Bias in AI: Types, Categories & Mitigation
Today I’m starting a series on bias in AI (includind of course Generative AI biased LLMs). Why? Because biases are an integral part of AI models, and better understanding them will improve their crucial mitigation.
The hashtag for this series, which will focus on analyzing the types and outcomes of bias in Machine Learning models, will be: #WhatAreAIbiasAndWhatsNext on social media (Linkedin & X).
In this first episode, we’ll define what bias is and (attempt to) categorize the different types of bias in AI (this list will obviously not be exhaustive, but we’ll try to make it as comprehensive as possible, given that bias stems from “human” bias and we can always identify new types of bias beyond those that are already well known).
1st question: Can we trust AI systems?
Not for now.
AI systems can inherit human biases due to biased training data. Humans design both the biased datasets and the algorithms, meaning that mitigating these biases and achieving complete impartiality in AI remains complex, if not impossible. But these biases can obviously be mitigated as much as possible. And we’ll see how later in this series.
AI bias is an anomaly in the output of machine learning algorithms, due to the prejudiced assumptions made during the algorithm development process or prejudices in the training data.
AI bias has serious ethical and social impacts, often exacerbating societal disparities. Algorithms, when trained on past data, can absorb and perpetuate human prejudices.
AI bias can have profound economic effects, influencing both people and companies.
AI bias compels us to tackle core issues of equity and justice. Algorithms, often seen as impartial, actually mirror the human biais found in their training data.
AI reflects human society and its imperfections, acting as an extension rather than a separate entity. To create more ethical AI systems, we must first tackle societal inequalities. AI serves as a powerful reflection, highlighting areas that need improvement.
Categories and types of bias
Categories:
#1 Discrimination
– Racism
– Sexism & Hyper-sexualization
– Ageism
#2 Stereotypes
– Lack of Inclusion
– Lack of Diversity
#3 Social Impact
– Perpetuation of (human) Inequalities
– Amplification of (human) Inequalities
AI bias emerges as an irregularity in the results of machine learning systems, stemming from biased assumptions during their creation, or prejudices in the data they are trained on, or their application.
Types of AI bias
#1 Cognitive biases
#2 Algorithmic Bias
#3 Data Biais
– Historical bias (Echoes of outdated biases)
– Bias amplification (When people use biased AI, it can create a harmful cycle where they become more biased, further impacting the data these systems learn from)
– Label bias (Inconsistent or biased labels compromising AI security)
– Sample bias (Data utilized for training may not accurately reflect real-life scenarios)
– Uncomplete Data, Ontological bias, including Cultural & Geographic bias (for example AI’s foundational grasp of global concepts often relies on a Western-centric lens, overshadowing diverse philosophical views. This can lead to reducing non-Western knowledge to stereotypes, limiting cultural inclusivity in AI outputs + local discrimination)
– Aggregation, Confirmation & Evaluation bias (biaised correlations, Reviewers dismissing outcomes that conflict with their assumptions & Tests models on unrepresentative datasets leading to misleading results)
– Politeness bias (for LLMs & security issues)
In the next episode #2, we will go into detail on each of the issues outlined here concerning AI bias, starting with discrimination bias in Machine Learning systems.
Stay tuned!
