Chasing the Hidden Truths of AI, Claim by Claim

Adityo Guni WaluyoOctober 5, 202601:12

Karen Hao's interview drew 5.4 million views. I traced its main claims to 24 sources: most held up, one collapsed, three remain unproven.

#fact-check #regulasi #data-center

TL;DR

Fact-checking a viral interview found lobbying and health claims held up, with $41 million spent in six months and measurable deaths linked to one data center. The regulatory gaps and delayed safety disclosures were also well documented. However, the viral 40% drop in entry-level jobs was contested, with primary data showing no broad AI-driven unemployment spike.

There was one second in Karen Hao's interview on The Diary Of A CEO that made me pause. She said AI companies spend hundreds of millions of dollars to kill legislation that gets in their way, and the video has been watched more than 5.4 million times. I didn't want to get angry just because the speaker sounded convincing. Tearing it apart without checking felt lazy too. So there was only one option: trace its main claims to the original sources [1].

My first guess about one number missed badly. The interview cites a roughly 40 percent drop in entry-level workers because of AI, and I assumed the figure came from Anthropic's research. Fair guess, since Anthropic does run its Economic Index. When I opened their latest publication, the text pointed the other way: *no systematic increase in unemployment for highly exposed workers since late 2022* [19]. The 40 percent figure matches no primary source still in circulation, including Anthropic's own.

The checking process holds up, so let me show it briefly. Eight search waves, starting from the first keyword to new research directions born from data gaps. Every source entered a tiered ledger: regulator documents, universities, and field studies at the top, media and opinion below. Of hundreds of hits, 24 sources passed with 27 verbatim quotes stored in full. One rule I refused to break: a number only enters if its page actually opened in front of me, never from a search snippet.

Claims that held up

It starts with narrative building. Karen Hao says internal documents show companies deliberately growing fear and euphoria around AGI to fundraise. I can't read those documents myself, but the documentary basis is real: the book is built on roughly 260 interviews, correspondence, and internal papers [17]. The pattern she criticizes also shows up on the regulatory side. The FTC once took action over training models on customer data without consent, up to a precedent requiring models trained on unlawfully obtained data to be deleted [8]. A Stanford paper closes the theory: today's transparency rules have three structural gaps, from specification to enforcement [7]. The hiding mechanism even has a shape: a Lawfare analysis shows companies can delay safety-report publication by invoking trade secrets, with no independent verification of that claim [10], and the NTIA adds the policy side: no standard AI accountability reporting framework exists yet [9].

Lobbying claims? The documented part alone is striking. Joint spending by 11 tech companies and trade associations hit 41 million dollars in the first half of 2026, up 8 percent year over year, with 324 registered lobbyists in Q2 alone [20]. Anthropic nearly tripled to 3.53 million dollars, OpenAI doubled to a record 2.22 million [20][14].

On health impacts, two new studies deliver numbers that didn't exist before. An analysis of a single data center facility in Loudoun County estimates 53 to 99 million dollars a year in health damages, with 3.4 to 6.5 additional premature deaths from on-site gas turbines [22]. Harvard research shows fine particulate pollution drives nearly 90 percent of air pollution's health burden and is often left out of standard energy or climate assessments [11]. Memphis makes it visible: residents were first to notice rows of unpermitted gas turbines on the xAI campus, with regulators moving late [23]. Virginia's government now runs a dedicated air monitoring program for data centers, a quiet admission that this is real [24].

The number that collapsed against its own primary source

Here's the part that bites. Anthropic's March 2026 publication states there has been no systematic rise in unemployment since late 2022, though hiring of young workers in AI-exposed occupations appears to have slowed [19]. The Stanford Digital Economy Lab team finds real entry-level declines, but small in aggregate [18]. EPI watches young unemployment rise in parallel for college and non-college grads over three years, which weakens the AI attribution [15]. The New York Fed data quoted by NPR: 5.7 percent recent-graduate unemployment against 4.1 percent overall [16].

My verdict on the 40 percent figure: contested, and possibly stale. The three primaries above partially contradict each other, while the number repeated everywhere matches none of them. A sympathetic whistleblower can still carry an outdated number. That's not a reason to toss the whole story; it's the reason to separate claims from facts.

What remains unproven

Three things stay open, and I'll say so plainly. Karen Hao's internal documents: the basis is documented [17], the contents nobody outside can inspect. The hundreds-of-millions magnitude for killing regulation: the documented federal floor is in the tens of millions a year [20], and the super-PAC layer stays unmapped. Epidemiology around data centers: the very studies calling for this research note the data doesn't exist yet.

Klaim di video	Status setelah dicek	Sumber
The AGI narrative is manufactured for fundraising	Strong	[1][7][8][17]
Record-breaking lobbying	Strong, 41 million dollars per 6 months	[14][20]
A data-center health crisis	Strong, quantified per facility	[11][22][23]
Entry-level down 40 percent because of AI	Contested	[15][16][18][19]
Hundreds of millions spent killing regulation	Partially proven	[1][20]

If you come across a similar AI claim, three questions are now my default filter. Is there a primary source? Does that primary still say the same thing today? And is the counterevidence you read independent, not just a rebuttal from an interested party? Pass all three, the number is quotable. Fail one, write it as a claim, not a fact.

What surprised me: the fact-check made Karen Hao's story stronger, not weaker. The claims that survived are far more concrete than what I heard in the video: not a vague sense of threat, but 41 million dollars of lobbying in six months and 3.4 premature deaths a year at a single facility. Honest numbers sell the story better than inflated ones.

Sources