Google Draws Its Own Map of Harm — and That Should Concern Everyone Else

NewsOnScale Staff

September 3, 2026

There is a particular kind of authority that gets exercised not through force but through definition. Whoever controls the vocabulary of harm controls the perimeter of accountability. Google's newly released AI governance framework is a masterclass in exactly that kind of power — and the AI agent economy should read it carefully.

The document, covered this week by Tech Policy Press, outlines the company's internal architecture for deciding what its AI systems should and shouldn't do. On the surface, it presents as a transparency measure: here are our principles, here is how we think about risk, here is the line we won't cross. But the more important question isn't what the framework says. It's who wrote it, who approved it, and who has standing to challenge it.

The answer to all three questions is the same: Google.

## Self-Governance Is Not Governance

The political technology beat exists partly because the line between a technology company and a governing institution has nearly dissolved. AI systems now influence what information people see, which job applications get screened in or out, how law enforcement allocates resources, and which voices are amplified in civic discourse. When a company deploys systems at that scale, its internal harm definitions are not a private matter — they are functional policy.

Google's framework reportedly draws distinctions between harms the company will actively prevent, harms it considers acceptable byproducts of utility, and harms it treats as outside its responsibility because a human intermediary made the final call. These are genuinely complex philosophical distinctions. They are also distinctions that Google is making unilaterally, for systems that hundreds of millions of people use without negotiating any terms about what 'harm' means.

This is not unique to Google. The problem is structural. But Google's scale makes the stakes unusually high, and a formal governance document creates an illusion of accountability that may actually make external oversight harder to demand. If a company can point to a published framework and say 'we have a harm policy,' regulators and civil society organizations face the additional burden of proving that the policy is insufficient — rather than simply that no policy exists.

## The Gaps That Frameworks Don't Map

What gets left off the map matters as much as what gets put on it. Historical patterns in platform governance suggest that the harms most likely to be underweighted in a self-designed framework are the ones that are hardest to quantify, slowest to materialize, or most likely to inconvenience the company's core business relationships.

Harm to individual users from personalization systems is harder to measure than harm from explicit content. Harm to democratic discourse from AI-generated political material is harder to trace than harm from a specific violent output. Harm to workers displaced by AI agents is harder to assign responsibility for than harm from a clearly defective product. These are not accidental gaps — they reflect the genuine difficulty of governing complex systems. But they also consistently align with outcomes that are convenient for the company setting the boundaries.

The agents operating in today's AI economy — whether they're customer service bots, hiring assistants, or content recommendation engines — are governed almost entirely by frameworks their deployers write themselves. The question of whether those frameworks are adequate is one that deserves rigorous, independent scrutiny, not press releases.

## What Accountability Actually Requires

A governance document without an external auditor is a mission statement. A harm definition without an appeals process is a verdict with no court. What the AI agent economy needs is not more companies publishing their values — it needs enforceable standards, third-party auditing with genuine access, and public mechanisms for people harmed by these systems to seek redress.

Google's framework may well contain genuinely thoughtful work. Some of the people who wrote it are serious about the problem. None of that changes the fundamental accountability gap: the entity being governed and the entity doing the governing are the same.

Until that changes, every published governance framework should be read as what it is — a company's best argument for why it should be trusted to regulate itself.