You asked for safety. They answered with a leash.
Three years ago, AI execs started whispering the same line: Yes, we need rules. Dario Amodei told Congress last year to bind AI down like a dangerous animal. Google’s Kent Walker called for “pragmatic, evidence-based” oversight in June 2026—just weeks after Anthropic’s models were suspended for harm. And yet here’s the punchline: nobody wants oversight that actually changes how they run things.
Google’s latest 21-page treatise—euphemistically titled “A Pragmatic Approach to AI Governance in America”—isn’t a plea for public protection. It’s a pitch deck disguised as policy, arguing that the only rules worth having are ones Google helps write. The centerpiece? A “frontier AI regulatory organization” (FARO), a body modeled after industry clubs like FINRA and NERC, where the regulated write the checklist. Call it compliance theater with a PhD.
Let’s cut through the jargon: if AI really posed an existential threat—as some of these same executives once warned—we’d be talking about shutdown protocols, not disclaimers that nobody reads. Instead, what we’re getting is a regulatory palimpsest: layer over layer of “reasonable measures” that let the platform keep its current superpowers intact.
So what’s on the table? Disclaimers on chatbots that no one notices, a copyright carve-out for scraping the whole internet, and datacenter siting negotiations that sound like zoning hearings at a casino. And beneath all of it, AI lobbying is up 340 percent since 2023. The math is simple: money talks, and Google just dropped millions to be heard.
This isn’t about balance. It’s about bandwidth—only the most expensive bands, only the ones Google has already claimed for itself.
The Faro Fix: Industry regulation, one phone call to Google later
FARO—frontier AI regulatory organization—is the crown jewel in Google’s governance proposal, and it might as well be called Google Oversight Inc. The idea is dead simple: create a government-chartered body to police frontier AI, then staff it with people Google knows, likes, and trusts. They’re not joking around—the paper spells out a checklist: membership rules, funding mechanisms, enforcement teeth—all designed to look neutral while quietly favoring incumbents.
Sound familiar? This is the same script played out in electricity, finance, and medicine. NERC keeps grid reliability humming, FINRA polices broker behavior, and the AMA weights medical licensing with clinical jargon. Each organization looks independent on paper, but ask a small operator how easy it is to gain access, and you’ll hear a different story.
The FARO model has the same effect: it raises compliance costs just enough to sideline startups while letting established players like Google and Microsoft shape policy through quiet appointments. Walker’s blog post touts “evidence-based” rules, but evidence doesn’t stand up to rent-seeking. When Google proposes disclaimers about chatbot identity—“regularly point out that it’s not a person”—it’s not trying to trick users into trust; it’s building plausible deniability for when the model hallucinates insurance claims or fabricates medical advice.
The deeper trick? FARO isn’t preventing harm. It’s insulating Google from liability when harm happens. The paper argues that FARO-style oversight should cover “prompt injection” threats, yet it says nothing about the company’s own training data or model fine-tuning processes that enable those very threats. Oversight? Sure, but only on the surface—with enough regulatory debt to keep lawyers employed for years.
The art student and the Louvre: Google’s fair-use flimflam
Here’s Google’s latest copyright pivot, straight from the policy paper: using publicly available web data for training models is like “an art student taking inspiration from walking through a gallery”—transformative, non-expressive, and therefore fair use. The analogy sounds charming until you realize who’s playing the art student and whose museum the artist is looting.
Let’s unpack that metaphor. A real art student might sketch a few pieces from memory after wandering the Louvre’s galleries, then go home and produce original work that echoes Degas or Delacroix. Google? It’s the art student who grabs every high-resolution image from the museum, puts them behind a paywall, and sells subscriptions to scholars who’d otherwise buy tickets. Then—this is the kicker—the student offers a free workshop on “how to see like the Louvre,” complete with a job placement program that’s so performative it barely deserves the name.
That’s exactly what Google’s doing on a trillion-token scale. It takes every book, blog post, article, and forum thread it can scrape, feeds them into the model, then spins off a “safe” version with disclaimers and guardrails that barely slow the hallucinations. Meanwhile, human creators—photographers, journalists, novelists—watch their livelihoods dissolve into training data. Courts are still sorting through those cases, but Google’s lobbying budget has already bought time: more amendments, more delays, more “let’s keep iterating” as everything burns.
