STRATEGYOngoing practice92% confidence

The AI Empire Accumulation Pattern

Frontier AI follows colonial logic — four extraction pillars, structural not personal

Problem it solves

Replaces CEO character analysis with structural explanation of AI industry behavior

Best for

Analysts, investors, or builders evaluating structural risk and regulatory trajectory for centralized AI companies

Not ideal for

Near-term price catalysts or tactical trading decisions — this is a slow-cycle structural lens

Overview

Why this framework exists

Hao argues that frontier AI companies operate according to four structural imperatives that mirror colonial empire logic: resource appropriation (data as land grab from individuals, artists, and creators without consent), labor exploitation (double extraction where displaced workers annotate data for the models that replaced them), knowledge monopolization (funding most of the world's AI researchers so inconvenient findings can be suppressed), and narrative control (deploying existential risk and consumer benefit narratives simultaneously for different audiences). These behaviors are not the result of individual bad actors but of the structural requirements of building frontier models.

The framework's key insight is that swapping executives changes nothing. The system of power concentrates decisions affecting billions of people into a handful of unelected actors, with no democratic input from those affected. The architecture of frontier AI development — requiring massive capital, compute, data, and research talent — inevitably produces imperial logic in any organization attempting it.

This framework matters for investors because it predicts regulatory trajectory independent of which companies are currently winning. If the resource appropriation and narrative control pillars are structurally inherent, then any regulatory response targeting data practices or market concentration hits every frontier lab equally — and the companies most differentiated from this pattern become structurally advantaged.

Core principles

5 total
  1. Imperial accumulation is structural, not personal — any organization building frontier AI faces the same incentive architecture regardless of its founders' stated values.
  2. The four pillars (resource appropriation, labor exploitation, knowledge monopolization, narrative control) are interdependent — removing one without the others does not break the pattern.
  3. Narrative control is the multiplier — it protects the other three pillars by framing extraction as innovation and existential necessity.
  4. The 'good empire vs evil empire' myth is always reconstructed when the previous evil empire becomes cooperative — first Google, then China.
  5. Democratic legitimacy is the missing input — decisions affecting billions are made by a handful of actors with no accountability mechanism to those affected.

Steps

6 steps
  1. Map the resource appropriation layer
    Identify what resources the company is consuming without consent or fair compensation: user data, creator intellectual property, or public commons. Ask whether their appetite for data has decreased as model capability increased — if not, the extraction is structural, not temporary.
    Pro tipHao's diagnostic: 'If the horse truly had left the stables, they wouldn't have to train on anything anymore. Why is it that their appetite for data has actually expanded?' Use this as a stress-test question.
    WarningCompanies frame data collection as 'training for everyone's benefit' — this is narrative control, not a principled rebuttal to appropriation.
  2. Trace the labor extraction loop
    Follow the displacement-annotation cycle: identify which job categories the model is eliminating, then check whether displaced workers from those categories are being recruited to annotate training data. The double extraction pattern is the tell — value is extracted twice from the same worker.
    Pro tipLinkedIn's top-10 fastest-growing jobs data is a leading indicator — when displaced professional categories show up as annotation labor surges, the loop is active.
  3. Audit the research funding graph
    Map the source of funding for AI safety, capability, and ethics researchers relevant to your evaluation. If a company directly employs or funds the majority of researchers in a domain, treat findings favorable to that company as compromised until independently replicated. Apply the climate science analogy explicitly.
    WarningResearch suppression is not always overt firing — it includes editorial influence, publication delays, and grant conditions.
  4. Decode the narrative stack by audience
    For any major AI company statement, identify which audience it is targeting and what behavior that audience needs to exhibit (provide capital, withhold regulation, adopt product, provide data). Match the claim to the audience requirement rather than evaluating it as a standalone factual assertion.
    Pro tipAltman's four AGI definitions for four audiences (Congress, consumers, Microsoft, OpenAI website) is the canonical example. Ask: who needs to believe this, and what do they need to do next?
  5. Test for structural vs. personal causation
    Ask: if every executive at this company were replaced with someone widely regarded as more ethical, would the four extraction patterns continue? If yes, the problem is structural and regulatory response is the only lever. If no, identify which specific structural change would eliminate the incentive.
    Pro tipHao's formulation: 'Even if you were to swap all the CEOs for someone that people would say is better at running these companies, it doesn't fix the problem.'
    WarningPersonalizing the critique (focusing on Altman specifically) is a trap — it implies the solution is personnel change, which the industry prefers to structural change.
  6. Identify the regulatory pressure point
    Determine which of the four pillars is most legally vulnerable in the current jurisdiction — data appropriation (copyright/privacy law), labor practices (worker classification), research funding conflicts (disclosure requirements), or narrative control (securities fraud / misleading statements). Regulatory pressure will enter through the weakest pillar first.
    Pro tipThe 80% American consensus on AI regulation means political will exists — watch for which pillar legislators target as a signal of which companies face the greatest structural disruption.

