Harvard's Graduate School of Education has released a tool that tells educators and parents what the research actually says about AI's effects on learning at a given age. The AI Guidance by Age Matrix draws on more than 5,000 pieces of research across 81 developmental areas, and it launched as part of a new initiative called Compass Collaborative on Nov. 10.
The timing matters. According to Education Week, research on how AI tools affect student learning is still emerging, and a separate audit of meta-analyses found the existing papers flawed and unreliable as guides for decision-makers. Schools are adopting AI anyway. The Harvard team's bet is that a structured synthesis beats an unstructured scramble.
For enterprise readers, the story is familiar. Any organization deploying AI into a workflow with weak evidence faces the same problem: decisions made in real time, tool by tool, without a shared framework. That is the gap Compass Collaborative says it wants to close, and it is the same gap showing up in workplace AI adoption, as tracked in our AI & Machine Learning coverage.
What actually shipped?
Compass Collaborative launched on Nov. 10 with one public artifact: the AI Guidance by Age Matrix. The website asks users to plug in their role, such as teacher, parent or researcher, a developmental domain such as cognitive foundations, creativity or relationships, and an age range from early childhood to adulthood.
In return, users get a quick take on what the research says about AI's possible effects on the skills that age range is developing, recommendations for how technology can support those skills, key developmental milestones, and where AI might help or hinder. The resource is part of a broader effort to close the gap between what developmental science says about learning and the decisions educators, policymakers and technology developers are making about AI.
Jeff Curley, chief of strategic initiatives in the dean's office at HGSE and co-director of the Compass Collaborative, described the problem bluntly. "There's a flood of attempts to understand how to navigate [AI] that's happening very differently from school to school," he said, "a lot of that being done without much evidence or research."
How thin is the evidence base?
The numbers in the Education Week report sketch a field still sorting itself out. Stanford University's AI Hub for Education reviewed the limited research on AI's effects on education and found mixed results that depend on how a tool is designed and used. A separate audit of meta-analyses went further, concluding the existing research papers are flawed and unreliable as guides for decision-makers.
Against that backdrop, the 5,000-study synthesis is less a definitive answer than a disciplined starting point. Curley said the team made a decision that "can be hard as an academic institution" to not wait and to meet the moment, releasing its best understanding while research continues. The guidance will be a living resource, updated as new studies on developmental science and AI's effects are published.
What does the guidance actually recommend?
The framework rests on two principles that co-director Tessa Forshaw, faculty chair of the workforce and workplace learning concentration at HGSE, called fundamental truths about how people learn. First, learning needs proximity to human relationships, meaning good relationships with more experienced others. Second, learners need opportunities to practice the skills they are developing.
Those principles produce a simple test. If an AI use case moves toward practice or human relationships, "generally you're in a really good place," Forshaw said. If it moves away from them, "that might be something you want to double-think about."
The age-specific entries make the tradeoffs concrete. For 9- to 11-year-olds, the guidance notes that children are typically still strengthening working memory, sustained attention and the ability to monitor their own thinking. Adaptive platforms might keep difficulty in the right range for meaningful practice, and evaluating AI output can build critical thinking that passive use bypasses. But AI homework and answer tools risk short-circuiting the retrieval that builds memory and leads to learning. This connects to our earlier piece, A Working List of AI Cybersecurity Tools by Function.
That same practice-versus-shortcut logic translates directly to workplace learning, where a deployment either builds employee capability or quietly replaces it. We looked at a parallel question in Machine Learning in Finance: What It Does, Where It Fails, Who Owns It. Readers following this should also see Machine Learning in Finance: What It Does, Where It Fails, Who Owns It.
Who is using it, and what do they think?
Early reactions in the Education Week report split along predictable lines. Heather Gauck, an elementary special education teacher in Grand Rapids, Michigan, called it a helpful tool. "There is still so much unknown with AI in education," she said in an email, adding that the platform answers educators' questions about how AI may affect students.
Tim Wells, an assistant principal at Mojave High School in North Las Vegas, was less convinced. He found the concept of breaking down audience, age and domain interesting but said the tool still feels basic. ChatGPT remains his preferred resource for figuring out how to use AI to develop student skills. "For now, I would not recommend it," he said, while noting the Harvard site may develop more in the near future.
That gap between a rigorous synthesis and a usable tool is worth watching. The resource is designed to be clear and actionable "at the point of decision for the educator or parent," Curley said. Whether it stays current enough to be that depends on the update cadence the team has committed to.
What should decision-makers take from this?
The operational takeaway generalizes beyond classrooms. When the evidence base is thin, anchor decisions to stable principles rather than tool-specific claims. Harvard's team used two: learning requires relationships with more experienced others, and it requires practice. Any AI deployment can be tested against principles like those before a procurement decision, a policy, or a pilot.
What remains unknown is how well the matrix holds up as the research matures. The underlying studies, by the audit's own finding, are flawed. Harvard has published its synthesis anyway, betting that a labeled, dated, revisable framework serves decision-makers better than silence. The next test is whether the updates arrive as fast as the evidence changes, and whether other sectors borrow the model. Our Tech News desk will follow the resource's development.
