Chapter 6 Further Reading: Computer Vision

Annotated recommendations for deeper exploration of computer vision, facial recognition, deepfakes, and their societal implications.

Foundational Understanding

"AI and the Future of Seeing" — MIT Technology Review An accessible overview of how computer vision has evolved from laboratory curiosity to pervasive technology. Good starting point for readers who want more technical depth without a computer science background.

Fei-Fei Li, "How We're Teaching Computers to Understand Pictures" — TED Talk (2015) A 17-minute talk by one of the founders of ImageNet, explaining how training AI on millions of images enabled breakthroughs in visual recognition. Li's presentation is clear, personal, and non-technical — ideal for understanding the "ImageNet moment" discussed in Section 6.2.

"The Humble Pixel: A Deep Dive into Digital Images" — 3Blue1Brown (YouTube) For readers who want a visual, mathematical explanation of how images are represented as numbers and how convolutions work, this video series provides elegant visualizations that build intuition without requiring advanced math.

Facial Recognition and Civil Liberties

Joy Buolamwini and Timnit Gebru, "Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification" (2018) The landmark study discussed in Section 6.4. Available through the MIT Media Lab. Essential reading for understanding how demographic disparities in training data translate to accuracy disparities in deployed systems.

Kashmir Hill, Your Face Belongs to Us: A Secretive Startup's Quest to End Privacy as We Know It (2023) A deeply reported investigation into Clearview AI, the company that scraped billions of photos from the internet to build a facial recognition database. Hill, a New York Times reporter, traces the technology's development, its adoption by law enforcement, and its implications for privacy. Highly readable and well-sourced.

Clare Garvie, Alvaro Bedoya, and Jonathan Frankle, "The Perpetual Line-Up: Unregulated Police Face Recognition in America" — Georgetown Law Center on Privacy & Technology (2016) A comprehensive report on law enforcement use of facial recognition in the United States. Documents the scale of deployment, the lack of regulation, and the civil liberties implications. Foundational for understanding the regulatory gap discussed in Case Study 2.

Evan Selinger and Woodrow Hartzog, "The Inconsentability of Facial Surveillance" — Loyola University Chicago Law Journal (2019) An academic article arguing that meaningful consent to facial recognition in public spaces is impossible, and that regulation rather than consent is the appropriate framework. Accessible and well-argued for a law journal article.

Deepfakes and Synthetic Media

Nina Schick, Deepfakes: The Coming Infocalypse (2020) A comprehensive overview of the deepfake landscape, from non-consensual pornography to political manipulation to fraud. Schick writes for a general audience and provides both technical context and policy analysis.

"Deepfakes Are Getting Better, But Are They Getting More Dangerous?" — MIT Technology Review (2024) An updated assessment of where deepfake technology stands, how detection efforts are progressing, and whether provenance-based approaches (like C2PA) offer a viable long-term solution.

Sam Gregory, "Deepfakes and Cheap Fakes: The Manipulation of Audio and Visual Evidence" — WITNESS (2020) A report from the human rights organization WITNESS, focused on how manipulated media affects marginalized communities and human rights documentation. Offers a global perspective often missing from U.S.-centric discussions.

Autonomous Vehicles and Vision Failures

National Transportation Safety Board, "Collision Between a Car Operating with Automated Vehicle Control Systems and a Pedestrian, Tempe, Arizona" — NTSB Report HAR-19/03 (2019) The official investigation report on the 2018 Uber crash discussed in Case Study 1. Detailed, technical, and publicly available. Reading the executive summary provides a clear picture of how multiple system failures combined to produce a fatal outcome.

Philip Koopman, How Safe Is Safe Enough? Measuring and Predicting Autonomous Vehicle Safety (2022) A technical but accessible book by a Carnegie Mellon researcher examining how to evaluate the safety of autonomous vehicles. Koopman argues that the industry's approach to safety validation is fundamentally inadequate. Good for readers interested in the "long tail" problem discussed in Section 6.6.

Technical Deep Dives (Optional, More Advanced)

"A Beginner's Guide to Convolutional Neural Networks" — Analytics Vidhya A step-by-step technical introduction to CNNs with diagrams and code examples. Suitable for readers who want to understand the architecture in more detail without taking a full machine learning course.

Ian Goodfellow, Yoshua Bengio, and Aaron Courville, Deep Learning (2016), Chapter 9: Convolutional Networks The standard graduate textbook treatment of CNNs. Heavy on mathematics but comprehensive. For readers with a strong quantitative background who want the full technical picture.

Christian Szegedy et al., "Intriguing Properties of Neural Networks" (2013) The original research paper on adversarial examples. Technical but historically important — this is the paper that first demonstrated that imperceptible perturbations could fool neural networks with high confidence.

Multimedia Resources

"Coded Bias" — Documentary film directed by Shalini Kantayya (2020) Follows Joy Buolamwini's journey from discovering facial recognition bias to advocating for regulation. Available on major streaming platforms. Powerful, accessible, and directly relevant to this chapter's themes.

"In Event of Moon Disaster" — MIT Center for Advanced Virtuality (2020) A deepfake demonstration project that created a convincing video of President Nixon delivering a speech announcing the death of the Apollo 11 astronauts — a speech that was written as a contingency but never delivered. The project was created to educate the public about deepfake technology and comes with educational materials.

For Your AI Audit Report

If your chosen AI system uses computer vision, the following resources may be particularly useful for your audit:

• The NIST Face Recognition Vendor Test (FRVT) reports provide ongoing independent evaluations of facial recognition system accuracy across demographics.
• The Object Detection on COCO benchmark (paperswithcode.com) tracks the current state of the art in object detection, giving you a sense of what's technically possible.
• Model Cards for Machine Learning Models (Mitchell et al., 2019) provides a framework for documenting the capabilities and limitations of AI systems — directly applicable to your audit report.