I want to talk about a pattern I am watching repeat itself across the artificial intelligence field, and I want to talk about it carefully, because the people who are repeating the pattern are not bad people. They are smart, well-resourced, and well-intentioned actors who genuinely believe they are solving a problem, and they are using a program design that has failed before, in the same shape, for the same reasons, in a closely related domain. They have not gone back to study what happened the last time the design was tried at scale, and they are about to spend the next decade learning the same lessons the social impact sector spent thirty years learning, inside a domain where the consequences of getting it wrong are significantly larger than they were the first time.

The pattern is this. A high-stakes field experiences a talent gap. The institutions inside the field decide to address the gap by recruiting young, smart, high-credentialed people from elite universities, putting them through a short training program, and dropping them into roles that have historically required years of practiced expertise, deep institutional fluency, and the kind of stakeholder navigation skills that can only be built by being in the rooms over time. The young, smart, high-credentialed people produce some good work, attract significant media attention, and exit the program after a fixed period to go pursue the next thing. The field reports the program as a success. The communities the program was supposed to serve report something different.

That pattern was the original design of Teach for America, which launched in 1990 with the premise that recent college graduates from elite universities could fill teacher shortages in under-resourced schools with five weeks of summer training, a two-year commitment, and the energy of youth. The program has now contracted significantly from its peak, having shrunk by nearly two-thirds from its high a decade ago. Across more than three decades of operation, the program has been documented by Brookings, the National Education Association, peer-reviewed studies in the American Political Science Review, and a range of investigative outlets to have produced higher staff turnover in the communities it served, inadequately prepared teachers placed in the most challenging classrooms, and a sustained alignment between the program's funders and a particular set of education policy positions, including charter expansion, weakened union protections, and standardized testing as the dominant evaluation tool. The pattern is not contested. The pattern is established.

The same design logic shaped City Year, AmeriCorps in many of its placements, and a significant portion of the capacity-building fellowship infrastructure that the social impact sector has built over the last three decades. Every time the model has been deployed at scale, the same critique has emerged. Young people without institutional fluency are placed into rooms where institutional fluency is what the work actually requires. The young people learn quickly, often impressively, but the learning happens at the expense of the communities the program was supposed to serve, because the communities are being asked to absorb the cost of someone else's training period. The institutional layer running the program reports the program as a success because the program's metrics are defined by the institution, while the communities most affected by the program have no standing to define what success looks like from their position.

I am telling you about this old pattern because the AI field is now reproducing it in a higher-stakes context, and almost nobody is naming the resemblance.

Anthropic, the AI company, currently runs a program called the Anthropic Fellows Program. The program is structured as a four-month research fellowship focused on technical AI safety research, with a weekly stipend of $3,850, compute funding of approximately $15,000 per month, and direct mentorship from Anthropic researchers. The published eligibility criteria are explicit. The program is looking for technical professionals with backgrounds in physics, mathematics, computer science, or cybersecurity. The program states that strong candidates have strong Python skills, the ability to make concrete progress on ambiguous technical problems, and the motivation to transition into empirical AI safety research, including the possibility of joining Anthropic full time afterward. The program is built around the goal of producing a public research output, typically a paper, by the end of the four months. The published materials state that more than forty percent of fellows in the first cohort subsequently joined Anthropic full time.

I want to be careful, because the people inside the program are doing real work, and the research outputs are real research outputs. The mentorship structure is serious. The compensation is meaningfully higher than the nonprofit fellowship model I will describe in a moment. The intent is sound. The published research has included contributions to mechanistic interpretability, model organisms of misalignment, scalable oversight, and adversarial robustness, and the public release of some of this work has been useful to the broader research community.

The problem is that the program design assumes that the most important contribution the field needs right now is technical AI safety research produced by early-career or career-transitioning technical talent who can pick up new skills quickly and ship a paper in four months. That assumption is the same assumption Teach for America made about teaching. The assumption is that the work is fundamentally a technical skills problem, and that smart people with the right training program can address it. The assumption is wrong, and the assumption is wrong for the same reason it was wrong about teaching.

The AI field does not actually have a technical skills shortage at the level the fellowship programs are designed to address. The AI field has an institutional governance shortage. The AI field has a stakeholder navigation shortage. The AI field has a community accountability shortage. The AI field has a senior expertise shortage in exactly the areas where the technology is touching down inside the lives of actual people, in actual schools, hospitals, public agencies, social service organizations, refugee response systems, child welfare systems, and the institutions that hold these communities together. The people who can do that work are not early-career technical professionals. The people who can do that work are senior practitioners with twenty-plus years of institutional fluency, who have built infrastructure inside the institutions the technology is now being deployed against, who have spent careers building the trust and the protocols and the systems that AI is now showing up to disrupt without consent.

