TLDR: The research paper “EVIDENCE WITHOUT INJUSTICE: A NEW COUNTERFACTUAL TEST FOR FAIR ALGORITHMS” by Loi, Di Bello, and Cangiotti introduces the Counterfactual Independence Principle (CIP). This principle evaluates whether the evidential value of an algorithm’s output would remain probative in a hypothetical world free from structural injustices. It helps distinguish between morally problematic algorithms, like predictive policing relying on historically unjust crime patterns, and less problematic ones, like camera-based systems tracking ongoing offenses. CIP suggests caution for punitive uses of evidence that fail the test, while encouraging supportive interventions. It also applies to healthcare, differentiating between predictors rooted in biological factors versus those tied to social disadvantages.
In the rapidly evolving landscape of artificial intelligence, ensuring fairness in algorithmic decision-making has become a critical challenge. A new research paper, “EVIDENCE WITHOUT INJUSTICE: A NEW COUNTERFACTUAL TEST FOR FAIR ALGORITHMS”, by Michele Loi, Marcello Di Bello, and Nicolò Cangiotti, introduces a novel framework to assess the moral acceptability of algorithms, particularly when they operate against a backdrop of historical and structural injustices.
The paper begins by presenting a compelling thought experiment involving two types of policing algorithms. The first is a predictive policing algorithm that uses historical crime data to forecast where future crimes are likely to occur, directing police patrols to those areas. The second is a camera-based system that records ongoing offenses and guides officers to the scene. Both algorithms, despite their different mechanisms, often lead to increased police presence in minority communities, resulting in disproportionate harms like stops, searches, and arrests.
Intuitively, many people find the camera-based system less morally troubling than the predictive policing algorithm, even if both lead to similar outcomes in terms of police deployment and disproportionate impact. The core question the authors tackle is why this moral distinction exists, especially when traditional fairness metrics—such as equalized odds, causal fairness, or even concepts like compounding injustice—fail to adequately explain it.
Existing statistical criteria often focus on equalizing error rates across groups, but both algorithms can produce similar disparities. Causal and counterfactual accounts, which examine whether an algorithm’s output would change if an individual’s race were different, also fall short because race is deeply intertwined with structural injustices, making it difficult to isolate as a single variable. Similarly, the idea of compounding injustice, where algorithms perpetuate prior wrongs, applies to both systems. Even arguments around individual agency or an algorithm’s ability to ‘track truth’ don’t fully capture the nuanced moral difference.
Introducing the Counterfactual Independence Principle (CIP)
Loi, Di Bello, and Cangiotti propose the Counterfactual Independence Principle (CIP) as a new lens. CIP asks whether the evidential value of an algorithmic output would retain its probative power in a hypothetical world free from relevant structural injustices. The principle suggests that for punitive uses of evidence (like dispatching police or restricting liberties), greater caution is needed when the evidence fails this test.
Applying CIP to the policing examples, the camera-based system passes the test. Its ability to capture perpetrators at crime scenes relies on physical mechanisms that would still hold true in a just society. The evidentiary value of video footage, including identifying characteristics, is independent of structural injustice. However, the predictive policing algorithm, which relies on historical crime data, often fails. Its accuracy is tied to the systematic spatial concentration of crime, which is a legacy of structural injustices like segregation, redlining, and economic exclusion. In a world without these injustices, past crime locations would be poor indicators of future crime locations.
The authors provide a step-by-step heuristic for applying the CIP test:
- Identify the mechanism linking evidence to an outcome.
- Determine if this mechanism is part of structural injustice.
- Evaluate a counterfactual world where the unjust mechanism is removed and ask if the evidentiary link still holds.
- If it holds, the evidence passes; otherwise, it fails.
For instance, the causal chain linking minority neighborhoods to crime via redlining, segregation, and disinvestment highlights how historical crime data’s predictive value depends on unjust policies. In contrast, crime clustering in busy entertainment districts, driven by crowd dynamics, would retain its predictive value even without structural injustice, thus passing the CIP test.
Applications Beyond Policing
The CIP extends beyond criminal justice. In healthcare, algorithms that predict treatment effectiveness based on genetic factors or biological comorbidities (even if socially caused, like type 2 diabetes) would pass the CIP test because the biological relationship would still hold in a just world. However, algorithms relying on social variables like lower income or housing instability to predict poor treatment outcomes would fail. These factors predict reduced effectiveness not due to biological limitations, but because unjust social conditions interfere with treatment adherence. In a fairer world, these factors would no longer predict reduced effectiveness.
Also Read:
- Rethinking AI Ethics: Why Current Evaluation Methods Fall Short in Measuring Systemic Harms
- Navigating Legal Precedents: A New Logic for AI in Law
Punitive vs. Supportive Interventions
A crucial distinction made by the paper is between punitive and supportive uses of algorithmic evidence. When evidence fails the CIP test (meaning its probative value rests on structural injustice), the principle recommends caution for punitive uses but encourages supportive interventions. For example, if an algorithm identifies patients at high risk of diabetes due to structural injustices, this information should be used to direct preventive care and support, not to deny treatment. Similarly, historical crime data, while problematic for punitive policing, could guide supportive community programs like improved street lighting or neighborhood watch support.
Conversely, when evidence passes the CIP test (its probative value is independent of injustice), the principle imposes no additional moral constraints, allowing its use in both punitive and supportive contexts, such as using blood alcohol concentration tests to sanction impaired drivers or offer counseling.
Ultimately, the Counterfactual Independence Principle acts as a moral filter, guiding how we should use evidence generated by algorithms. It pushes us to scrutinize the foundations of our data and to ensure that our pursuit of efficiency does not inadvertently perpetuate or exacerbate existing injustices.
