The New York City Department of Education was supposed to release final guidance on artificial intelligence in public schools by the end of June. The draft guidance came out in March, used a traffic light framework that mapped what teachers could and could not do with AI tools inside their classrooms, and was met with nearly 6,500 public comments, packed public meetings, and a coalition of parents, teachers, students, and artists rallying outside City Hall demanding a two-year moratorium on generative AI in city schools. More than half of City Council members have signed a letter urging Mayor Mamdani and Chancellor Samuels to pause AI use in schools entirely. The Public Advocate joined the rally. A broader petition seeking a moratorium has collected thousands of signatures.

In response to all of this, the Department announced on Wednesday that the final guidance is being delayed.

The news coverage is calling it a delay and the official statements are calling it a course correction. What it actually is, when you read the situation through the right frame, is the institution discovering after the fact that it tried to make a policy decision about families without the families being in the room. That distinction matters, because the same dynamic is playing out in school districts across the country, and what New York decides over the next twelve months is going to set a pattern that other districts are going to follow whether they admit it or not.

I want to walk through what is actually happening here, because the structural problem on display in New York is the same structural problem AI governance is wrestling with in every sector, and the framework I have spent the last several years building, which I call the translation layer, was developed precisely to give institutions a way to navigate situations like this one without ending up at the rally outside City Hall.

The previous administration, under Chancellor David Banks, was warm to AI. Banks publicly said that the technology could revolutionize how the school system functions, and he allowed schools to experiment widely without releasing a clear policy. The absence of policy created a vacuum that individual schools filled with their own decisions, which meant AI tools were already operating across the city before any central guidance existed. When the current administration came in, they inherited an installed base of AI tools that families had never been asked to weigh in on, deployed in classrooms whose teachers had not been trained to evaluate them, surfacing in homework and grading and student counseling and IEP development in ways that nobody at the central office had a comprehensive view of. The draft policy from March was the first attempt to retroactively impose a framework on a deployment that was already underway. Families finally responded with the volume of a rally at City Hall because they had not been asked anything to begin with, and they had figured out that the policy being drafted was not actually going to give them authority to stop what was already happening in their children's classrooms.

That is not a process failure. That is the structural problem the translation layer was developed to name.

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. Existing AI governance frameworks frequently conflate participation with power, and the translation layer distinguishes between the two. What happened in New York City is what happens when participation is performed without power being transferred. Families were given 6,500 comment slots on a policy that had already been drafted. They were not given the authority to stop deployment before it started, because the deployment had been started without their knowledge. They were not given the authority to require the institution to demonstrate that the tools were safe before adoption, because the tools were already adopted. They were given the option to comment on what had already been done. That is the difference between participation and power, and that is the gap the parents standing outside City Hall on Wednesday were trying to close with their bodies because the institution had not built the infrastructure that would have allowed them to close it through the process.

Three components define the translation layer in operational terms, and applying them to the New York situation surfaces what should have happened and what still could.

The first component is participation infrastructure. Participation infrastructure means standing bodies with reserved authority for affected communities, and the distinction matters because participation infrastructure is not advisory panels, not consultation processes, not town halls. In a school district, participation infrastructure means a standing parent body with reserved decision-making power over technology adoption inside classrooms. Reserved decision-making power means that if the parent body says no, deployment does not move forward, even if the central office or the vendor or the chancellor wants it to. That is the distinction the field has been avoiding for years, because building that kind of authority requires the institution to give up power it currently holds, and the institution has consistently been unwilling to do that. The 6,500 public comments the city received were the system telling the institution, after the fact, that the participation infrastructure should have existed before any tool went live.

The second component is evidence redefinition. When a parent at the rally on Wednesday described an app being used in her son's school that records when he goes to the bathroom and how much sleep he gets and his mood and photos of him, what that parent was bringing was not a story. It was data. It was a documented use case showing what the technology actually does inside the classroom, gathered by the person closest to the deployment and the person most accountable for the child's safety. The translation layer treats that testimony as evidence with the same weight as any vendor's white paper, any superintendent's risk assessment, any consultant's audit report. Existing AI governance frameworks treat that testimony as input. The distinction sounds subtle. It is not. Input is something an institution receives. Evidence is something an institution must respond to.

The third component is enforcement authority. Enforcement authority means real power over evaluation metrics, equity constraints, and the timing and conditions of deployment, including the authority to pause or decommission tools after they are already running. In the New York case, no enforcement authority existed at the family level before the protest began. The only enforcement authority that materialized was political pressure expressed through a rally and a Council letter, which is to say, the families had to manufacture enforcement authority through public mobilization because the institution had not built it into the governance structure. Manufactured enforcement authority is exhausting and unstable, because it depends on the affected community having the capacity to organize on top of everything else they are already doing. Built-in enforcement authority is sustainable, because the affected community already has standing when an issue arises.

Without all three of those components operating together, the translation layer does not exist, and what is happening in a school district is not governance. It is the management of a public relations problem that becomes governance only when the affected community gets loud enough to force the conversation. New York has now reached the point where the conversation is being forced. The question is what the institution does next.

I want to be careful here, because what is happening in New York City is not uniquely bad. What is happening in New York is happening in every district in the country, and New York is the district where the conversation is being held publicly. That is to the credit of the parent coalitions that organized, the Council members who signed the letter, the Public Advocate who showed up at the rally, and the chancellor who is at least willing to delay the policy rather than ram it through. Most districts are not having this conversation. Most districts are deploying the same tools quietly, watching the same dynamics unfold inside their classrooms that New York is watching, and waiting to see what New York decides so they can use it as cover for whatever they do next.

The two-year moratorium that parents and Council members are calling for is a reasonable response to a situation that has gotten away from the institution, but it is a response, not a solution. A moratorium without a translation layer underneath it is just a delay before the same dynamic plays out again, because the underlying problem is not the technology. The underlying problem is that the institution adopted the technology without building the governance infrastructure that would have allowed families to participate in the decision as decision-makers rather than as commenters.

A two-year pause that is used to build the translation layer is governance.

A two-year pause that is used to negotiate the same vendor contracts with the same lack of family authority is theater.

The work the city has to do now is the work that should have been done before any of this started. A standing parent advisory body with reserved authority over technology decisions, established at the district level and replicated at the school level, would have surfaced these concerns at a fraction of the political cost and with significantly more institutional credibility, because the conversation would have happened with the people who had standing to be in it rather than with the people who had to take to the streets to claim standing they were never given. That body would have evaluated the apps that are already operating in classrooms, including the ones recording bathroom visits and sleep and mood and photos. That body would have established the conditions under which any tool could go live, the criteria for ongoing review, and the procedure for decommissioning a tool that fails. That body would have given the institution something it currently does not have, which is the ability to deploy AI in schools with the affirmative consent of the families whose children are most affected.

The question every parent and educator and Council member and chancellor watching this conversation should be asking, including the families whose children are already being recorded by classroom apps that nobody asked them to consent to, is this. Who has the authority to stop a deployment, and what would families need to say to exercise that authority? If the answer is that the only way to stop a deployment is to organize a rally at City Hall, the institution does not have governance. The institution has a public relations problem that becomes governance only when the affected community is loud enough.

The translation layer is the infrastructure that closes the distance between the people who are credentialed to make AI decisions about communities and the people who actually live with the consequences. In a school district, the people who live with the consequences are families. The translation layer says they belong in the room with binding authority, not at the rally with a sign.

Build that infrastructure, and the rally becomes unnecessary. The rally is the system telling you that you did not build it.