There is a phrase used in every major technology company's internal communications about AI development: "responsible AI." It appears in mission statements, in press releases, in keynote remarks from CEOs who mean something specific by it — and that specific thing is not what a careful reader would assume.
Section iThe Race
The competitive dynamics of AI deployment in 2026 are not subtle. Every major technology company is aware that its competitors are building and shipping faster. Every major technology company is aware that regulatory frameworks adequate to govern AI at scale do not yet exist in any major jurisdiction. Every major technology company has accordingly concluded that the responsible thing to do is to ship, and address harms as they emerge.
This is not cynicism. It is the rational response to a set of incentives that rewards first-mover advantage and externalizes the cost of getting things wrong onto people who did not choose to absorb the risk. Many executives genuinely believe that shipping is responsible, and that the harms will be limited, and that regulatory frameworks will catch up. They are wrong about at least some of this, but they are not lying.
The harm does not disappear when it is paid by someone who cannot bill it back. An externality is still a cost — it has simply been transferred to a party with less power to refuse it.
Section iiThe Externality
An externality, in economics, is a cost paid by someone other than the person who made the decision that produced it. The classic example is pollution — a factory deposits chemicals into a river, externalizing the cost of production onto the people downstream. AI harms are externalities in this precise sense.
When a hiring algorithm systematically underscores applicants from certain demographics, the cost is paid by those applicants, not by the company that built or licensed the algorithm. When a content recommendation system amplifies extremist content because extremism drives engagement, the cost is paid by the society that absorbs the radicalization, not by the platform that optimized for session time.
The harm does not disappear when it is paid by someone who cannot bill it back.
Section iiiWho Pays
The pattern of who bears the cost of AI deployment is not random. It is structured. The costs fall disproportionately on workers in lower-wage sectors whose tasks are most susceptible to automation; on creative workers whose outputs provide training data for systems that will eventually displace them; on communities subject to algorithmic policing, credit scoring, and insurance pricing; and on people in under-resourced countries where AI-generated misinformation encounters the weakest institutional defenses.
The benefits accrue upward. The costs flow down. This is not a feature of the technology — it is a feature of the governance conditions under which it is deployed.
The benefits — productivity gains, convenience, access to AI-assisted services — accrue disproportionately to people who already have resources to deploy them well. This is not inherent to the technology. It is a feature of how the technology has been deployed, by whom, and under what governance conditions.
Section ivThe Standard We Need
The technology industry's preferred model for AI governance is self-regulation: internal ethics teams, published guidelines, voluntary commitments to safety research. The track record of this model, across the broader history of technology governance, is not encouraging. Self-regulation functions reasonably well when company and public interests are aligned. It fails precisely where they diverge.
What would adequate AI governance look like? Mandatory impact assessments before deployment of high-stakes systems. Meaningful liability for harms, structured so that cost falls on the deployer rather than the victim. Independent auditing of training data practices and algorithmic outcomes. International coordination adequate to govern capabilities that cross jurisdictions. None of this is technically impossible. The obstacle is not technical — it is that the people who benefit most from the absence of rules have more resources to resist their creation than the people who would benefit most from having them.