Artificial Intelligence Ethicology (WILL A.I. CRASH OUT?) with Abeba Birhane

Episode Summary

Executive Summary: The episode argues that AI is not a sentient force to fear, but a human-made system shaped by data, incentives, and power. Dr. Ababa Burhani explains how embodied cognition informs AI ethics, why generative AI is often unreliable and biased, how corporations exploit data and hype, and why the real risks are surveillance, labor displacement, environmental cost, and misuse by governments and companies.

Main Topics: What AI is, and what it is not (Priority: 5/5): The conversation distinguishes traditional machine learning, NLP, and generative AI, emphasizing that current systems predict and generate patterns rather than possess intention, understanding, or a soul. Embodied cognition and AI ethics (Priority: 4/5): Burhani explains embodied cognitive science: cognition is shaped by bodies, environments, and communities, not just isolated brains. This frames AI as a socially embedded human artifact rather than an independent mind. Data extraction, copyright, and bias (Priority: 5/5): The episode details how AI companies scrape web data without meaningful consent, how artists and writers are used to train models, and how racist, sexist, and pornographic content can be embedded in training sets and outputs. Public hype versus technical reality (Priority: 4/5): A major theme is that AI companies overstate capability. Burhani argues that hallucinations, poor reliability, and model collapse are real limitations, and that public education is needed to counter PR-driven narratives. Power, surveillance, and militarization (Priority: 5/5): The episode stresses that the danger is not AI itself but powerful actors using AI for surveillance, policing, border enforcement, and military purposes, often under the banner of national security. Labor, education, healthcare, and the environment (Priority: 4/5): Listeners’ questions focus on job displacement, school use, healthcare tools, and energy/water consumption. Burhani says AI can help in some domains, but only with human oversight, transparency, and better regulation. Practical resistance and alternatives (Priority: 3/5): The discussion suggests partial resistance through Signal, open-source tools, data-poisoning defenses like Nightshade and tar pits, and legal/regulatory action, though abstention is often difficult because AI is embedded in infrastructure.

Key Arguments: AI systems are not conscious or intentional; they do not care, desire, or fear anything, and should not be treated as sentient beings. The major ethical issue is not an AI apocalypse, but how humans and institutions deploy AI for profit, surveillance, and control. Training data is central: if models are trained on biased, toxic, or unconsented data, those harms will be reproduced and amplified in outputs. Generative AI is often unreliable, with hallucinations, made-up citations, and degraded performance that users need humans to verify. AI companies function as commercial entities and generally resist transparency about their datasets, making external auditing difficult. The internet is a deeply skewed training source, so AI often reproduces historical racism, sexism, xenophobia, and pornographic distortions. Data center power use and water cooling create substantial environmental costs, especially for generative AI. AI can still be beneficial in areas like disaster mapping, soil monitoring, and some healthcare applications, but only if patient/public needs outrank profit motives. Students may get short-term help from chatbots, but long-term learning and critical thinking can be harmed when AI substitutes for engagement. The most realistic path forward is regulation, transparency, public education, and human oversight rather than techno-optimism or fear of sentient machines.

Data Points: ChatGPT hallucination rate in one test: as high as 79% - Referenced while discussing newer AI systems producing incorrect information more often. Energy use vs. Google search: about 5 to 10 times more electricity - Google AI and Burhani both cited estimates comparing ChatGPT-like queries to a standard search. Energy use vs. basic Google search: up to 30 times more energy - ChatGPT’s own response about version 4's electricity use. Irish household energy comparison: compute resources required equal or exceed total energy needed to run Irish households - Burhani described Ireland’s data centers as major power hogs. Data center power demand increase: 160% - A Goldman Sachs report cited by Google AI projected AI could drive this rise in data center power demand. Data set audited: 80 Million Tiny Images - Burhani mentioned an audit that found thousands of problematic labels and images in this MIT-held dataset. AI assistants knowledge cutoff: around 2021 or 2022 - Used to explain why models cannot reliably answer with real-time awareness. Google Walkout year: 2018 - Mentioned in the context of Meredith Whitaker’s organizing work before leading Signal. Age of core AI foundations: 1950s–1980s - Burhani noted many foundational AI ideas predate the recent generative boom. Student control study size: over 3,000 students - A study discussed on chatbot use in math learning and later performance decline. Job-related data extraction timeline: AI training and analysis work often outsourced to developing countries - Burhani noted data labor in Kenya, Nigeria, Ethiopia, India, and related regions. Lead cause of mortality for U.S. teens and children: firearms - Referenced as an analogy about the false neutrality of technology and the limits of “people, not tools” arguments.

Pivotal Quotes: "There is no intention. There is no understanding. There is no soul in these machines." — Dr. Ababa Burhani: Explaining why chatbots should not be treated as sentient or morally accountable beings. "The real worry is powerful people using AI to do terrible things." — Dr. Ababa Burhani: Clarifying that the main risk is human misuse of AI for surveillance, war, and rights suppression. "You always need humans for AI to function and operate as it's supposed to." — Dr. Ababa Burhani: On labor, oversight, and why AI remains dependent on human data work, curation, and verification.

Implications: Listeners are urged to treat AI as a human-built system with real harms: bias, labor exploitation, surveillance, and environmental cost. The future depends less on sentient-machine fears and more on regulation, transparency, and refusing hype.