The Odyssey of Artificial Intelligence: Pride and Prejudice

Artificial intelligence could become a major tool in the struggle against cognitive biases that foster discrimination, particularly within the judicial system.

AI is neither good nor bad except insofar as we allow it to be. As researchers, programmers, and users, we must remain vigilant about the biases we may unknowingly encode into our AI systems. Only under this condition can we truly harness AI's potential to reduce prejudice and discrimination, and move toward a fairer world and a more inclusive and impartial environment.

Artificial intelligence is not inherently good or bad. It is a tool shaped by human intention. The same technology that can inadvertently amplify human biases can also be leveraged to reduce them and promote greater fairness. It is our responsibility to guide AI in the right direction.

Unlike humans, whose biases can become deeply entrenched over time, AI can be adjusted rapidly and efficiently as soon as a bias is detected. There is no need for lengthy educational or societal campaigns to change AI behavior. A programmer can implement corrections with just a few lines of code.

When applying artificial intelligence to the judicial domain, the goal is not to replace human judgment but to enhance it, making us more aware of our inherent biases and helping us counteract them. Just as a spellchecker alerts us to a misspelled word, AI could alert us to potential biases, guiding us toward a more just and equitable society.

Judicial decisions, however, are systematically affected by racial and gender biases. We know this from sophisticated analyses of historical evidence, but also from simple observation of the real world. For example, federal judges appointed by Republicans tend to impose harsher sentences on Black defendants and more lenient sentences on female defendants. Federal judges also appear to behave more politically in the period leading up to presidential elections, particularly those residing in highly competitive states. A judge's political affiliation can even be predicted from the citations they choose to support their decisions.

Hope

These findings reveal persistent bias. On the one hand, a judge's identity can lead to arbitrary decisions. For example, a judge's racial identity is predictive of disparities in sentencing outcomes. On the other hand, seemingly trivial factors, such as whether the local football team won or lost, or whether it is a litigant's birthday, can also influence judicial decisions. Moreover, marginalized groups consistently bear the punitive consequences of these departures from objectivity.

As societies continue to grapple with prejudice, a controversial yet critical battleground has emerged: artificial intelligence. The digital world mirrors the analog one, and AI systems are vulnerable to the biases embedded within our societies.

Yet there is reason for optimism. Unlike humans, AI is fundamentally flexible. It can be reprogrammed and refined to mitigate biases, a far more straightforward process than reshaping deeply rooted human prejudices.

Diagnosing Prejudices

Drawing on a study of gender attitudes in the United States Courts of Appeals, I will illustrate how AI can counter prejudice more effectively than humans. It can do so by diagnosing bias in ways that humans cannot.

The study in question employs natural language processing (NLP), a branch of artificial intelligence, to detect judges' attitudes toward women. Researchers developed a measure of "gender bias" to assess the extent to which judges associate men with careers and women with family roles in their written opinions. This nuanced approach uncovered subtle sexist biases that direct analysis of judicial decisions had failed to detect.

AI's unique strength lies in its ability to aggregate and objectively analyze vast amounts of data far beyond human capacity. NLP provided a quantitative and unambiguous measure of gender bias by examining 380,000 published judicial opinions, a task that would be impractical, if not impossible, for humans to complete within a comparable timeframe. AI therefore holds exceptional potential for detecting bias beyond human capabilities.

Gender Bias

Gender bias is a compelling indicator of partiality. Female and younger judges tend to exhibit lower levels of gender bias. Having a daughter is also associated with reduced sexist bias. Lower levels of bias are linked to more frequent use of gender-neutral pronouns such as "he or she." Finally, judges with stronger gender biases tend to express less empathy toward women in their written opinions.

Not only do judges differ systematically in how they write about gender, but these differences are also predictive of how they rule in women's rights cases and how they interact with female colleagues. The study examines how judges with differing attitudes toward gender interact with female judges in three areas: reversing lower-court decisions, assigning opinions, and citing judicial opinions.

The findings show that judges with stronger gender biases are more likely to overturn decisions issued by female district judges, less likely to assign opinions to female judges, and less likely to cite opinions written by women. These judges also tend to vote more conservatively in gender-related cases. The results suggest that sexist attitudes may hinder the career progression of female judges and reinforce gender disparities within the judicial system.

A Systemic Problem

The underrepresentation of women at the highest levels of the legal profession has received considerable attention in the United States. It is striking that, although women have represented nearly 45% of law school graduates since the 1990s, they account for only 20% of partners in major law firms and approximately 30% of federal and state judges. These disparities point to a systemic issue: the differential treatment of female judges, potentially driven by sexist attitudes among their colleagues.

Gender attitudes, meaning the biases and stereotypes individuals hold about social groups, particularly women and racial minorities, are known to significantly influence judgments and decisions. These biases affect outcomes across a wide range of contexts, from medical treatment and hiring decisions to employer-employee relations and even teaching effectiveness. If such attitudes result in differential treatment of female judges, they may contribute to the underrepresentation of women within the judiciary.

Studying these issues among judicial actors is challenging because traditional measures of gender attitudes are generally unavailable for judges. Researchers have therefore innovatively applied recent advances in natural language processing to develop a new measure of gender attitudes. By analyzing a vast corpus of appellate judges' writings, they constructed a measure of sexist bias based on the strength of associations between men and careers and women and family in judicial texts. Using a technological tool known as "word embeddings," the researchers calculated a judge-specific measure of gender bias.

As we move from analysis to practical application, AI could be used to counter the sexist biases it detects. Systems could be programmed to prompt judges to reflect upon and reconsider potential implicit biases in their reasoning.

Rather than replacing human judgment, AI could serve as an intelligent safeguard, helping decision-makers recognize and overcome hidden prejudices. If designed responsibly, it could become a powerful ally in the pursuit of a more impartial, equitable, and inclusive justice system.

Full article in L'Opinion, published on August 11, 2023.
Illustration: Unsplash