Who Controls AI Ethics?
The Nine-Day Board, $36M in Lobbying,
and the Answer Nobody Likes

Google dissolved its AI ethics advisory board before it held a single meeting. Big Tech spent $36 million lobbying against AI oversight while simultaneously publishing responsible AI principles. The EU’s enforcement clock hits zero in August 2026. And most organizations still can’t name who has authority to halt their highest-risk AI system.

Part I The Nine-Day Board

The ATEAC case gets used sparingly in AI governance writing. That’s a mistake, because it is the cleaner data point. When Google fired Timnit Gebru in December 2020, you could at least argue about the causal chain — whether the paper’s content drove the termination, whether other factors were in play. People did argue. The structural problem was clear but the specific causal claim was contested.

The ATEAC case requires no inference. No contested causal chain. The sequence is public and documented: Google assembled external oversight, encountered a governance disagreement before the council had done anything, and dissolved it. The council had not held a meeting. It had produced no position on any subject. Council member Joanna Bryson wrote afterward that the group had been preparing to stress-test Google’s internal facial recognition policy — that the people she spoke to at Google “seemed profoundly disturbed by what face recognition could do.” That work never happened. The council was nine days old. It was still easier to end than to maintain.

Seventeen months later, the Gebru case supplied the second data point: this time showing what happens when internal ethics research, rather than external oversight, threatens a commercial interest. Together they establish a two-case pattern. External oversight dissolved at first friction. Internal oversight suppressed at first commercial conflict. Two separate mechanisms, same result. Neither is a one-off bad actor story. Both are what the incentive structure produces.

Part II $36 Million and What It Actually Buys

The gap between what AI companies publish and what they fund is not a secret. It’s just rarely described in the same sentence, which is convenient for everyone involved.

During Q2 2025, Big Tech tried to push a provision into Trump’s spending bill that would have stripped states of the power to regulate AI and social media algorithms for the next ten years. A blanket, decade-long federal preemption of the most active regulatory jurisdiction in the country — while federal frameworks remained incomplete and enforcement mechanisms untested.

Here’s the part that deserves explicit naming, because it usually gets reported as two separate stories: the same companies pushing that preemption were simultaneously publishing responsible AI principles and commissioning ethics research. Both things are true. The ethics commitments are not fabricated. The lobbying strategy is not fabricated. They coexist because the incentive structure allows it — and in fact rewards it. Commit to ethics in the domain you control. Lobby against oversight in the domain you don’t.

Part III Regulatory Capture Isn’t a Conspiracy. It’s a Mechanism.

The 2025 AI & Society peer-reviewed analysis doesn’t use the word “conspiracy.” It uses “structural condition.” The distinction matters.

Regulatory capture is what you get when the regulated entities have more resources, more technical expertise, and more sustained attention to regulatory processes than the bodies overseeing them. The FDA version has been documented for decades: senior officials cycle into industry roles, taking institutional expertise and relationships with them. The AI version operates the same way — except the technical complexity is higher, the development pace is faster, and the standards organizations nominally overseeing AI are frequently funded by the industry they’re meant to check.

The auditor analogy is the clearest frame. A professional auditor at a public company cannot simultaneously hold equity in the company being audited. That rule exists not because auditors are assumed dishonest — it exists because the structural conflict corrupts independent judgment in honest people, gradually, in ways the person experiencing the conflict usually can’t detect. A hospital-funded safety review of hospital practices produces systematically more favorable findings than an independently funded one. Not lying. Drift: in the framing of research questions, the selection of comparison cases, the interpretation of ambiguous data. All of it drifts toward the funder’s interest without anyone deciding it should.

Part IV The Three Actors — and Why None Can Hold the Key Alone

There are three actors with a genuine claim on AI ethics authority. Each has real advantages. Each has a structural failure mode. Understanding all three is required to understand why the current governance architecture produces ethics performance rather than ethics substance.

The complicating finding — the one that works against a simple “regulate harder” answer — is that the EU AI Act is already showing implementation strain. The Council on Foreign Relations’ January 2026 analysis identifies a regulatory arbitrage dynamic: countries crafting permissive environments attract agentic AI investment the same way offshore financial centers attracted capital. The countries with meaningful enforcement face a competitive disadvantage relative to those without it, regardless of governance quality.

China resolved its version of this problem by replacing commercial concentration with party oversight. That eliminates the power asymmetry between labs and the state — and creates a different accountability failure. Neither model resolves the foundational problem.

Part V The Number That Reveals the Whole Architecture

A cross-source finding that doesn’t appear in any single published analysis — but emerges when you put three datasets in the same frame.

These are not the same governance instrument. One of them is growing at 84% annually. The other isn’t growing. The governance architecture being built in response to regulatory pressure is producing the appearance of accountability more reliably than accountability itself.

The lobbying data, the board-oversight data, and the regulatory capture literature, read together, show the same structural dynamic from three separate angles. Companies are not adding ethics boards — they’re adding board-level oversight disclosures. The distinction between “the board monitors AI risk” and “there exists an independent body that can produce findings the company must act on” is the entire substance of the governance question. In most organizations in 2026, the former exists. The latter frequently doesn’t.

The Nine-Day Test Is Coming for Every Organization

No single actor holds the key to AI ethics authority. The evidence doesn’t support the model where one should. What it supports is a distributed accountability architecture where labs are constrained by binding external regulation they cannot lobby away, regulators are constrained by independent technical audit capacity they don’t currently possess, and civil society is constrained by the funding and legal infrastructure its independence requires.

That architecture doesn’t currently exist. What exists is performance of distributed accountability: ethics boards that dissolve in nine days, oversight disclosures growing 84% annually without corresponding independent oversight bodies, standards organizations funded by the entities they nominally govern.

The foundational assumption is this: structural independence is not a design preference. It is the minimum condition under which oversight produces information the organization doesn’t already prefer. The ATEAC council wasn’t ended because it had produced a problematic finding. It was ended because maintaining it required managing a disagreement the company could avoid by dissolving it. The entity that holds the authority to end oversight when it becomes inconvenient holds the master key. In 2026, that entity is still, in most cases, the company itself.

Every organization deploying high-risk AI in the next 18 months faces a version of the nine-day test. Not one imposed by external pressure, but the internal moment when ethical oversight becomes operationally inconvenient — when a system flags a concern that would delay a launch, when an audit function surfaces a finding that conflicts with a revenue projection, when a governance review creates friction with a government contract timeline. Most organizations haven’t been tested yet. The enforcement timeline is fixed. The test is coming.