are now integral in predicting regression risks and aiding in parole decisions (Brennan, et al., 2009).

Flash forward 50 years and we’ve implemented procedures such as predictive policing. Not unlike the exaggerated concept depicted in the popular film Minority Report (2002), the controversial idea of predictive policing is one that’s been actively integrated across many of today’s criminal justice systems. These algorithms aim to forecast where and when crimes are likely to occur by analyzing historical crime data and other factors to allocate police resources more efficiently. But, the efficiency comes with a caveat - the potential perpetuation of historical biases. If past policing efforts tended to be concentrated in minority neighborhoods, AI could potentially interpret this as a higher crime area, leading to a self-fulfilling prophecy of continuous over-policing. 

Unveiling and defining racial bias in AI

But what does racial bias in AI actually mean? This concept refers to situations where AI systems exhibit prejudice based on race, often due to the data they're trained on. This can manifest in various ways, from facial or voice recognition systems misidentifying people of color to risk assessment tools disproportionately flagging minority individuals as high-risk. It's a complex issue rooted in the very fabric of how AI learns and makes decisions.

The key to remember with AI systems is that they’re only as good as the data they’re fed. When the input data is biased, the output is likely to inherit that bias. This issue is particularly poignant in the context of racial bias. A study by Dressel and Farid (2018) found that widely used risk assessment algorithms in the criminal justice system showed little to no racial disparity in their predictions. The problem, however, lies in the data used to train these algorithms, which often reflects systemic racial biases in arrest and conviction rates.

Real-life examples: AI's bias in action

One of the most glaring examples of racial bias in AI within criminal justice is the case of facial recognition software. A study by Buolamwini and Gebru (2018) found significant inaccuracies in commercial gender classification systems, particularly for darker-skinned females. These errors aren't just numbers; they translate into real-world consequences, like wrongful arrests based on misidentification.

The ProPublica investigation (2016) into the COMPAS risk assessment tool highlighted a similar, disturbing find: Black defendants were often predicted to be at a higher risk of reoffending than they actually were, compared to white defendants. This, of course, indicates a deep-seated issue in the way AI interprets and learns from racially skewed data.

The consequences of these biases are definitely far-reaching. According to the Sentencing Project, Black Americans are incarcerated at more than 5 times the rate of white Americans. When biased AI tools are used in sentencing, bail, or parole decisions, they may exacerbate these existing disparities, creating a cycle of disadvantage for certain racial groups.

A vivid example occurred in Detroit, where Robert Williams, an African American man, was wrongfully arrested in 2020 due to a facial recognition mismatch. This case, reported by the ACLU, demonstrates how such biases can lead to real and devastating errors, impacting innocent individuals' lives and freedom.

Digging deeper: where does the bias originate?

The root of this bias often lies, as previously mentioned, in the data used to train AI systems. If the data used to train these AI systems reflects historical inequalities - like disproportionate arrest rates for certain racial groups - the AI will likely mirror these biases going forward and base its decisions and suggestions off of this skewed data. Moreover, the decision-making process in AI algorithms can be opaque, making it challenging to pinpoint how and why these biases occur.

Societal impact: a cycle of disadvantage

The ripple effects of AI bias extend into broader societal consequences. Biased AI systems in criminal justice can reinforce and amplify existing stereotypes and inequalities. This not only affects the perception of and treatment towards minority groups but also influences policy decisions, resource allocation, and public opinion, perpetuating a cycle of inequality and mistrust.

These biases raise significant legal and ethical concerns, as they challenge the principle of equal treatment under the law and question the fairness and impartiality of AI-assisted decisions. The legitimacy of AI tools in criminal justice should, of course, be questioned - how can AI be made more fair when applied within the criminal justice system? 

Tackling the bias: steps towards fairer AI

The first step in mitigating racial bias in AI is acknowledging its existence and understanding its roots. This typically involves first critically examining the data sets used for training AI, and ensuring they’re both fairly representative as well as free of historical biases. 

Some more recent efforts to combat AI bias have to do with regulatory measures and ethical frameworks, implemented by organizations such as the European Union (EU). The EU’s proposed Artificial Intelligence Act emphasizes strict rules for high-risk AI systems, including those used in criminal justice (European Commission, 2021). Ethical guidelines, like those from the AI Now Institute, advocate for transparency and accountability in AI systems.

Diversity in the teams developing AI systems is also crucial. By including voices from various racial and cultural backgrounds, AI development can better anticipate and address potential biases. Another important step to take in creating fairer AI systems includes implementing transparency in AI algorithms that can help identify where these biases occur. This, coupled with the right accountability measures (be those personnel-wise or algorithm-wise), ensures that when AI systems do exhibit bias, there are mechanisms to correct them.

Combating bias in AI is an ongoing process - it requires continuous monitoring and updating of AI systems to ensure they evolve with our understanding of fairness and justice.  Experts from various fields such as computer science, sociology, and law, emphasize a multidisciplinary approach, advocating for diverse teams and inclusive data sets to ensure AI systems don't perpetuate existing biases.

On the tech front, in recent years researchers across multiple industries are starting to develop algorithms that can detect and mitigate bias - this includes 'fairness-aware' machine learning, which adjusts algorithms to account for imbalances in data and decision-making processes.

A comparative case study: success vs. failure of AI in criminal justice

Case of bias: racial disparities in voice recognition systems

A less commonly discussed but significant example of AI bias is found in voice recognition systems used in criminal investigations and courtrooms. Research has shown that these systems have higher error rates for ‘African American vernacular English’ compared to ‘General American English’ (Koenecke et al., 2020). The study by Koenecke et al. (2020) revealed a critical issue in voice recognition technologies, particularly in their application within criminal investigations and courtrooms. The research focused on five major tech companies' speech recognition systems - Amazon, Apple, Google, IBM, and Microsoft - and found that these systems had error rates of 35% for ‘African American vernacular English’, compared to just 19% for ‘General American English’.

