Large Language Models, Artificial Intelligence and the Future of Law
Session 8: How do we account for algorithmic bias in AI applications?
Embedded Files
What is Algorithmic Bias?
What is Algorithmic Bias?
Algorithmic bias refers to systematic and repeatable errors in a computer system that create unfair outcomes, such as privileging one arbitrary group of users over others. This bias in algorithms can stem from a variety of sources, including but not limited to:
Data Bias: If the data used to train an algorithm reflects existing prejudices or lacks diversity, the algorithm may replicate or even amplify these biases. For instance, if a facial recognition system is trained primarily on images of people from a particular ethnic group, it may perform poorly on individuals from other ethnic groups.
Design Bias: The way an algorithm is designed and the objectives it is set to achieve can introduce bias. For example, if an algorithm is designed to maximize engagement on a social platform without considering the quality of content, it might favor sensational or divisive content.
Human Bias: Human decision-makers can inadvertently influence algorithms through their own biases. This can happen during the programming phase, where the choices about which features to consider or exclude can reflect subjective preferences.
Feedback Loops: Algorithms can sometimes create a feedback loop where they learn from outputs they generated themselves, which can reinforce initial biases. For example, predictive policing tools might send more police to neighbourhoods they predict will have more crime, thereby increasing the likelihood of recording more crime there simply due to increased police presence.
Source: British Medical Journal from https://research.aimultiple.com/ai-bias/
Amazons AI Recruitment Tool
Amazons AI Recruitment Tool
In 2014, Amazon started using an AI-based tool to shortlist CVs.
See also Bertrand, Marianne, and Sendhil Mullainathan. "Are Emily and Greg more employable than Lakisha and Jamal? A field experiment on labor market discrimination." American economic review 94, no. 4 (2004): 991-1013.
Facebooks Ad Algorithm
Facebooks Ad Algorithm
Facebook uses user information including demographic details to recommend ads to those users.
They developed an AI algorithm for these targeted ads.
See also Google’s online advertising system
PredPol Predictive Policing
PredPol Predictive Policing
UCLA and National Science Foundation developed PredPol- now one of the most widely used predictive policing algorithm.
See also Heavan, W. D, (2021) Predictive policing is still racist—whatever data it uses
COMPAS (Correctional Offender Management Profiling for Alternative Sanctions)
COMPAS (Correctional Offender Management Profiling for Alternative Sanctions)
See Also
Angwin, Julia; Larson, Jeff (2016-05-23). "Machine Bias". ProPublica.
Flores, Anthony; Lowenkamp, Christopher; Bechtel, Kristin. "False Positives, False Negatives, and False Analyses" (PDF). Community Resources for Justice.
But How do we Detect Bias?
But How do we Detect Bias?
Source: Angwin and Larson (2016)
Source: Flores et. al (2016)
Source: Engel et al. (2024)
How do you correct or account for Algorithmic Bias?
1. Diverse and Representative Data:
Ensure that the data used to train algorithms is representative of all groups that will interact with or be affected by the system. This involves including diverse demographics, ensuring gender balance, and considering various socio-economic backgrounds. Oversample if required.
2. Representative Design:
Integrate bias mitigation directly into the algorithmic training process. This could involve adding fairness constraints or regularization terms to the learning algorithm to penalize biased decisions.
3. Correct in RLGF
Adjust the output of algorithms to ensure fairness. For example, if a hiring algorithm shows disparities in recommendations between different groups, post-processing can adjust these recommendations to ensure parity.
4. Bias and Fairness Audits:
Regularly conduct audits to check for biases. This can be done through statistical tests to measure and quantify fairness in the system’s outputs.
5. Transparency and Explainability:
Develop algorithms that are not only accurate but also interpretable. This means being able to explain in understandable terms how the algorithm makes decisions, which can help identify where biases may occur.
Page updated
Google Sites
Report abuse