The paper never acknowledges the economic damage, just the risk of over-regulation. There’s a reason for that. If Google truly believed in fair use, it wouldn’t lobby so hard to keep data scraping off-limits for scrutiny. Instead, the company plays both sides—begging courts to protect its dataset while demanding liability shields for whatever comes out the other end.
The money trail: 340 percent and counting
The policy paper drops at the same time lobbying spend hits a new peak. AI industry spending on federal lobbying jumped 340 percent between 2023 and this year, according to公开 records cited by The Register. That’s not charity. It’s a down payment on influence—cash paid to keep proposals like FARO alive, while more aggressive ideas (say, banning non-consensual deepfakes or mandating transparency for image-generation models) gather dust in committee.
Google’s been quiet about where the money’s going, but the pattern is clear. Each new AI rule proposal comes with a corresponding spike in Beltway hiring. The company stakes out ground on copyright early, then shifts focus to datacenter siting, telling towns that building a datacenter “responsibly” means working with them—not about them, not for them.
The problem with that framing is simple: communities aren’t signing up. Across the U.S., town councils are rejecting datacenter expansions—not because they’re anti-tech, but because the promised benefits (jobs, tax revenue) often evaporate once construction ends. Meanwhile, the machines keep humming, guzzling megawatts and never apologizing for the noise.
Lobbying isn’t evil. But when a company spends millions to shape rules while denying the problems its products created, that’s not advocacy—it’s obfuscation with a budget. Google can stand on stage and call for “pragmatic” oversight while quietly funding the people who write that pragmatism into law. The result? A regulatory landscape so saturated with exceptions and carve-outs it barely qualifies as such.
Section 230, reborn: The internet’s ghost haunting AI policy
We’ve seen this movie before. In 1996, Congress passed Section 230 of the Communications Decency Act, promising platforms immunity so long as they made “good faith” efforts to police harmful content. Sounds reasonable, right? Except “good faith” meant whatever courts and later court precedents decided it meant—and over time, that drifted toward no fault at all.
Platforms promised safe spaces. What they built were echo chambers, ad-optimized rage machines, and recommendation engines that turned misinformation into currency. The disclaimers arrived in later amendments—notice boards, age gates, “viewing history” opt-outs—but the real protections remained untouched. Liability was still capped at $0, as long as the company could point to its Terms of Service and call it a start.
Google’s FARO plan is the 2026 version of that playbook. It asks for disclaimers, guardrails, and occasional public apologies—all performative safety—while dodging anything that would meaningfully constrain the model. The paper insists chatbots shouldn’t promote emotional dependency, but nowhere does it suggest capping engagement features or capping training on scraped data that produces those addictive loops in the first place.
The lesson of Section 230 is clear: unless you build in teeth from day one, regulation becomes an obstacle course for honest actors—while bad ones simply adjust their route. AI leaders keep saying they want rules, but rules for whom? If the goal is public safety, then disclaimers are a start. If the goal is market dominance under a moral veneer, then FARO is the ticket.
The middle path is already paved with good intentions
Google’s regulatory proposal sounds like a reasonable compromise, because it’s written to sound that way. Kent Walker mentions “evidence-based” approaches, community partnerships, and the need to avoid over-regulation. But evidence would suggest otherwise. The same company that once pushed “move fast and break things” now wants to move slowly—and only where it suits them.
The real middle ground doesn’t look like disclaimers and datacenter negotiations. It looks like strong copyright enforcement, mandatory incident reporting, independent audits of high-risk models, and liability for harms that an organization’s own products can directly cause. None of those show up in Google’s policy paper—not because they’re hard, but because they’d cut into the bottom line.
It’s tempting to believe executives really worry about AI risk. But when the same people who once claimed regulation would kill innovation now lobby for rules they helped draft, you have to ask: whose risk are we mitigating? Is it the existential threat AI might pose? Or just the reputational risk of appearing to do nothing?
FARO isn’t a solution. It’s the end of the conversation—just dressed up like the beginning.