A note on process: I used an AI writing assistant (Claude) to help articulate and refine this post through multiple drafts. The ideas, arguments, framework, and lived experience here are entirely my own. Every sentence reflects my thinking, beliefs, actions and my theory of change.
I spent twenty years building institutional trust before I ever heard the term AI safety. Not trust in the abstract, but the kind that lives in community meetings, union halls, neighborhood associations, and working-class communities across the United States. The kind that takes years to earn and minutes to lose, and one where your word is the only currency that matters.
Then a year ago, I pivoted into AI safety and I found something that troubled me in a specific way. The communities I had spent two decades working inside were nowhere near the rooms where the response was being shaped and decisions were being made. As I immersed myself in the research and analysis, a career mentor asked me to think about my contribution to the ecosystem and develop a theory of change. I noticed a gap between the communities I had been working in and every room I was now walking into. The communities most vulnerable to AI-driven harm were absent from every conversation about the response. I did not see anyone building the pathway to bring them in and that absence is what I built Hokmah to address.
I want to be clear from the start about who I am and what I am not. I am a founder and an infrastructure builder, not a researcher. I came to this work with deep experience in constituency building, institutional trust, program management, communications and civic organizing, now combined with formal AI safety training through BlueDot Impact, the Successif AI Policy Strategy Fellowship, CEA Bootcamp, the EA community, and EA Global 2026. My edge is not technical, though I am learning, at my own pace. However, it is the ability to build civic infrastructure and community trust in practice.
The assumption underneath every disclosure regime
Every major AI transparency initiative in the world, from the EU AI Act to platform labeling policies to nearly every disclosure provision before Congress, rests on a single untested assumption: that telling people a system is synthetic protects them, I believe that assumption is wrong.
The space between disclosure and actual protection is not a communication failure. It is the mechanism through which trust, belief, and agency migrate toward systems that did not earn them and were never meaningfully consented to. I refer to this as the comprehension gap. In most cases it is structurally invisible within current regulatory frameworks. That invisibility is not incidental, and it is where power concentrates.
When AI systems can reshape belief and erode civic trust at scale in communities that lack the resources, literacy, or representation to recognize or resist it, that is not a consumer harm. It is a structural condition that makes unchecked AI power concentration more durable and more difficult to reverse. The comprehension gap is a longtermism problem, not only a governance one.
The scope signal we should not ignore
In May 2026, Pope Leo XIV released his first encyclical dedicated specifically to artificial intelligence. He opened it with the Tower of Babel and named explicitly the danger of technological concentration without moral accountability. I am not making a theological argument. I am making a scope argument. When one of the world's oldest continuously operating moral institutions enters this conversation with that kind of urgency, it tells you something about the communities already sensing the threat with no framework yet for naming it. Those communities are not peripheral to this moment. They are its core constituency.
Faith institutions reach more than 65 percent of Americans through trusted messenger networks built over generations. That infrastructure cannot be manufactured, purchased, or replicated by any lab, funder, or policy shop. It is almost entirely absent from AI governance conversations. That is a structural problem, and it is the one I am working to solve.
What Hokmah is building
Hokmah, drawn from the Hebrew word for wisdom and the figure in Proverbs 8 who stands at the city gates calling out to everyone, is building the constituency infrastructure that meaningful AI governance requires. Not an awareness campaign. Not a literacy initiative. The civic capacity, trusted messenger architecture, and community-scale evidence base that will allow affected communities to recognize synthetic persuasion and hold meaningful power over the systems entering their lives.
The core of this work is a methodology I developed and stress-tested through ongoing conversations with researchers and fellow cohort members across CEA and BlueDot Impact, grounded in research-ethics principles, to examine a specific question: does comprehension protect? Not whether disclosure exists. Whether it produces the understanding current law assumes it does.
The legislative destination is Community Consent. A framework requiring AI systems to demonstrate meaningful comprehension, not merely technical disclosure, before engaging communities of faith and vulnerable populations at scale. The precedents exist: informed consent in medicine, plain-language requirements in consumer finance, FDA labeling efficacy standards. AI is the only high-stakes domain where disclosure theater satisfies the legal standard, and that will not hold.
Where I am now
I am in active pre-launch, building the online infrastructure and research foundation before scaling into community partnerships. I deliberately slowed the build to stress-test the methodology. That is where I am, and I consider that discipline part of the work.
What I am looking for
I am here because this community takes these questions seriously. I am specifically interested in connecting with people working on AI persuasion and manipulation at community or population scale, empirical research on disclosure efficacy, power concentration as a governance mechanism, and Community Consent or informed consent frameworks applied to AI systems.
I am also genuinely interested in the hardest objections to this work. I would rather find the places where the argument does not hold now rather than later.
If this resonates, I would be glad to hear from you. This is my first EA post and I am facing a fear by sharing with the community.
Karen Maria Alston Founder, Hokmah wearehokmah.org