This disparity becomes especially problematic in legal settings. For instance, voice recognition is increasingly used for evidence collection, witness statements, and even in courtroom proceedings. If a system fails to accurately transcribe a Black person's speech, it could lead to crucial misunderstandings. This, in turn, means that misinterpretations of spoken evidence or testimonies due to AI errors could potentially affect the outcomes of trials and legal proceedings, leading to unfair sentencing or wrongful convictions.

The broader implication we can see from this failure is a trust gap in AI technologies - when minority groups cannot rely on these systems to understand and represent them accurately, it not only perpetuates existing biases but also erodes confidence in the fairness of the legal system.

The other side of the coin: successfully mitigating bias through the New Jersey bail reform

In 2017, New Jersey implemented a significant reform in its bail system, shifting away from cash bail to a risk-based assessment, largely driven by an AI tool called the Public Safety Assessment (PSA). This tool was designed to assist judges in making informed decisions about pretrial release based on risk rather than financial ability. The PSA evaluates risk using factors like age, current charge, and prior convictions, but notably, it does not include race or socioeconomic status as variables. This design choice was intentional to minimize potential biases.

The impact of this reform was substantial: by 2019, New Jersey’s pretrial jail population decreased by 43.9% from its 2015 levels. An important outcome to note is that this decrease did not lead to an increase in crime rates. Reports indicated a 16% decrease in violent crimes statewide, showcasing that public safety was not compromised (New Jersey Courts, 2019).

Moreover, the reform addressed racial disparities in the justice system. The New Jersey Judiciary's 2019 report noted that the percentage of Black individuals in pretrial detention decreased after implementing the PSA. This suggests that the AI tool, alongside the broader reform, contributed to reducing the disproportionate impact of the bail system on minority communities.

Analysis: understanding the difference between these cases

These two cases illustrate the varied implications of AI in criminal justice - while voice recognition systems demonstrate how biased training data can lead to discriminatory outcomes, the PSA in New Jersey shows that AI, when underpinned by diverse and comprehensive data, can enhance fairness. This contrast highlights the need for careful evaluation and continuous monitoring of AI systems in order to ensure they serve justice without prejudice.

The path forward: better understanding AI in criminal justice

The integration of AI in criminal justice systems presents opportunities rich with potential, yet fraught with ethical challenges. Our exploration of this topic reveals the dual nature of AI: a powerful tool capable of enhancing efficiency in legal processes, but also a mirror reflecting societal biases.

Consider the evidence from the study by Koenecke et al. (2020), highlighting the racial biases in voice recognition systems, and contrast this with the success of New Jersey's bail reform. Do these cases highlight a fundamental truth about AI? That perhaps its impact is largely dependent on how it is implemented and trained? Paying close and careful attention to ethical and unbiased processes may be the key to ensuring fair AI use throughout the criminal justice system. 

For readers, these insights should offer a window into the complexities of AI in criminal justice - the knowledge that AI systems, like the ones in New Jersey, can be used to promote fairness should serve as a catalyst for demanding responsible AI deployment across other industries and sectors. 

It seems clear that policymakers and AI developers are at a crucial point, where decisions they make shape the future of justice and technology. Prioritizing the creation of AI systems that are not only intelligent but also equitable should be key - and this includes ongoing efforts to diversify training data, transparency in AI algorithms, and a commitment to multidisciplinary collaboration.

And this call to action extends beyond professionals to every reader: stay informed, critically evaluate the AI technologies in your environment, and advocate for systems that promote fairness and justice for everyone. The advancement of AI should not just be a testament to human ingenuity, but also a reflection of our dedication to equality and ethical responsibility.


AI Now Institute. (n.d.). AI Now 2019 Report. Retrieved from

Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016). Machine Bias. ProPublica. Retrieved from

Brennan, T., Dieterich, W., & Ehret, B. (2009). Evaluating the Predictive Validity of the COMPAS Risk and Needs Assessment System. Criminal Justice and Behavior, 36(1), 21-40.

Buolamwini, J., & Gebru, T. (2018). Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. Proceedings of Machine Learning Research, 81, 1-15.

Dressel, J., & Farid, H. (2018). The accuracy, fairness, and limits of predicting recidivism. Science Advances, 4(1), eaao5580.

European Commission. (2021). Proposal for a Regulation laying down harmonized rules on artificial intelligence (Artificial Intelligence Act). Retrieved from

Grother, P., Ngan, M., & Hanaoka, K. (2019). Face Recognition Vendor Test Part 3: Demographic Effects. National Institute of Standards and Technology. Retrieved from

Koenecke, A., Nam, A., Lake, E., Nudell, J., Quartey, C., Xiong, C., ... & Raji, I. D. (2020). Racial Disparities in Automated Speech Recognition. Proceedings of the National Academy of Sciences, 117(14), 7684-7689.

MDRC. (2020). Evaluation of the Public Safety Assessment: Results from Three New Jersey Counties. Retrieved from

New Jersey Courts. (2019). 2019 Report to the Governor and the Legislature. Retrieved from

The Sentencing Project. (n.d.). Criminal Justice Facts. Retrieved from

Alex Olsson

Alex has six+ years of experience in copywriting, translation, and content editing, and has a broad background in different fields of knowledge such as international business, digital marketing, psychology, criminology, and forensic science. She works as a content specialist at a digital marketing consultancy in Stockholm and has experience writing content for clients in the financial, e-commerce, education, and media industries. Alongside her employment, Alex works with non-profits that engage with human rights topics such as advocacy for sustainable development as well as for victims of sexual assault.

35 views0 comments
bottom of page