Chapter 20 – When things go wrong

Contents

20.1  Overview
20.2  Introduction
20.3  Wrong on Purpose?
20.3.1  Intentional Bad Use
20.3.2  Unintentional Problems
20.4  General Strategies
20.4.1  Transparency and Trust
20.4.2  Algorithmic Accountability
20.4.3  Levels of Opacity
20.5  Sources of Algorithmic Bias
20.5.1  What Is Bias?
20.5.2  Stages in Machine Learning
20.5.3  Bias in the Training Data
20.5.4  Bias in the Objective Function
20.5.5  Bias in the Accurate Result
20.5.6  Proxy Measures
20.5.7  Input Feature Choice
20.5.8  Bias and Human Reasoning
20.5.9  Avoiding Bias
20.6  Privacy
20.6.1  Anonymisation
20.6.2  Obfuscation
20.6.3  Aggregation
20.6.4  Adversarial Privacy
20.6.5  Federated Learning
20.7  Communication, Information and Misinformation
20.7.1  Social Media
20.7.2  Deliberate Misinformation
20.7.3  Filter Bubbles
20.7.4  Poor Information
20.8  Summary

Glossary items referenced in this chapter

accountability, accuracy, adversarial attacks, adversarial learning, aggregation for privacy, algorithmic accountability, anonymisation, anti-discrimination laws, automated decision, autonomous car, autonomous vehicles, autonomous weapons, backpropagation, base rate, bias, big data, black-box machine learning, Cambridge Analytica scandal, centroid, chatbot, choice of features, clustering, COMPAS, cyberattack, cyberwarfare, cyberweapons, de-bias, deep neural network, deliberate misinformation, delta, denial of service (DoS), echo chambers, ethics, expert system, explainable AI, explanation, Facebook, facial recognition, fake news, false negative, false positive, federated learning, fitness function, GDPR, generative AI, Google, Google search, GPT-4, human bias, human labelling, human-in-the-loop, ID3, identity theft, image processing, image recognition, k-means algorithm, labelling, machine learning, Microsoft, Microsoft Tay, misinformation, misinformation detection, natural language algorithms, natural language processing, network analysis, neural network, obfuscation, OpenAI, optimal classification, overfitting, PageRank, perturbation techniques, pragmatic, privacy, privacy preserving algorithms, protected characteristic, proxy indicator, pseudonymisation, search engine, search engine personalisation, semi-autonomous car, simulated data, social media, sources of bias, statistical bias, statistics, Stuxnet, symbolic systems, threshold, transparency, trust, Twitter bot, unintended bias, unique identifier, user interface, visualisation, web search