Those people exist. They are not being recruited into the fellowship programs that the AI companies are running. The fellowship programs are being designed for a different demographic, a different career stage, and a different theory of what the field needs, and the result is going to be the same result that Teach for America produced. Smart young people produce papers, get hired at the company that ran the program, attract media coverage about how the program is filling the gap, and then the communities the technology is being deployed against are left to absorb the consequences of decisions made by people who do not know the institutions, do not know the communities, and have not built the stakeholder relationships that would have given them the standing to make those decisions in the first place.

That is the structural critique. I want to name it precisely, because the people running these programs are about to spend the next decade making the same mistake the social impact sector spent the last three decades making, and the AI version of the mistake is going to be significantly more consequential, because the systems being built and deployed have far more reach than a single classroom and operate with significantly less oversight.

I want to be clear about what I am not saying, because the people defending these programs are going to misread the critique if I am not precise.

I am not saying young people should not be in AI safety research. They should be. They bring fresh thinking, technical fluency with current systems, and the willingness to question assumptions that older practitioners have stopped questioning. That contribution is real and necessary. I am not saying technical AI safety research is not important. It is. The work the Anthropic Fellows are producing on mechanistic interpretability, scalable oversight, and adversarial robustness is serious technical work that the field needs.

What I am saying is that the design of the program excludes the expertise the field most urgently needs, and the exclusion is structural rather than incidental. The program states it is looking for people transitioning into AI safety research. It is not looking for senior practitioners bringing twenty years of institutional governance experience into the conversation. The eligibility criteria, the technical bias, the four-month duration, and the conversion-to-full-time-research-role pathway all signal that the program is designed for one career stage and one type of expertise, and a different career stage and a different type of expertise are being filtered out before the application is even submitted.

There is a different program design that would address the gap the field actually has. It would recruit senior practitioners with deep institutional experience inside the sectors where AI is being deployed. It would pay them at the level their senior expertise commands, rather than at the level of an entry-level stipend. It would structure their engagement around the actual problem, which is not producing a paper but building the governance infrastructure that lets affected communities exercise authority over the systems being deployed against them. It would treat their decades of institutional fluency as the asset, rather than asking them to acquire Python skills as the prerequisite. It would build their work into the public record in a way that compounds across the field, rather than disappearing into a four-month deliverable.

That program does not currently exist at any of the major AI labs running fellowships, and the parallel infrastructure that does exist for senior practitioners, including various nonprofit capacity-building fellowships across the social impact sector that match experienced professionals with mission-driven organizations, is structured around significantly lower compensation than the AI lab fellowships pay early-career technical talent. A senior practitioner with twenty years of expertise can access a nonprofit fellowship that pays roughly twenty-five thousand dollars for a thousand hours of work, or twenty-five dollars per hour, in placement with a single organization for six to twelve months. The same field is paying early-career technical fellows roughly four times that hourly rate to produce research papers at a major AI lab. The math tells you everything about which expertise the field is choosing to invest in.

The translation layer thesis I have been building over the last several years is partly a response to this gap. The translation layer is the governance infrastructure that enables communities most affected by AI systems to exercise actual decision-making authority over the development, evaluation, and deployment of those systems. Building the translation layer requires senior institutional expertise. It cannot be built by people who are still learning the institutions. The work of building it cannot be compressed into a four-month fellowship that ends with a paper. The work of building it is what experienced practitioners do over years, inside complex institutions, with the trust and the patience and the longitudinal vantage that only a long career can produce.

The field knows this. The field has not yet acted on what it knows.

The question worth asking is not whether the Anthropic Fellows Program should exist. It should. Technical AI safety research is genuinely important, and the program is doing some of it. The question worth asking is whether the program design represents the totality of what the field needs, or whether the program design is a partial answer being treated as a complete one. The answer is clearly the latter, and the gap between what the program addresses and what the field actually needs is being filled, right now, by senior practitioners working without institutional support, without structured compensation, without mentorship infrastructure, and without the platform that the AI labs are providing exclusively to a different demographic.

That is the imbalance I want the field to name. The social impact sector spent thirty years learning the lessons that came from making this mistake at scale. The AI field has the opportunity to study what happened and design something better. The question is whether the institutions running the programs are interested in doing that work, or whether they are committed to the program design they already have because the design serves their pipeline more than it serves the field.

I think it is the latter. I am also willing to be wrong about that, and I am writing this piece publicly because I want to be told if I am, by anyone running one of these programs who can show me the design choices I have not seen and the senior practitioners I have not noticed being recruited.

What I want for the people reading this is the same thing I want for the institutions running these programs. I want all of us to study what happened the last time this model was tried at scale, name the lessons that the social impact sector spent thirty years learning, and refuse to spend the next decade repeating them inside a domain where the consequences of getting it wrong are going to be considerably larger than they were the last time around.