Visual summary of operating lessons from Karen Hao.

Lessons from Karen Hao

Karen Hao is an investigative journalist and the author of Empire of AI. Drawing on reporting inside AI companies and in communities affected by their supply chains, she argues that the current race to build very large AI systems concentrates power while imposing labor, environmental, and democratic costs. Her critique targets that scale-at-all-costs approach, not every use of AI. — UNSW Interview.

Part 1: The AI Empire Metaphor

  1. Think of AI firms as political actors: Hao argues that the largest AI companies should be examined not only as product businesses but as institutions accumulating economic and political power on a scale that can affect public decisions. — UNSW Interview.
  2. Treat data as an extracted resource: Hao compares the collection of people’s data and creative work for model training with older patterns of claiming resources without giving affected communities comparable control or benefit. — UNSW Interview.
  3. Follow the infrastructure footprint: Hao argues that large-scale AI changes physical as well as digital environments: its data centers, energy demand, water use, and supply chains have consequences in the communities where they are placed. — Democracy Now Interview.
  4. Scale is a choice, not the definition of progress: Hao distinguishes the resource-intensive race for very large models from smaller, specialized AI systems that can address specific public needs with less infrastructure. — AI Now Interview.
  5. Watch the concentration of research: Hao argues that wealthy AI firms can draw researchers and computing resources away from independent institutions, narrowing who studies the technology and who can challenge claims about its limits. — UNSW Interview.
  6. Separate the technology from the dominant business model: Hao’s criticism is directed at Silicon Valley’s scale-at-all-costs development model. She points to smaller, task-specific AI as evidence that the benefits of AI need not depend on the same degree of extraction or concentration. — Democracy Now Interview.
  7. Track where costs and gains fall: Hao’s colonialism comparison focuses on firms sourcing labor, data, minerals, and infrastructure from less powerful places while much of the economic value accrues elsewhere. — AI Now Interview.
  8. Defend communities’ ability to choose: Hao links the empire metaphor to democracy: people lose agency when companies make decisions about their data, land, water, or labor without meaningful local control. She points to community resistance as a way to regain that voice. — Democracy Now Extended Interview.
  9. Build AI around local needs: Hao argues for smaller, specialized systems designed with the communities that use or are affected by them, rather than assuming every country should replicate Silicon Valley’s largest-model race. — AI Now Interview.

Part 2: OpenAI and the Race for AGI

  1. AGI shaped OpenAI’s mission: Hao reported that OpenAI defined its early mission around reaching artificial general intelligence and ensuring that such a system would be developed safely and widely beneficially. Those were the organization’s stated ambitions, not achieved outcomes. — MIT Technology Review Investigation.
  2. The funding model changed: Hao traced how the computational demands of OpenAI’s chosen research path led it to create a capped-profit arm in 2019 under a nonprofit board, raising questions about how that structure would preserve its founding commitments. — MIT Technology Review Investigation.
  3. Define what success demands: Hao’s reporting links OpenAI’s race to develop AGI with a compute-heavy strategy, a need to raise capital, and growing secrecy. She presents these as strategic choices with trade-offs, not inevitable requirements of all AI research. — MIT Technology Review Investigation.
  4. Safety work needs more than a promise: Hao documented OpenAI teams working on values, interpretability, and learning without harm while also noting the uncertainty around what AGI would be and whether those methods would generalize. — MIT Technology Review Investigation.
  5. Treat safety claims as testable decisions: Hao’s account of OpenAI’s GPT-2 release debate shows how safety concerns, publication norms, and outside scrutiny collided over one model. The episode calls for examining concrete release choices rather than assuming either safety or speed always wins. — MIT Technology Review Investigation.
  6. Internal debates accompanied release choices: Hao reported disagreements around the GPT-2 release decision and the lab’s move toward less open publication. The tensions concerned both safety explanations and criticism that secrecy could serve competitive interests. — MIT Technology Review Investigation.
  7. AGI remains a contested definition: Hao noted that researchers disagree about what artificial general intelligence would look like, what minimum abilities would count, and whether there is a single form of general intelligence at all. — MIT Technology Review Investigation.
  8. Commercial pressure can change research priorities: Hao reported that, after OpenAI created a capped-profit arm and took major investment, employees faced increasing pressure to produce revenue-supporting research even as leaders maintained that the mission would continue. — MIT Technology Review Investigation.
  9. Ask who is represented in AI research: Hao observed that OpenAI’s early staff lacked representation from many groups likely to be affected by its systems. She questioned how a lab could credibly promise benefits for everyone without engaging those perspectives. — MIT Technology Review Investigation.