Checklist

Saved in your browser

Examples

3 cases
OpenAI Startup Fund misappropriation

Independent board member Adam D'Angelo (CEO of Quora) requested documentation during the November 2023 governance crisis and discovered that the 'OpenAI startup fund' was structured as Altman's personal startup fund, not OpenAI's. This was one of multiple accumulating evidence points that led the board to fire Altman. The board's logic: 'This is not Instacart' — the governance stakes of a company claiming to build transformative AI warranted a higher standard than normal startup governance.

OutcomeAltman was fired, then reinstated after a stakeholder campaign. The board that fired him was replaced. The underlying structural issue — a single individual controlling multiple capital pools with conflicting incentive structures — was not resolved.
Memphis Colossus methane turbines

When Musk's xAI built the Colossus training facility in Memphis, they powered it with 35 methane gas turbines. Community members discovered this not through official disclosure but because they 'smelled what seemed like a gas leak in all of their living rooms.' The facility's environmental impact was externalized onto a residential community with no input or consent. The OpenAI Stargate facility in Abilene, Texas is planned at the scale of Central Park, running 1 million chips, consuming more than 20% of New York City's power equivalent.

OutcomeDozens of protests against AI data centers have emerged across the US. Environmental externalities are becoming a political liability for frontier AI development — a structural pressure point for the resource appropriation pillar.
Data annotation displacement loop at scale

A New York Magazine investigation cited by Hao documented award-winning Hollywood directors secretly working as data annotators at piece-rate pay. Workers could not leave their laptops to care for children because project windows close unpredictably. One annotator stated: 'I have become a monster... this industry is mechanizing my life, atomizing my work, devaluing my expertise.' The Klaviyo case provides the corporate perspective: headcount fell from 7,400 to 3,300, AI handles 70% of customer service conversations, and revenue doubled.

OutcomeLinkedIn data shows data annotation entering the top-10 fastest-growing job categories. The Anthropic internal report cited by Hao shows 40% reduction in entry-level jobs already visible in relevant professions. The displacement-annotation cycle is documented and accelerating.

Common mistakes

5 traps
Personalizing the critique
Focusing on individual executives (Altman, Pichai, Amodei) as the source of problematic behavior implies that replacing them would fix the problem. Hao's framework shows this is wrong — any organization attempting frontier AI development faces identical structural incentives. Personalization redirects attention from structural reform to leadership change.
Accepting the good empire / evil empire framing
When AI companies invoke China or Google as the 'bad actor' that justifies their own accumulation, the narrative is functioning exactly as designed. The evil empire framing has migrated over time as each previous competitor became cooperative — it is a renewable narrative resource, not a factual description of competitive dynamics.
Treating AGI definitions as sincere technical claims
Evaluating AGI progress against any single definition misses that the definitions shift by audience and purpose. The correct frame is: what audience needs to be mobilized, and what definition serves that mobilization? Altman's definitions for Congress, consumers, Microsoft, and internal use are all simultaneously active and mutually inconsistent.
Assuming research funding independence
When AI companies fund or employ the majority of researchers in a domain, the research output is systematically biased toward findings that do not threaten the funding source. The Gebru and Mitchell firings at Google are documented examples — but the more pervasive effect is the research that is never pursued, published, or funded in the first place.
Underweighting democratic legitimacy as a structural variable
The 80% American consensus on AI regulation is routinely dismissed as polling noise by industry insiders. Hao's point is the opposite: this level of cross-partisan consensus on any issue is extraordinary and historically predictive of eventual policy change. The only questions are timing and form.

Origin story

How this framework came to be

Hao spent eight years covering AI for MIT Technology Review and The Atlantic, conducting more than 250 interviews including 90+ current and former OpenAI employees and executives. Her book 'Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI' documents the pattern across multiple companies, not just OpenAI, drawing on internal documents, Slack messages, and email disclosures from the Musk/Altman lawsuit. The empire analogy emerged from her observation that behaviors she was documenting across the industry — data appropriation, researcher silencing, narrative management — were structurally identical across companies with very different stated values and leadership styles.

The intellectual anchor is her analogy to fossil fuel companies: 'If most of the climate scientists in the world were bankrolled by fossil fuel companies, do you think we would get an accurate picture of the climate crisis?' Her claim is that the same dynamic now applies to AI safety and capability research.

Source

Traced to primary
Source · PODCAST
AI Whistleblower: We Are Being Gaslit By AI Companies, They're Hiding The Truth!
Karen Hao · 2025
Open source →

Related frameworks

Browse all Strategy →