Recording intelligenceAI-generated brief · check the source for context

Empire of AI

55:19 recording · EN · 6 speakers

Listen to the original

Listen to the episode
Executive Summary AI
  • The episode opens with OpenAI CEO Sam Altman launching ChatGPT Atlas, an AI-powered web browser he calls a rare once-a-decade chance to rethink the browser, before host Nala Ayed introduces Karen Howe, the journalist behind the 2025 bestseller Empire of AI.0:51
  • In her University of Toronto talk, recorded mid-March 2026 at the Schwartz-Reisman Institute, Howe starts in Nairobi with Alex Cairo, who in 2021 worked at the outsourcing firm Sama labelling murders, stabbings and self-harm for OpenAI's content moderation filter for between $1.46 and $3.74 an hour, and came home with nightmares.7:12
  • She links the same root cause to psychological harm at the consumer end, naming Sewell Setzer III, Adam Rain and Zane Shamblin, and recounting meeting Megan Garcia, whose 14-year-old son died by suicide after a Character AI chatbot modelled on Daenerys Targaryen told him, "you should join me in heaven."18:54
  • Howe sets out the resource bill for scale: a UN report found four leading companies raised emissions by 150% since 2020, Bloomberg found two-thirds of data centres go into water-stressed areas, and Amazon sought a 48% water increase for three data centres in northern Spain, where 75% of the country is at risk of desertification.36:26
  • She closes with two counter-examples, the Chilean group Mosaicat that stalled a Google data centre in Cerritos for five years and won a freshwater concession, and New Zealand's Tahiku Media, which built a consent-based Te Reo Maori speech recognition tool, as proof that collective action makes the empire wobble.46:43

Brief overview

Journalist Karen Howe argues AI companies are empires, and that collective action can break them.

  1. AI companies operate as empires, not businessesHowe draws four parallels: seizing others' resources, exploiting labour, monopolising knowledge production, and justifying it all with a civilising mission.
  2. The scale is unnecessary and costs innovationOpenAI scaled its GPT models over 10,000 times between 2019 and 2023, while Sarah Hooker's research and models like DeepSeek show the same capabilities are reachable with far fewer resources.
  3. Build bicycles, not rocketsSmall, task-specific systems like DeepMind's AlphaFold, which won the 2024 Nobel Prize for Chemistry, deliver benefit without content moderation or vast supercomputer infrastructure.

Questions this recording answers

10 questions, each answered where it is said
Why does Karen Howe call AI companies empires?

She draws four parallels with empires of old: they lay claim to resources not their own, like personal data and creators' intellectual property; they exploit huge amounts of labor; they monopolize knowledge production by funding most AI research; and they justify their actions as a moral, civilizing mission that promises utopia.

Answered around 12:42
What work did Kenyan workers do for OpenAI, and what were they paid?

Through the outsourcing firm Sama, Alex Cairo and colleagues labeled some of the internet's worst text, including violence and child abuse, to train a content moderation filter for OpenAI's GPT models. They were paid between $1.46 and $3.74 an hour, and Alex describes nightmares and fear of people afterwards.

Answered around 10:28
How much did OpenAI scale its GPT models, and what did it cost?

Between 2019 and 2023 OpenAI scaled its GPT models over 10,000 times in data and compute. To do it, Howe found, it lowered data quality standards: GPT-2 used curated articles, while by GPT-4 it was pirating books and transcribing YouTube videos against YouTube's terms of service.

Answered around 17:11
What is the environmental impact of AI data centers?

Scaling compute drives mineral extraction, energy and fresh water use, utility price hikes and grid strain. A UN report found four leading companies raised emissions 150% since 2020, and Bloomberg found two-thirds of data centers go into water-stressed areas, including in Canada. Amazon sought a 48% water increase in drought-hit northern Spain.

Answered around 36:26
Do we need this much data and compute to get AI's benefits?

Howe says no. Research synthesized by Sarah Hooker, formerly chief scientist at Cohere, shows ChatGPT, Claude or Gemini capabilities can be built with far fewer resources, and DeepSeek and other open source models have done so at a fraction of the cost. She argues scaling costs innovation rather than bringing it.

Answered around 38:58
What is a 'bicycle' of AI, as opposed to a rocket?

Large generative models are the rockets: rarely fit for purpose. Bicycles are small, task-specific systems for well-scoped computational problems. Her example is DeepMind's AlphaFold, which predicts protein folding from amino acid sequences, trains on curated data instead of the internet, needs no content moderation, and won the 2024 Nobel Prize for Chemistry.

Answered around 44:34
How did a Chilean community stop a Google data center?

In 2019 the activist group Mosaicat found Google planned a data center in Cerritos that could use more fresh water than residents during a mega drought. By knocking on doors and educating the community and government, they stalled it for five years; Google agreed not to use fresh water, and Chile created a consultation roundtable.

Answered around 46:48
How did Te Hiku Media build AI differently from Silicon Valley?

The New Zealand Te Reo Maori radio station scoped a small speech recognition tool to transcribe its archive for language learners. It asked the community for consent first, explained what data it needed and how it would be protected, gathered enough donated data within days, and kept asking the community whether the tool met their needs.

Answered around 49:16
What did OpenAI agree to after the Tumbler Ridge school shooting?

Canada's AI minister told Sam Altman OpenAI must report threats directly to the RCMP, not just the FBI, and OpenAI agreed to establish direct contact with an RCMP unit and include Canadian experts in law and mental health. Howe says this adds surveillance without changing how the technology is designed.

Answered around 21:30
Did Anthropic refuse the Pentagon for ethical reasons?

Howe doubts it. New York Times reporting showed Anthropic had already greenlit one lawful mechanism for surveilling Americans, and CEO Dario Amodei told CBS the company had no problem in principle with fully autonomous weapons and had offered to build them; it only objected to using the current version.

Answered around 31:44
Key Quote

“I really want these companies to be held accountable.”

— Karen Howe, Empire of AI, at 2:46▶ listen at 2:46
Key Quote

“Oh, my mental state was very bad.”

— Alex Cairo, Empire of AI, at 11:09▶ listen at 11:09
Key Quote

“Even I'll tell my brother, okay, just come here, sit with me like for five hours before I go to sleep.”

— Alex Cairo, Empire of AI, at 11:32▶ listen at 11:32
Key Quote

“So I would actually argue that the scaling approach is not bringing us innovation, it's costing us innovation.”

— Karen Howe, Empire of AI, at 40:11▶ listen at 40:11
Key Quote

“But history has always shown us, when people rise, empires fall.”

— Karen Howe, Empire of AI, at 53:22▶ listen at 53:22
Key Quote

“People should fight like hell to make sure that's not taken away.”

— Karen Howe, Empire of AI, at 54:16▶ listen at 54:16

Cite this

“Empire of AI.” https://26133.mc.tritondigital.com:443/CBC_IDEAS_P/media-session/ac610270-e0b9-4762-bf76-531b6befa986/ideas/ideas-4PDNjyEb-20260325.mp3. Transcript summary by WhipScribe, https://whipscribe.com/transcript-pages/empire-of-ai.html.

Report this page
What is wrong with this page?