Part 3: The Ghost Workers of the Global South

  1. Trace the workers behind the model: Hao reported that Kenyan contractors who helped make ChatGPT safer were paid roughly two to three dollars an hour while reviewing disturbing material. The wages and exposure were part of a specific OpenAI–Sama contract, not a measured rate for every AI worker. — WSJ Journal Podcast.
  2. Automation still depends on human work: The moderation system used for ChatGPT depended on people who labeled harmful examples so the model could learn to recognize them. Hao’s reporting makes that labor visible behind an apparently automated product. — WSJ Journal Podcast.
  3. Account for psychological exposure: Hao’s reporting describes Kenyan moderators repeatedly encountering graphic content while preparing training examples for OpenAI’s safety systems. The documented case is Kenya; the inherited claim about Colombia is not established here. — WSJ Journal Podcast.
  4. Follow who captures the value: Hao contrasts the labor and resources sourced from less powerful communities with the wealth and strategic advantage concentrated in the companies developing large AI systems. — AI Now Interview.
  5. Worker protections are a live demand: After describing the Kenyan contractors’ work, Hao reported that workers had begun organizing across outsourcing companies and petitioned Kenya’s parliament for stronger protections. The petition is not evidence that those protections were enacted. — WSJ Journal Podcast.
  6. Do not mistake safety automation for labor-free AI: In Hao’s ChatGPT investigation, people manually labeled violent and abusive examples used to train a safety filter. That work does not mean all model behavior is continuously corrected by humans in real time. — WSJ Journal Podcast.
  7. Make the contracting chain visible: Workers in the Kenyan moderation project were employed through the outsourcing firm Sama rather than directly by OpenAI. Hao’s reporting traced that arrangement and the conditions under which the work was done. — WSJ Journal Podcast.
  8. Test safety beyond English: Hao’s podcast investigation found that the safety filter worked more effectively in English than in Swahili in the reported ChatGPT case, revealing a concrete language-specific gap. — WSJ Journal Podcast.
  9. Workers can organize across contractors: Hao reported in 2023 that Kenyan data and moderation workers had begun building a union across several outsourcing companies and sought legislative protection for the psychological toll of the work. — WSJ Journal Podcast.

Part 4: Natural Resources and Hyperscale Computing

  1. The cloud has a physical footprint: Hao emphasizes that large AI models rely on data centers, electricity, cooling water, and hardware supply chains. Their effects depend on where and how that infrastructure is built. — The Atlantic Investigation.
  2. Examine site-specific water demand: In her Arizona reporting, Hao documented projected water demand for a Microsoft data-center campus in a dry region. The figure was a site-specific projection, not a universal amount required to train a model. — The Atlantic Investigation.
  3. Let local communities scrutinize siting: Hao described Chilean activists challenging a proposed data center’s water plans, while her Arizona reporting examined water use in a dry region. Neither case establishes that current residents lost drinking water to those projects. — Democracy Now Interview.
  4. Demand site-level environmental disclosure: Hao found that public environmental reporting did not always make the water and energy costs of individual AI data centers easy to assess. The transparency gap is narrower than a proven industry-wide pattern of hidden emissions. — The Atlantic Investigation.
  5. Plan for electricity demand: Hao’s data-center reporting describes how growing AI computing demand drives new power and cooling infrastructure. The burden on a particular grid should be evaluated locally, not assumed from a global model count. — The Atlantic Investigation.
  6. Ask what a community gives up: Hao’s Arizona field report treats data-center water demand as a local planning question involving a finite regional resource. It does not establish the inherited comparison between one model and whole neighborhoods. — The Atlantic Investigation.
  7. Test sustainability claims against local impacts: Hao distinguishes company-wide sustainability commitments from the water and electricity demanded by a specific data-center campus. Local environmental effects need their own assessment. — The Atlantic Investigation.
  8. Include minerals and hardware in the account: Hao’s extractivism framework includes the minerals and material infrastructure needed for AI hardware as well as data, labor, and data centers. Her comparison is an argument about patterns of resource control, not a claim that today’s supply chain is identical to nineteenth-century mining. — AI Now Interview.
  9. Compare computing costs by task: Hao reports that generative-AI systems can be much more resource-intensive than narrower AI approaches. An energy comparison should identify the model, task, and infrastructure rather than assume every generated image or essay has one fixed cost. — UNSW Interview.

