A Principle From the Wrong Direction

The Haudenosaunee Confederacy, which influenced the political thinking of Benjamin Franklin and several American founders, operated according to a principle that their oral tradition describes as considering the effects of decisions on seven generations into the future. The principle is sometimes cited in environmental contexts and occasionally in technology ethics discussions, usually as a poetic illustration of long-term thinking. What gets lost in most citations is the specific character of the obligation. Seven-generation thinking is not just about thinking further ahead. It is about accepting responsibility for consequences you will not live to see and that will be experienced by people who have no voice in your decision. It is about the moral weight of what you choose to hand forward. Applied to AI development, this framing changes the question considerably.

What Seven Generations Actually Means

If a generation is roughly twenty-five years, seven generations is a hundred and seventy-five years. Nobody building AI systems today will live to see the world of 2200. Neither will anyone currently living anywhere. The people who will inhabit that world are not consulted in any of the decisions being made now. They have no representatives at regulatory hearings, no voices in corporate governance meetings, no ability to resist choices that close off their options. This is, of course, true of all consequential decisions made in the present. The people who built industrial systems in the 19th century did not consult the communities that would live downstream of their pollution. The people who made land use decisions in the early 20th century did not consult the cities that would struggle with the consequences in the 21st. But AI is different from those precedents in one specific way: it is a technology that may shape the conditions of its own development going forward. The AI systems built now influence what AI systems will be built next. The values and norms embedded now may propagate and compound. The path dependencies established in the early period may be very difficult to alter later.

The Compounding Problem

Compounding is straightforward in finance: small differences in returns accumulate into large differences in outcomes over time. The same logic applies to value alignment in AI development. A small bias in how current AI systems represent certain groups, evaluate certain kinds of knowledge, or weight certain considerations may seem marginal today. Over successive generations of AI development, where new systems are trained partly on outputs of current systems, that bias compounds. By the time it is obvious, it is deeply embedded. Research from Oxford's Future of Humanity Institute examining value propagation in AI development over successive model generations has found that errors in value representation in early training data propagate and often amplify through subsequent generations of models trained on synthetic or AI-augmented data. The implication is that choices made in the early period of AI development have disproportionate downstream influence — exactly what seven-generation thinking would predict.

The Tangent: What Olmstead Designed For

Frederick Law Olmstead, who designed Central Park in the 1850s, planted trees in configurations that would take fifty to a hundred years to mature. He was designing for people who had not yet been born. His plans included maintenance schedules and usage patterns anticipated for generations of New Yorkers he would never meet. The park that exists today reflects choices made by someone who explicitly understood his role as a trustee for future users, not just a provider for current ones. This orientation — trustee for the future rather than provider for the present — is almost entirely absent from AI development culture. The timescale of product development, the quarterly earnings cycle, and the competitive dynamics of the AI sector all enforce a present orientation. The people making consequential design choices are evaluated on what those choices produce now.

What Institutions Would Need to Change

Taking seven-generation thinking seriously in AI development would require institutional changes that run against current incentives. It would require incorporating long-term consequence modeling into design review processes — not as compliance theater but as genuine decision input. It would require governance mechanisms that give weight to interests of future users alongside interests of current users and shareholders. Some jurisdictions are beginning to experiment with long-term impact assessment requirements for AI systems. Research from the Center for the Governance of AI examining mandatory impact assessment frameworks across jurisdictions has found that requirements that specify multi-decade consequence modeling produce substantively different design decisions from requirements that focus only on immediate impacts. The longer the mandated horizon, the more design choices shift away from pure engagement optimization and toward resilience and reversibility.

The Obligation Worth Accepting

Seven-generation thinking does not require certainty about what the world of 2200 will need. It requires epistemic humility about the limits of current knowledge, preference for reversible choices over irreversible ones, and genuine concern for consequences that will be experienced by people who cannot speak for themselves in current decisions. These are orientations that require deliberate cultivation. They run against the natural human tendency to heavily discount future costs relative to present benefits. But they represent exactly the kind of moral seriousness that a technology with AI's potential for long-run consequence demands.