Empire of Evidence: Testing Karen Hao's 9 AI Claims Against Governance Infrastructure
The Other AI: Audio Briefings on Augmented Intelligence and AI GovernanceShow Notes
Investigative journalist Karen Hao spent eight years and conducted over 300 interviews examining the AI industry. Her book Empire of AI won the National Book Critics Circle Award for Nonfiction, reached the New York Times bestseller list, and earned her a place on TIME's TIME100 AI list. In her March 2026 interview on The Diary of a CEO, she made nine specific claims about how major AI companies operate. This episode is an audio examination of all nine claims, testing each against available evidence and mapping the strongest findings to published open-source AI governance architecture. Five claims held under scrutiny: Knowledge Production Control. AI companies fund the scientists who study their own systems and censor researchers who produce inconvenient findings. Google fired AI ethics co-leads Dr. Timnit Gebru and Margaret Mitchell. Congress cited Hao's reporting five times. AGI Definition Shifting. OpenAI describes artificial general intelligence differently depending on the audience. The OpenAI Charter, the Microsoft contractual threshold, the Congressional framing, and the consumer marketing describe fundamentally incompatible systems. Revenue-Driven Capability Selection. Internal documents show companies advance capabilities based on which industries pay the most, not on scientific priority. Data Annotation Labor Conditions. The annotation industry absorbs displaced workers and drives conditions downward through structural competition on speed and cost. Environmental Externalities. AI data centers consume massive resources. The Memphis Colossus facility runs on 35 gas turbines. Hao acknowledged a 1,000x unit error on one Chilean water figure, but the broader environmental reporting remains substantiated. Four claims required challenge: The Empire Analogy works as a structural lens but breaks down at literal comparison with colonial empires that enforced power through military violence. Self-Driving Car Predictions. Waymo reports 92% fewer serious-injury crashes, but those miles are in five US cities under mapped conditions. New York City rush hour, Bangkok traffic, and unpaved mountain roads in Peru would produce fundamentally different data. Bicycles vs. Rockets. AlphaFold was built by Google DeepMind on Google's TPU clusters. The "bicycle" came from the same corporate infrastructure Hao critiques. Intelligence Scaling. The mechanism debate is real, but measurable capability improvements in coding, reasoning, and planning are not hypothetical. The examination maps findings to AI Provider Plurality, the Economic Override Pattern, the Constitutional Wall Principle, and Multi-Provider Divergence through HAIA-CAIPR. All are published working concepts, not production-validated systems. Other governance approaches may address the same structural problems. Full white paper with complete sources and APA references: https://basilpuglisi.com/empire-of-evidence-testing-karen-hao-claims-governance-infrastructure/ Karen Hao's Diary of a CEO interview: https://www.youtube.com/watch?v=Cn8HBj8QAbk Open-source governance frameworks: https://github.com/basilpuglisi/HAIA AI Content Disclosure: This audio was generated by Google NotebookLM from the published article. NotebookLM audio cannot be edited after generation. The guidance instructions provided beforehand are the only editorial control available. Proper noun pronunciation varies in AI-generated audio. #AIassisted using the HAIA Ecosystem
More from The Other AI: Audio Briefings on Augmented Intelligence and AI Governance
Every organization that puts AI to work makes the same choice, whether it names the choice or not. It can let agents carry the work with machines checking machines, or keep human agency in the work, with a named person holding authority at checkpoints and answering for what comes out. This Deep Dive on Basil C. Puglisi's article walks through Responsible AI as the factory and AI Governance as the handmade product, where each one belongs, and the rule each one carries. It covers why presence is not authority, the inspection mark a handmade product can carry, and cases that run from an airline chatbot to Australia's Robodebt scheme. It also weighs what research on disclosure and labels has found and what none of it has tested, along with the procurement question any officer can answer. It then covers the EU AI Act and how an SEC enforcement action, Colorado's new law, and New York's court rule reach the same line by other routes. https://basilpuglisi.com/agents-agency-transparency/ Podcast: The Other AI: Audio Briefings on Augmented Intelligence and AI Governance Spotify: https://open.spotify.com/show/033dvhzMIcWLdY7IUgsu7F Apple Podcasts: https://podcasts.apple.com/us/podcast/id1896506152 Amazon Music: https://music.amazon.com/podcasts/923d1a79-533f-4623-bae3-e2ba83453dfb YouTube playlist: https://www.youtube.com/playlist?list=PLchpU2bIYoEEBh2hdY-BVP9ckyTPiHOAQ Generated using NotebookLM. Content may have inaccuracies. For full detail refer to the original paper, document, or article. These are AI generated under NotebookLM as audio overviews not polished products. #AIassisted using the HAIA Ecosystem
In January 2026, one vocabulary covered both automation and human control, so a system that only watched itself could be sold as governed, and a person who only watched a screen could be counted as a governor. This episode follows the path from that definition problem to running code, from Checkpoint-Based Governance and the three layers of Ethical AI, Responsible AI, and AI Governance to the formal definition of who answers. The hosts walk the three HAIA Agents models. Agent Model 1, released in Alpha as the CBG Model 1 Runner, is Responsible AI written as code: a held-out Navigator, parallel review on three or more AI platforms, verification against pages the runner fetched, scoring on the author's own instruments, and a gate it cannot move, with a handoff to the human when a piece cannot pass. Agent Model 2, released as the CBG Gateway, is the hybrid that runs through to AI Governance at a signed human checkpoint after every role. The episode closes on Agent Model 3, Manual Human AI Governance, which built both releases, and on what the Alpha does not claim: neither program has been deployed, neither has called a live AI provider, and a signature establishes what was signed, not what was read. Read the full article: https://basilpuglisi.com/haia-agents-running-code/ Podcast: The Other AI: Audio Briefings on Augmented Intelligence and AI Governance Spotify: https://open.spotify.com/show/033dvhzMIcWLdY7IUgsu7F Apple Podcasts: https://podcasts.apple.com/us/podcast/id1896506152 Amazon Music: https://music.amazon.com/podcasts/923d1a79-533f-4623-bae3-e2ba83453dfb YouTube playlist: https://www.youtube.com/playlist?list=PLchpU2bIYoEEBh2hdY-BVP9ckyTPiHOAQ Generated using NotebookLM. Content may have inaccuracies. For full detail refer to the original paper, document, or article. These are AI generated under NotebookLM as audio overviews not polished products. #AIassisted using the HAIA Ecosystem