Part 5: Algorithms, Attention, and Facebook

  1. Engagement can reward misinformation: Hao’s Facebook investigation found that ranking systems optimized for engagement could elevate divisive or misleading material even when the company had teams trying to curb it. — MIT Technology Review Investigation.
  2. Look at the business incentive: Hao traced how engagement-based ranking and growth incentives shaped which integrity projects advanced at Facebook. The harms were tied to product and business choices, not solely isolated coding errors. — MIT Technology Review Investigation.
  3. Recommendation systems can amplify extremes: Hao reported that Facebook’s engagement-oriented recommendations promoted inflammatory and extremist material in documented cases. That finding is about the systems and period investigated, not all social networks. — MIT Technology Review Investigation.
  4. Moderation cannot ignore ranking: Hao described a tension at Facebook: teams worked to remove harmful posts while engagement-driven recommendation systems could continue to promote similar material. — MIT Technology Review Investigation.
  5. Take internal warnings seriously: Hao documented Facebook employees and internal studies warning about harmful recommendations and polarization. Proposed fixes faced resistance when they threatened engagement or growth. — MIT Technology Review Investigation.
  6. Incentives affect what gets fixed: Hao found that engagement-linked priorities and leadership decisions shaped which Facebook recommendation and safety projects received support. Engineers did not simply lack awareness of offline harm. — MIT Technology Review Investigation.
  7. Inspect recommendation paths: Facebook’s internal research, as reported by Hao, showed that recommendation systems could lead users toward more extreme communities and content. The finding is more specific than a universal claim that feeds destroy shared reality. — MIT Technology Review Investigation.
  8. Ranking choices matter for polarization: Hao reported internal evidence and staff concern that Facebook’s engagement-driven ranking could intensify political polarization. The causal claim should remain tied to those documented systems and studies. — MIT Technology Review Investigation.
  9. Do not blame users alone: Hao’s reporting shows that a platform’s recommendation design can help distribute harmful content; user behavior is not the only variable to investigate. — MIT Technology Review Investigation.

Part 6: Geopolitics and Global AI Development

  1. Examine the AI-race narrative: Hao argues that casting AI development as a contest to be won can justify rapid scale-up and closer ties between technology companies and governments. Her interview does not establish that every funding decision lacks oversight. — Tech Policy Press Interview.
  2. Global AI development is interconnected: Hao’s account of U.S.–China competition discusses export controls, computing constraints, and alternative approaches in China. That evidence supports interdependence and policy trade-offs, not the categorical claim that technological decoupling is impossible. — Democracy Now Extended Interview.
  3. Patriotic framing can narrow debate: Hao describes how an AI race framed as a national imperative can make scale and speed appear nonnegotiable. That is her critique of the rhetoric, not evidence that every critic is silenced. — Tech Policy Press Interview.
  4. Give affected communities a role: Hao argues that regions supplying data, minerals, labor, and infrastructure should have a real say in AI development rather than gaining access only by offering resources for extraction. — AI Now Interview.
  5. Investigate biometric deployments locally: In explaining her empire framework, Hao points to a South African facial-recognition case as an example of AI built or deployed without adequate attention to the people it affects. The case does not establish the inherited worldwide disparity claim. — Tech Policy Press Interview.
  6. Open models involve governance choices: Hao supports wider independent access to AI models so researchers can scrutinize them, while acknowledging that the form and conditions of release matter. The direct account does not resolve every national-security concern. — UNSW Interview.
  7. Question what winning would mean: Hao argues that a race to build the biggest AI system can crowd out discussion of who benefits and which smaller or public-interest alternatives deserve support. — Tech Policy Press Interview.

