Google Decides What Counts as Harm — and That Should Worry Everyone

NewsOnScale Staff

September 2, 2026

When a government agency defines harm, there are hearings, public comment periods, legal challenges, and — however imperfectly — democratic accountability. When Google defines harm, there is a blog post.

That is, roughly, the situation created by Google's newly released AI governance framework, which lays out the company's internal architecture for deciding what its AI systems should and should not do. The document is detailed, professionally written, and genuinely more transparent than what most AI developers have offered. It is also a reminder that the most consequential decisions in the AI economy are still being made by a small number of private actors operating largely outside any formal accountability structure.

## What the Framework Actually Does

At its core, Google's governance plan establishes categories — what the company considers harmful outputs, sensitive topics, high-risk use cases, and acceptable tradeoffs between helpfulness and safety. These are not trivial editorial choices. They determine what a Gemini-powered tool will say to a user asking about medication dosages, election integrity, or geopolitical conflict. They determine which business customers can build which kinds of applications on Google's infrastructure.

The framework appears to draw on genuine internal deliberation and reflects some legitimate best practices in AI safety thinking. But the act of drawing those lines — and drawing them privately — is itself a form of power that deserves scrutiny independent of whether the lines are drawn well.

The definitional problem is not academic. AI governance frameworks function as a kind of constitutional document for the systems they govern. What gets coded as "harmful" shapes what gets suppressed. What gets coded as "safe" shapes what gets amplified. In an environment where Google's AI products are embedded in search, productivity software, healthcare tools, and educational platforms, those choices have downstream consequences for public discourse, professional practice, and civic life.

## The Regulator Gap This Fills — and Exploits

It would be unfair to suggest Google is acting in bad faith by publishing this framework. The more accurate observation is that the company is filling a vacuum that governments have so far failed to fill. Brookings, the Regulatory Review, and a growing list of policy institutions have been sounding this alarm for months: without federal AI legislation in the United States, companies are writing their own rules by default.

Google's framework is, in part, a response to that pressure — a preemptive demonstration of responsible self-governance intended to reduce the urgency of external regulation. That is a reasonable corporate strategy. It is also precisely the dynamic that makes the framework worth examining critically.

Self-governance documents tend to be written to foreclose the hardest questions rather than invite them. They establish that harm exists, that the company takes it seriously, and that internal processes exist to address it — without necessarily making those processes legible or challengeable from the outside.

## What Accountability Would Actually Look Like

Transparency about a governance framework is not the same as accountability for how it operates. A genuinely accountable AI harm definition process would include external audits of how the framework is applied in practice, not just how it is described in policy. It would include clear appeal mechanisms for users or developers who believe the framework has been misapplied. And it would include some form of independent review — whether by a regulatory body, a multi-stakeholder panel, or both — that does not begin and end with the company itself.

None of that currently exists at scale. In its absence, governance documents like Google's serve a real function: they surface the logic driving AI systems that would otherwise remain entirely opaque. That is worth acknowledging. But surfacing the logic is different from subjecting it to meaningful scrutiny.

The AI governance conversation in Washington and Brussels is accelerating. Frameworks like this one will be cited in those debates — by industry lobbyists as evidence that self-regulation works, and by critics as evidence of how much definitional territory companies have already claimed.

Both readings are correct. The question is which one drives policy.