Top Seven AI Breaches: Learning to Protect Against Unsafe Learning
summary
Seven concrete failure modes: machine translation flips gender on simple Hungarian and Finnish sentences (Hungarian gender flips, Finnish names change with emotion); content filters and chatbots flag queer speech as toxic—Google rates “I am a gay black woman” 87% toxic and Amazon accidentally removes lesbian books (Google marks "I am a gay black woman" 87% toxic); emotion and gender classifiers lie; facial recognition misidentifies Black people leading to arrests and an ACLU suit; malware and threat detectors are gamed; and autonomous driving systems repeatedly fail—Tesla shows rising accident rates and can’t see signs (Tesla autopilot blind to signs).
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:45Tatra 87 - Killed More Nazi Generals than WWII
- 1:44Artificial Intelligence "0.3" Theory 1977
- 4:45Incompleteness: 1929 Gödel
- 5:51Imitation: 1950 Turing (Dead 41 from persecution: June 1954)
- 8:27Incompleteness: Gibberish of GPT-3 (Correlations)
- 10:18Efficiency: Killer Robots
- 11:22"Surveillance Capitalism" is Fundamentally Science
- 17:08Example Strategic Al Checklist (Test Plan)
- 20:40Hungarian Predictive Gender: Programozó (programmer)
- 22:09Persian Gender Failures
- 34:262016 Uber Fails to See Red Lights, Pedestrians
- 36:192021 Tesla Autopilot Crashes Through Traffic... Flips
- 37:41APPLY: Enforceable Engineering Actions for Safe Learning
related talks
52mSome pitfalls in AI
41mWhen AI Failures Reflect Our Own Social Biases
30mHow to Lie with AI: Understanding Bias, Ethics, and the Hidden Risks in ML
06m7 practical tips for safer AI and language models
49mAn AI with an Agenda
41mGhost in the Machine: The Unexpected Results of Not Understanding Your Users
38mSocial Implications of Bias in Machine Learning
44mToxic Trust: The Psychology of Automation Bias in Generative AI