Tesla now calls itself a physical AI company, and its own record makes it the clearest live test of what autonomous really means. This deep dive follows the Optimus humanoid robot from the 2024 We, Robot demonstrations, where humans remotely operated much of the interaction, to Elon Musk's January 2026 statement that Optimus was not contributing to manufacturing work in any material way, and on to the new Fremont lines that feed an Optimus Academy for training data. It then follows the Robotaxi from its June 2025 launch with a passenger-seat safety monitor to the March 2026 letter in which Tesla told Senator Edward Markey how remote operators can redirect, command, and, as a last resort, control its vehicles. The episode covers NHTSA's audit of the Cybercab self-certification, the engineering analysis of Tesla's camera-based driving system, the capital behind the bet, and the investors and researchers who read the same record in opposite ways. The takeaway is that one word, autonomous, covers six operating modes, and every claim needs its mode, its date, its source, and a named human who decides. Read the full article: https://basilpuglisi.com/physical-ai-platform-shift/ Podcast: The Other AI: Audio Briefings on Augmented Intelligence and AI Governance Spotify: https://open.spotify.com/show/033dvhzMIcWLdY7IUgsu7F Apple Podcasts: https://podcasts.apple.com/us/podcast/id1896506152 Amazon Music: https://music.amazon.com/podcasts/923d1a79-533f-4623-bae3-e2ba83453dfb YouTube playlist: https://www.youtube.com/playlist?list=PLchpU2bIYoEEBh2hdY-BVP9ckyTPiHOAQ Generated using NotebookLM. Content may have inaccuracies. For full detail refer to the original paper, document, or article. These are AI generated under NotebookLM as audio overviews not polished products. #AIassisted using the HAIA Ecosystem
Chat made AI legible to every board. Robots that see, plan, and act now make it consequential in warehouses, labs, hospitals, and homes, where a wrong output moves mass instead of words. This episode works through the case that physical AI is the next platform shift, and through the reasons the deployment ledger still trails it. It covers the investor thesis that robot learning, autonomous science, and new human-machine interfaces are the next compounding domains beyond language and code. It weighs that thesis against independent economic evidence that robots can perform most physical work tasks in at least some settings yet are cost-competitive for only a small fraction of job tasks today. The hosts walk the model build path from NVIDIA Isaac GR00T, Google DeepMind's Gemini Robotics, and Physical Intelligence's π₀ through their 2026 successors, and why vendor results stay condition-specific. They work through the reliability arithmetic that decides whether a multi-step robot task survives production, and what a real deployment's hardware ledger reveals. They also cover why classical automation keeps the burden of proof on the learning system, and why every learned cell needs a tested fallback. The episode closes on governance. A weekly evidence cadence ends with a named person who decides to accept, modify, or reject, and who answers for the result. Model capability, deployment economics, and functional safety stay separate, and the checkpoint sits above engineered safety rather than replacing it. Read the full article: https://basilpuglisi.com/physical-ai-platform-shift/ Podcast: The Other AI: Audio Briefings on Augmented Intelligence and AI Governance Spotify: https://open.spotify.com/show/033dvhzMIcWLdY7IUgsu7F Apple Podcasts: https://podcasts.apple.com/us/podcast/id1896506152 Amazon Music: https://music.amazon.com/podcasts/923d1a79-533f-4623-bae3-e2ba83453dfb YouTube playlist: https://www.youtube.com/playlist?list=PLchpU2bIYoEEBh2hdY-BVP9ckyTPiHOAQ Generated using NotebookLM. Content may have inaccuracies. For full detail refer to the original paper, document, or article. These are AI generated under NotebookLM as audio overviews not polished products. #AIassisted using the HAIA Ecosystem
This episode is a story, told for entertainment. Teresa writes grant proposals for a food bank with a machine open on the second screen. Her brother-in-law Dale builds the air systems that keep hospitals breathing and has never typed a prompt. Over one backyard argument he asks the question every AI user hears, and her answer runs through Tron in 1982, Phil Tippett and the dinosaurs of Jurassic Park in 1993, and the two adaptations of The Shining, before it lands on the three old reasons people refuse a new tool. The history in the story is real and sourced. The people at the table are invented. Read the full story at https://basilpuglisi.com/intelligent-why-use-ai/ Podcast: The Other AI: Audio Briefings on Augmented Intelligence and AI Governance Spotify: https://open.spotify.com/show/033dvhzMIcWLdY7IUgsu7F Apple Podcasts: https://podcasts.apple.com/us/podcast/id1896506152 Amazon Music: https://music.amazon.com/podcasts/923d1a79-533f-4623-bae3-e2ba83453dfb YouTube playlist: https://www.youtube.com/playlist?list=PLchpU2bIYoEEBh2hdY-BVP9ckyTPiHOAQ Generated using NotebookLM. Content may have inaccuracies. For full detail refer to the original paper, document, or article. These are AI generated under NotebookLM as audio overviews not polished products. #AIassisted using the HAIA Ecosystem