The word disruption has stopped meaning anything.

It was a useful word, once. Clayton Christensen used it, in the 1990s, to describe a specific phenomenon — the way incumbent businesses could be displaced by new entrants that started at the low end of the market and moved upward, on a different cost structure, with a different business model. The word was precise. The word described something real.

The word was then adopted, in the 2010s, by every technology company in Silicon Valley, in the precise language of marketing, to mean “we are going to make a lot of money by replacing something that currently exists, and we want you to feel good about it.” The word was stretched. The word was diluted. The word became, in the precise language of the situation, a marketing artifact, and the meaning drained out of it.

The same thing happened to transformation. The word was useful, once. It described the work of taking an institution apart at the seams and rebuilding it for a different world. The word was adopted, in the 2010s, by every consulting firm in the world, to mean “we are going to charge you a lot of money to install software that will not work, and we want you to feel good about it.” The word was stretched. The word was diluted. The word became a marketing artifact, and the meaning drained out of it.

The same thing is now happening to AI.

The word AI is being adopted, in the 2020s, by every technology company in the world, to mean “we are going to make a lot of money by automating something that currently exists, and we want you to feel good about it.” The word is being stretched. The word is being diluted. The word is becoming, in the precise language of the situation, a marketing artifact, and the meaning is draining out of it.

I want to propose, in this post, a different word. The word is Regenerative AI. The word is, I believe, the right word for what AI should be, in the second half of this decade and the first half of the next. The word is the foundation of a category I have been building, in my work and in the book I am writing on AI sovereignty for Southeast Asia, for the last three years.

This is the manifesto. This is the category-creation piece. This is the post I hope will be the one that people quote, when they want to explain what I am for.

What Regenerative Means

The word regenerative comes from ecology. It describes systems that do not merely sustain themselves, but actively restore, renew, and rebuild the ecosystems in which they operate. A regenerative farm does not just avoid depleting the soil. A regenerative farm increases the fertility of the soil, the biodiversity of the surrounding ecosystem, and the resilience of the community that depends on the farm. A regenerative economy does not just minimize harm. A regenerative economy produces outcomes that are, over time, measurably better than the outcomes the economy was producing before.

The word is not, in the precise language of ecology, a synonym for “sustainable.” Sustainability is the lower bar. A sustainable system maintains itself. A regenerative system improves the system it is part of.

I want to propose, in this manifesto, that the same distinction applies to AI.

A sustainable AI is an AI that minimizes harm. It is trained on data that was ethically sourced. It is deployed in a way that does not displace workers. It is governed by principles that respect the autonomy of the people it affects. Sustainable AI is, in the precise language of the situation, the responsible AI that the major technology companies have been promising, in their responsible AI reports, for the last five years. Sustainable AI is necessary. Sustainable AI is not sufficient.

A Regenerative AI is an AI that actively improves the system it is part of. It is trained on data that strengthens the cultural identity of the communities that produced it. It is deployed in a way that increases the capacity, the agency, and the economic power of the people it serves. It is governed by the communities it affects, in a language the communities understand, under institutions the communities own.

Regenerative AI does not just avoid extracting value from a community. Regenerative AI produces value that circulates within the community, in a way that strengthens the community’s capacity to produce more value, in a virtuous cycle that compounds over time.

This is the category. This is the word. This is the standard against which I want AI, in the second half of this decade, to be measured.

The Four Tests for Regenerative AI

I have been developing, in my work with institutions across Southeast Asia, a set of four tests that any AI system can be put through, to determine whether it is regenerative or merely sustainable. The tests are not, in the precise language of the situation, perfect. They are, in the precise language of the work I have been doing, useful.

Test One: Does the AI increase local capacity, or does it replace local capacity? A regenerative AI is one that, when deployed in a community, leaves the community more capable than it found it. A teacher who uses AI to produce lesson plans faster is not, by this test, using regenerative AI, unless the AI is also teaching the teacher to produce better lesson plans independently. A farmer who uses AI to optimize irrigation is not, by this test, using regenerative AI, unless the AI is also teaching the farmer to understand the soil in ways the farmer could not understand before. The distinction is the distinction between a tool and a teacher. Regenerative AI is, by this test, a teacher.

Test Two: Does the AI circulate value within the community, or does it extract value to an external owner? A regenerative AI is one that, when deployed in a community, leaves more economic value in the community than it takes out. The data cooperative model, in which the data of 70 million Indonesian UMKM is aggregated, governed by the members, and used to commission models that serve the cooperative’s interests, is regenerative AI by this test. The platform model, in which the data of 70 million Indonesian UMKM is harvested, governed by a foreign shareholder, and used to train models that serve the platform’s interests, is not regenerative AI by this test. The distinction is the distinction between ownership and tenancy. Regenerative AI is, by this test, owned.

Test Three: Does the AI strengthen cultural identity, or does it homogenize? A regenerative AI is one that, when deployed in a community, leaves the community’s cultural identity stronger than it found it. A model that knows the difference between a Javanese rice terrace and a Thai one, that knows that a Filipino grandmother cooks sinigang differently from a Visayan grandmother, that knows the quê hương of a Mekong Delta village in the language the village actually uses — that model is regenerative AI by this test. A model that produces, in response to any prompt about a Southeast Asian village, the same generic image of “the Orient” — that model is not regenerative AI by this test. The distinction is the distinction between representation and erasure. Regenerative AI is, by this test, representative.

