AI Industry Creates New Age of Imperial Extraction
Journalist Karen Hao, author of Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI, said in a Democracy Now! interview rebroadcast July 3, 2026 that AI development is tied to labor and environmental harms in Kenya and Chile. According to Hao, contract workers in Kenya were paid a few dollars an hour or less to review graphic AI-generated text for OpenAI's safety filters, and a Google-linked data center project near Santiago, Chile sought roughly a thousand times a local community's annual freshwater use. "This is an extraordinary type of AI development that is causing a lot of social, labor and environmental harms," Hao said. For AI teams, the account is a reminder to extend vendor due diligence beyond model performance to labor conditions and resource disclosures.
For AI practitioners and enterprise buyers, this interview is a reminder that AI's risk surface extends well beyond model performance into labor sourcing and infrastructure siting. Journalist Karen Hao's reporting, the basis for her book Empire of AI, ties specific data-labeling and data-center practices to labor and environmental harms that carry direct vendor-due-diligence implications for any team sourcing training data or leasing large-scale compute.
What happened
In a Democracy Now! interview rebroadcast July 3, 2026 (revisiting an earlier conversation with hosts Amy Goodman and Juan Gonzalez), Hao, a former Wall Street Journal and MIT Technology Review reporter, described findings from her multi-year investigation of OpenAI and the wider AI industry. According to Hao, OpenAI contracted data-annotation firms in Kenya whose workers reviewed graphic, AI-generated text for a few dollars an hour or less to help build the content-moderation filters behind ChatGPT, leaving many workers psychologically traumatized. She also described a Google-linked data center proposal near Santiago, Chile that, per her reporting, sought to draw roughly a thousand times the freshwater a local community uses annually from one of the region's last public freshwater sources; sustained community activism reportedly delayed the project for four to five years and won residents a seat at government-mediated talks. "This is an extraordinary type of AI development that is causing a lot of social, labor and environmental harms," Hao said.
For practitioners
The due-diligence questions generalize beyond any single vendor: verifying pay and working conditions in outsourced data-annotation pipelines, requesting water- and energy-use disclosures for data center sites under consideration, and factoring community or regulatory pushback into infrastructure timelines. These increasingly show up as standard vendor-risk and ESG checks for enterprises procuring AI infrastructure or labeled data at scale.
What to watch
Whether government-mediated negotiation frameworks like the one Hao described in Chile spread to other water- or power-constrained data center markets, and whether AI data-labeling and content-moderation labor practices face the kind of scrutiny social media content moderation did in the previous decade. Note this account draws primarily on Hao's own reporting and book rather than an independently audited source; specific figures such as the freshwater comparison are her reporting's characterization.
Key Points
- 1Journalist Karen Hao's Democracy Now interview says OpenAI-linked data-labeling work in Kenya paid workers a few dollars an hour reviewing graphic content.
- 2Hao reports a Google-linked Chile data center sought roughly a thousand times a community's annual freshwater use, prompting years of local activism.
- 3The account signals AI buyers should extend vendor due diligence to labor conditions and water and energy disclosures, not just model performance.
Scoring Rationale
Solid, well-attributed account of AI-industry labor and environmental harms (Kenya data-labeling conditions, Chile data-center water use) from a credentialed journalist's book and reporting, relevant to practitioner vendor due diligence. Kept in the 5.0-6.4 band rather than higher because this is a rebroadcast of earlier reporting rather than new news, and the specific figures rest on a single reporter's account not independently audited.
Sources
Public references used for this report.
Practice with real Ad Tech data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Ad Tech problems