Part 7: Safety, Alignment, and Silicon Valley Logic

  1. Ask whose values count: Hao’s OpenAI investigation notes the difficulty of deciding which values a general-purpose AI should embody and whether a small, unrepresentative group can make those choices for everyone affected. — MIT Technology Review Investigation.
  2. Do not overlook present harms: Hao argues that discussion of future existential AI risk should not eclipse the labor, environmental, and democratic effects already visible in the current development model. — Democracy Now Interview.
  3. Scrutinize claims of a universal mission: Hao traced how OpenAI’s stated mission to benefit humanity coexisted with commercial pressure and growing secrecy. Her reporting asks who gets to define that mission and hold the lab accountable. — MIT Technology Review Investigation.
  4. Representation shapes safety questions: Hao observed that OpenAI’s early staff was not representative of the many communities its systems might affect. She raised this as a problem for defining benefits and harms, without claiming a measured failure rate for non-Western contexts. — MIT Technology Review Investigation.
  5. Define AGI before measuring progress: Hao reported that researchers disagree about what artificial general intelligence would entail or how to know it had been reached. Claims that a benchmark proves human-level general reasoning therefore need careful qualification. — MIT Technology Review Investigation.
  6. Watch the safety–growth tension: Hao’s reporting on OpenAI’s 2023 board crisis describes senior leaders’ concerns about safety processes alongside pressure to ship products and grow the business. It documents a concrete tension, not an iron law that profit always wins. — The Atlantic Book Excerpt.
  7. Release reviews need independent scrutiny: In Hao’s account of OpenAI’s 2023 conflict, senior staff raised concerns that expedited product decisions could bypass established deployment-safety review. The claim is tied to reported incidents, not a verdict on all internal testing. — The Atlantic Book Excerpt.

Part 8: Journalism, Accountability, and the Future

  1. Report beyond company announcements: Hao’s own investigations followed workers, internal decisions, and data-center infrastructure rather than relying only on corporate descriptions of AI. That is a method visible in her reporting, not an unverified quotation from her. — The Atlantic Investigation.
  2. Separate capability claims from observed effects: Hao argues that both utopian and catastrophic AI narratives can skip over the steps by which claimed futures would actually occur. Reporting should test those pathways and the effects visible now. — Democracy Now Extended Interview.
  3. Make AI decisions understandable to the public: Hao connects public understanding of AI with people’s ability to challenge decisions about data, land, labor, and technology. Her argument is about democratic agency, not a demonstrated requirement for every citizen to master algorithms. — Democracy Now Extended Interview.
  4. Make independent scrutiny possible: Hao favors broader access to models and research so independent experts can examine AI systems outside the companies that build them. The inspected source does not establish a specific legal power to inspect every training dataset. — UNSW Interview.
  5. Treat AI as a set of choices: Hao’s reporting traces how decisions about model scale, publication, commercialization, and data-center siting shape AI’s effects. Those choices can be reported and contested rather than treated as inevitable. — Tech Policy Press Interview.
  6. Go where the costs are felt: Hao visited an Arizona data-center site and reported on the local water context; her Kenyan labor investigation also foregrounded workers who performed AI safety work. Both show why place-specific reporting matters. — The Atlantic Investigation.
  7. Investigate secrecy at the source: Hao reported that OpenAI became less open about its research as it pursued a compute-heavy strategy and commercial funding. The source does not support the inherited claim about opaque decisions in housing, employment, and justice. — MIT Technology Review Investigation.
  8. Communities can challenge the default path: Hao points to workers, artists, and local activists contesting how AI is built and deployed. Their actions show that the current scale-and-extraction model is a choice that can be challenged. — Democracy Now Extended Interview.
  9. A smaller, community-led path is possible: Hao highlights specialized AI projects shaped by local needs and data sovereignty, including language and climate initiatives. They offer alternatives to treating ever-larger models as the only route to progress. — AI Now Interview.