Test Four: Does the AI increase the agency’s autonomy, or does it increase the system’s dependence? A regenerative AI is one that, when deployed in a community, leaves the community more capable of acting independently than it found it. A student who uses AI to learn the material is, by this test, using regenerative AI. A student who uses AI to replace the work of learning the material is not. A doctor who uses AI to make a better diagnosis is, by this test, using regenerative AI. A doctor who uses AI to replace the work of making a diagnosis is not. The distinction is the distinction between a tool that increases the user’s capacity and a tool that decreases it. Regenerative AI is, by this test, capacity-increasing.

A model that passes all four tests is regenerative. A model that passes one or two is sustainable. A model that passes none is extractive. The category I am proposing is the regenerative category. The work of the next decade, in my view, is to move as many models as possible, in as many contexts as possible, up the gradient from extractive to sustainable to regenerative.

What Regenerative AI Looks Like in Practice

I want to make this concrete, because abstract manifestos are easy to write and useless to implement. Here are three examples, from the work I have been doing, of what Regenerative AI looks like in practice.

The first example is the data cooperative. A data cooperative, in the precise language of the model I have been developing, is a member-owned institution that aggregates the data of smallholders, governs the data under the members’ control, and uses the data to commission AI models that serve the members’ interests. The 70 million Indonesian UMKM, organized into a cooperative structure, can produce a training dataset worth billions of dollars. The dataset, governed by the cooperative, can be used to train models that help the members optimize inventory, predict demand, access credit, and navigate regulatory complexity. The value circulates within the cooperative. The capacity of the members increases. The cultural identity of the communities the members serve is preserved. The members’ agency is increased, not decreased. The data cooperative is, by all four tests, regenerative AI.

The second example is the sovereign model repository. A sovereign model repository, in the precise language of the model I have been developing, is a regional institution that hosts open-weight language models trained on Southeast Asian data, governed by Southeast Asian institutions, available to Southeast Asian developers at the cost of inference. The ASEAN AI Commons, which I have proposed elsewhere as a regional cooperation framework, is a sovereign model repository. The models, trained on the linguistic and cultural diversity of the region, are regenerative by all four tests. The value circulates within the region. The capacity of the region’s developers increases. The cultural identity of the region’s communities is preserved. The agency’s autonomy of the region’s institutions is increased, not decreased.

The third example is the pedagogical AI tutor. A pedagogical AI tutor, in the precise language of the model I have been developing, is an AI system designed not to replace the teacher, but to increase the teacher’s capacity. The tutor does not produce lesson plans. The tutor produces questions that the teacher can use to provoke the student’s thinking. The tutor does not grade the student’s work. The tutor produces feedback that helps the student understand what the work is for. The tutor does not replace the teacher’s relationship with the student. The tutor helps the teacher understand the student better, in the time the teacher has, in the context the teacher is in. The pedagogical AI tutor is, by all four tests, regenerative AI.

These three examples are not, in the precise language of the situation, the only examples. They are, in the precise language of the work I have been doing, the examples I have found most useful, in conversations with institutions across the region, when the conversation turns from “what is Regenerative AI” to “what does it look like in our context.”

Why This Category, Why Now

The category-creation work I am doing with Regenerative AI is, I believe, the work of this moment, for three reasons.

The first reason is that the existing categories — disruption, transformation, even responsible AI — are exhausted. The words have stopped meaning anything. The institutions, the policymakers, the technologists, the citizens who are trying to think clearly about what AI should be are running out of vocabulary. The vocabulary has to be renewed. Regenerative AI is, I believe, the right renewal.

The second reason is that the cost of getting the category wrong is asymmetric. If we adopt the category of disruption, the AI we build will disrupt. If we adopt the category of transformation, the AI we build will transform. If we adopt the category of regenerative, the AI we build will regenerate. The category is not a label. The category is a commitment. The category shapes the institutions, the policies, the investments, the careers, the lives that get built. The category is, in the precise language of the situation, the most consequential decision any of us will make about AI, in the next five years.

The third reason is that the alternative is extractive AI by default. If we do not, deliberately, build regenerative AI, we will, by default, build extractive AI. The platforms are already built. The models are already trained. The data colonization is already underway. The default future is, in the precise language of the situation, a future in which 675 million people in Southeast Asia are tenants in someone else’s digital building, paying rent in data for the privilege of being “digitized.” Regenerative AI is the alternative. Regenerative AI is the work.

A Letter to the People Building AI

I want to close this manifesto with a direct address to the people who are building AI, in the precise language of the situation, in every institution, in every country, in every sector.

You are, in the language of the moment, the most consequential workforce in human history. The systems you build will, in the next decade, shape the cognitive landscape of every human being on the planet. The systems you build will either regenerate the communities they touch, or they will extract from them. The systems you build will either increase the agency’s autonomy of the people they serve, or they will decrease it. The systems you build will either strengthen the cultural identity of the communities they are part of, or they will homogenize them.

You do not get to be neutral. The systems you build are not neutral. The systems you build are, in the precise language of the situation, political acts. The systems you build will, in the next decade, be evaluated, in retrospect, by the people they affect, in the language those people use, under the standards those people hold.

Build regenerative AI. Build AI that increases local capacity. Build AI that circulates value within the community. Build AI that strengthens cultural identity. Build AI that increases the user’s agency. Build AI that the community can govern, in a language the community understands, under institutions the community owns.

Build the AI that the world, in twenty years, will be grateful you built.

The category is open. The work is starting. The word is regenerative.

I will defend it, in every conversation I have, for as long as I am in this work.