AI Guidance by Age

Explore how AI might help or hinder human development.

I’m a role…

I’d like to learn about

developmental domain…

for age range…

Developmental domain

Age range

Example output

Compass Collaborative · AI Guidance by Age

Quick Take

Our Recommendations

Key Developmental Processes and Milestones

How AI Might Hinder or Help Development at This Stage

*grounded in the longstanding science of human development, how might we think about the role of AI?

These hypotheses are based on longstanding research in human development, together with emerging research on AI. AI’s impact on human development hasn’t been studied systematically for long, so these hypotheses do not represent conclusive guidance. Instead, they’re meant to model the kind of careful thinking and deliberation required when considering AI use across different ages and in different areas of development.

Coming soon

The Human Loop

The usual advice is to keep a human in the loop. But look at what that phrasing assumes. If the human is in the loop, the loop belongs to something else, and the person is a checkpoint inside it. We think it should be the other way around. Keep the loop human, then work out where AI fits inside it in a way that supports human flourishing.

Methodology

How AI Guidance by Age is built

AI Guidance by Age, created by Compass Collaborative at the Harvard Graduate School of Education, makes hypotheses about where AI might help learning and development, where it likely hinders, and how the answer might change with age.

Why we built it

Settled evidence on how AI shapes a developing mind may take a decade or more. We have seen this before: a generation grew up with social media before longitudinal research arrived. Because choices made now will be hard to reverse, we start with what we already know. Decades of research on how people learn and develop let us form grounded hypotheses about AI that can inform decisions today.

How it works

AI Guidance by Age explores two different axes. 1/ a learner’s age and 2/ the developmental capacity they’re building, like early literacy or managing emotions. Where those two axes meet (e.g., 0-2 and numeracy), we craft a unique perspective because a five-year-old and a fifteen-year-old are developmentally different, and what’s happening for them in each developmental domain is different too. Every one of these perspectives explores four questions:

What is developing at this age, in this area?
What emerging research exists on AI here?
How might AI hinder this development?
How might AI help this development?

How we answer “help or hinder?”

With little established literature on AI and human development, we start with well-established research on how each capacity develops at each age. That research is the base of every entry. Early AI research informs our thinking. Much of it is new, small, or not yet repeated, so we weigh it against what is already known about development. From the two, we reason to the most plausible ways AI might help or might get in the way of development. This is abductive reasoning: inference from incomplete evidence that seeks the most likely explanation. Each entry is a hypothesis, not a firm finding, and a starting point for future research.

How each entry was created

Our ethos is to put AI in the human loop, not humans in the AI loop. At the same time, we see clear value in AI tools, and so we want to be clear about how each entry was created. In our “human loop” process, human researchers set the structure, write the guidance, and make every decision. AI does two jobs, and human researchers bookend them. Our process was as follows:

We defined the ages and domains the tool covers, and built a library of more than 5,000 academic research papers. This library consisted of two groups of papers. First, developmental science, learning science, psychology, cognitive science, and neuroscience research from HGSE faculty or PhD affiliates, plus any papers they include on their syllabi and on our doctoral exam reading list. Second, early research on AI in education.
For each entry, AI drew only from that library to summarize what the research says about that domain at that age and populate those papers into a document.
Using this input and their own expertise a research assistant wrote the entry, and a second research assistant checked it.
We worked with an abductive reasoning coach to push past status quo bias and design fixation: the pull to assume things will stay as they are, and to stop at our first ideas. This helped us think more widely about how these technologies might help or hinder development.
An editor cut each entry, often from five pages to two, and translated academic language into plain English. The author then confirmed the meaning hadn’t changed.
A PhD-level researcher or faculty member reviewed every entry.
AI checked every citation against its source and flagged anything that didn’t match.
A research assistant checked every cited paper and decided what to change.
A PhD-level researcher reviewed each entry again before publication. All sources are listed at the bottom of each entry.
We revise entries and add new ones as feedback comes in and new research is published.

We tested entries with early childhood, K–12, higher education, and workforce educators at several points, and their feedback guided each round of changes. Their feedback is the reason each entry opens with a summary and ends with recommendations.

What this version covers

This version focuses primarily on neurotypical learners under age 40. Most existing developmental research is based on populations in the Global North (sometimes called WEIRD populations), and thus has a bias towards these contexts. Future versions and our emerging research agenda will work to widen that scope.