AI Empire Audit
Test whether an AI business creates value or extracts it without fair exchange
- Difficulty
- Advanced
- Time to result
- ~weeks to results
- Steps
- 6
- Confidence
- 98%
The AI Empire Audit translates Hao's empire metaphor into a repeatable institutional review. Start with the resources an organization claims, including data, intellectual property, land, energy, and water. Follow the labor involved in training and deployment, then inspect who funds research and whether critical work can be constrained. Examine the public story that presents one organization as the necessary protector against a worse rival. Finally, compare what the organization extracts from workers, users, creators, and communities with what each group receives and how much influence they retain. The mechanism shifts attention away from whether an executive seems benevolent and toward the structure of exchange and decision-making. A pattern of extraction, knowledge control, legitimating myths, and weak public participation indicates imperial behavior in Hao's analysis.
Origin
Karen Hao says empire is the only metaphor she found that captures the scale, resource claims, labor practices, knowledge control, and public narratives she observed while reporting on major AI companies.
Core principles
- 01Evaluate systems of power rather than relying on a leader's character
- 02Track who supplies resources and who receives the benefits
- 03Treat labor conditions as part of the technology's design
- 04Question narratives that make concentrated control seem necessary
- 05Require a fair exchange of value across the supply chain
How to run it
- 1
Map claimed resources
List the data, intellectual property, land, energy, water, and capital required by the system. Record who originally owns or depends on each resource.
Pro tip Include resources outside the company's balance sheet, such as community infrastructure and creators' work.
Watch out A resource being technically accessible does not establish that the exchange is fair.
- 2
Trace the labor chain
Identify employees, contractors, and data-annotation workers who make the system possible. Examine pay, stability, agency, and how automation changes their later opportunities.
Pro tip Follow subcontracting layers rather than stopping at the company named on the product.
Watch out New job counts alone can conceal lower pay, weaker conditions, or a broken career ladder.
- 3
Audit knowledge control
Map who finances research about the system's capabilities and harms. Look for access pressure, agenda setting, or reported attempts to suppress inconvenient findings.
Pro tip Separate independent evidence from research funded or controlled by interested firms.
Watch out Industry expertise is useful, but it should not be treated as the only legitimate source of knowledge.
- 4
Decode the legitimating story
Identify claims that concentrated control is necessary to deliver abundance or prevent a rival from causing catastrophe. Ask what resources, permissions, or regulatory advantages the story is intended to mobilize.
Pro tip Compare how the goal is defined for consumers, governments, and investors.
Watch out A dramatic prediction is not evidence that the predicted future is inevitable.
- 5
Test the exchange
Compare the value each group contributes with the value, rights, and voice it receives. Flag arrangements where gains concentrate while costs are shifted to workers, creators, users, or local communities.
Pro tip Ask affected groups directly rather than inferring their interests from company statements.
- 6
Review decision rights
Determine who can approve, contest, or stop decisions that affect the public. Treat a lack of meaningful participation as a governance defect even when leaders claim benevolent intentions.
Watch out Replacing one executive does not repair a structure that remains highly concentrated.
In the wild
Hao cites Timnit Gebru and Margaret Mitchell while alleging that Google tried to stop critical research about large language models and then fired the two ethical-AI co-leads. In her audit, this is presented as an example of a company constraining knowledge that conflicts with its agenda; the episode gives Hao's account rather than Google's response.
→ The knowledge-control dimension becomes visible alongside the technology's technical claims.
Hao describes reporting in which displaced professionals take lower-status data-annotation work that trains models on the jobs they previously held. She argues that this can both degrade working conditions and help automate further work in the same field.
→ The audit captures job quality and career progression rather than counting replacement jobs alone.
Common mistakes
Judging only the chief executive
Hao argues that a leader's perceived moral character cannot fix a governance structure that concentrates decisions affecting billions of people.
Counting benefits but externalizing costs
A useful product can still impose labor, resource, or community costs that must be included in the assessment.
Treating predictions as destiny
Hao says industry forecasts also function as persuasive speech that can mobilize capital and power, so they should be examined rather than assumed.
Is it for you?
Best for
Policy reviews, procurement decisions, journalism, investment diligence, and governance discussions about large AI companies.
Not ideal for
A technical model evaluation that does not examine labor, resources, institutional power, or affected communities.
From the transcript
“They lay claim to resources that are not their own in the pursuit of training these models.”
“Third, they monopolize knowledge production.”
“the bigger question is is the governance structure that we've created a sound one”
From the episode
AI Whistleblower: We Are Being Gaslit By The AI Companies! They’re Hiding The Truth About AI