Artificial intelligence is moving into government faster than many states anticipated. After an initial period of caution, when agencies were primarily focused on understanding the risks and establishing guardrails, we’re now seeing greater interest in how AI might improve how government works.
That shift is particularly important in education, especially early childhood. State systems manage enormous amounts of information while supporting families, providers, and children through processes that are often complicated, understaffed, and supported by aging infrastructure. This presents a legitimate opportunity for AI to reduce some of that administrative burden, but it’s a bit more complicated than simply adopting new technology.
States need to be clear about where AI can add value, where it should remain limited, and what safeguards should continue to guide its use. It’s not about embracing AI indiscriminately or banning it altogether—both those ships have sailed. What we need to do now is ensure AI is used in ways that make systems work better while still protecting the people those systems serve.
Moving from caution to responsible adoption
That shift is showing up in recent requests for proposals (RFPs) at the local and state levels. Vendors are now being asked not only to explain their AI guidelines and models, but also how they plan to implement it to improve services and make government more effective.
That represents a meaningful change from even a year or two ago, when much of the focus was on disclosure and control. Early pilots and test cases have continued to prove both feasibility and value, leading to an environment in which the conversation has moved from “what if?” to “what else?”
In early childhood, the most promising opportunities are practical, behind-the-scenes uses that can help people do their jobs more efficiently. That could mean reducing some of the manual work involved in subsidy applications, complex licensing workflows, case management, and routine communications, giving staff more time for the work that requires experience, judgment, and direct interaction with families.
We need to define where AI belongs and where it doesn’t
That means being particularly cautious about using AI as an instructional tool for very young children. We need to be aware that there is no strong evidence that suggests AI-driven instruction improves outcomes for children from birth to age five. Early childhood brain development still depends heavily on relationships, interaction, play, and hands-on experiences.
The same caution applies to decisions that can have significant consequences for children, families, or providers. AI might help organize information, flag something for review, or surface patterns in a large volume of records. But decisions about whether a family receives benefits, whether a provider receives a license, or other high-stakes outcomes call for human judgment and accountability.
Human oversight has to be more than a checkbox
Nearly every responsible AI framework includes some version of “human in the loop” guardrails. But meaningful oversight requires more than just having a person review an AI-generated recommendation. The person reviewing that output needs enough information to understand how the system arrived at it, enough context to recognize when something doesn’t look right, and enough authority to reject or disregard it when necessary.
There are already well-known examples of what can happen when organizations fail to account for AI’s limitations. Amazon, for example, abandoned an experimental recruiting system after discovering that its training data had resulted in the system disadvantaging women. Because the model learned from historical hiring patterns in a male-dominated workforce, it reproduced some of those patterns in its recommendations.
The lesson for the government is not that AI cannot be used in important or sensitive processes. It is that AI does not automatically make a process objective or unbiased. Systems can and do reflect the assumptions and inequities embedded in the data and processes from which they learn, making monitoring and human accountability essential.
Protecting children’s data requires a different level of care
Privacy becomes even more important when AI is introduced into systems that handle information about children and families.
One useful way to think about an AI system is as another potential user of information: if a human employee needs an appropriate reason and authorization to access sensitive information, an AI system supporting that employee should not automatically receive broader access simply because it can. That starts with data minimization. An AI system drafting a general family newsletter does not need access to a child’s complete record. An AI system helping a caseworker review an application may need access to certain information, but that access can be limited to what is necessary for the task.
States and vendors also need to understand where information is processed, how it is retained, and whether it can be used to train another model. Sensitive child and family information cannot simply be entered into public AI systems in the same way you or I would ask ChatGPT for a pancake recipe.
Private or isolated AI environments provide an additional layer of protection when sensitive information must be processed. Role-based access controls, audit logs, and clearly defined retention policies can be built into the architecture from the beginning rather than added later.
Transparency with families and providers is another consideration. When AI is used to process sensitive information, people need to understand that it is being used and, where appropriate, have meaningful choices about that use. As these technologies become more integrated into public services, expectations around consent and the ability to opt out are likely to become even more important.
Putting AI to work behind the scenes
Realizing the vast potential of AI requires a thoughtful approach. Treating the technology as inherently dangerous could prevent useful applications. Assuming it will automatically make government more efficient could create its own problems. Neither approach gets at the practical decisions agencies now have to make about where the technology fits and where it doesn’t.
States are entering a more productive phase of the AI conversation, one focused less on whether the technology should exist and more on how it can be used well. That means identifying problems AI is actually equipped to help solve, establishing clear boundaries around its use, and building transparency and oversight into the systems from the beginning.
The technology will continue to change, and state policies will need to change with it. What I hope remains consistent is the principle that AI should strengthen the people and processes responsible for serving children and families, rather than becoming a substitute for the judgment, relationships, and accountability those systems require.
