In 1984, educational psychologist Benjamin Bloom published a paper that should have upended how we think about teaching. It was called "The 2 Sigma Problem," and the finding was equal parts breakthrough and dead end: students who received one-on-one tutoring performed two full standard deviations better than those in conventional classroom instruction. Two sigma. That's the difference between average and exceptional: the 50th percentile and the 98th.

The good news: we knew exactly what worked. The bad news: we couldn't do anything about it. One-on-one tutoring doesn't scale. It never has. You can't assign a personal tutor to every learner in a certificate program, a university course, or a corporate training initiative. The economics don't work. The logistics don't work. So for forty years, educators and instructional designers, myself included, accepted a quiet compromise: design for the middle, hope the outliers find their way.

For the next forty years, the field tried to crack it. Mastery learning, Bloom's own proposed solution, moved the needle by about one sigma. Adaptive learning platforms promised more. Most delivered thin personalization: adjust the pacing here, serve a remedial module there. Not the same as genuinely adapting the learning experience to who the learner actually is.

Before I get to what changed, it's worth understanding why closing the gap is harder than it looks, and why most organizations trying to leverage AI for learning will fall short.

Most instructional design operates almost entirely within Bloom's cognitive taxonomy: a six-level framework moving from basic recall through understanding, application, analysis, evaluation, and finally creation. It's a powerful tool. I used it for years to build assessments that moved learners up the ladder.

But cognitive learning is only half the picture.

In 1964, David Krathwohl, Bloom's own collaborator, published a companion taxonomy for the affective domain: the dimension of learning that involves values, attitudes, professional identity, and character. Its five levels move from simply receiving new ideas, through responding and valuing, to organizing those values into a coherent belief system, and finally to characterizing: acting consistently with internalized values as a matter of professional identity.

Affective learning was part of our thinking throughout the program's development. But midway through building the capstone — the Chartered Marketer Summit Course — I wanted to tackle soft skills directly. Marketing demands them more than almost any discipline, and I wasn't willing to gesture at them. Something still wasn't sitting right as we worked through it. We were conflating two fundamentally different kinds of learning. How to structure a pitch is cognitive. Becoming someone who believes they can pitch, who has internalized the values of professional communication, that's affective.

I stopped development completely and split the team: one group working on the cognitive objectives, one working on the affective. It's a distinction I've explained to students and practitioners the same way ever since: you only learn in the affective domain by experience. If I tried to teach you how to present, I'd just be teaching you my skills. Go take an improv class. Create conditions for the experience. Don't lecture your way to it.

This distinction matters enormously for AI adaptive learning. Most adaptive systems today, even sophisticated ones, operate exclusively in the cognitive domain. They adjust difficulty, sequence content, and personalize pacing. What they cannot do is design for the affective dimension: the values, self-awareness, and professional identity that determine whether a learner becomes genuinely capable, not just technically competent.

I started using ChatGPT in early December 2023. By early 2024, you could upload PDFs.

Years of building original courses meant I had a hard drive full of them: complete, unabridged, word-for-word course content. Not summaries. Not outlines. The real thing. A five-semester professional designation program built over four and a half years with a team of twenty-eight expert marketers, instructors, and subject matter experts, every one of them personally recruited. All original content, both cognitive and affective objectives, was designed from the start.

I wrote five learner personas, not archetypes invented at a desk, but people I'd watched move through seven cohorts. The one who read every module the day it dropped. The one who disappeared until the deadline. The one who aced the quizzes but couldn't see the forest for the trees. The one with a blind spot they didn't know they had.

I uploaded the curriculum. I uploaded the personas. Then I asked ChatGPT to write the assessment instructions for one major assignment. Twice. Once for learner three. Once for learner five.

What came back was pedagogically sound and completely different for each one.

ChatGPT did not know it was solving a “2 Sigma” problem. It did not know why the curriculum was structured the way it was, or why one persona’s entitlement belonged in the emotional layer rather than being treated as a simple behavior. It just worked with the inputs I gave it.

Those inputs were the result of years of building, watching, pausing when something felt off, and iterating. The curriculum was not a course outline. It was thousands of hours of original material built by twenty eight people under real delivery pressure. The personas were not demographic profiles. They were composites drawn from real learners I had worked with, cohort after cohort, presentation after presentation.

That is the part most organizations will struggle to replicate. Not because the AI is hard to access. It is becoming cheaper and more widely deployed every month. The bottleneck is the depth of the inputs. You need curriculum built with craft, learner models grounded in observation, and someone who understands that learning is not only cognitive.

Most organizations still treat learning like content distribution. They upload a slide deck, generate a quiz, and call it personalized because the pacing changes. That is not adaptive learning.

The challenge AI-driven adaptive learning creates for traditional institutions is not mainly technical. It is structural. For decades, scale has been the justification: we cannot personalize because it is too expensive. AI weakens that argument. If an instructor can produce meaningfully different explanations, practice sets, and feedback for each learner, then “we can’t personalize” becomes less about resources and more about priorities.

The harder problem sits above individual faculty. It is the institution’s instructional architecture. Traditional teaching is built around one expert, one voice, one sequence. Adaptive learning at scale asks for a different capability: combining expertise from multiple people into a single system that can respond differently to different learners. How do you take the knowledge of five subject matter experts and translate it into a curriculum that adapts across thirty learner profiles at once? That is not just a tooling decision. It is a curriculum design and governance problem that many institutions have not had to solve.

The educators who will do well here will not be the ones with the cleverest prompts. They will be the ones with a precise understanding of how people learn, including the cognitive and emotional parts, and the craft to build learning experiences that hold up. The teacher’s role does not disappear. It shifts from delivering information to designing the conditions for practice, feedback, and progress.

The 2 Sigma gap hasn't closed because AI arrived. It's closed now because the question finally shifted.

For forty years, the question was: how do we scale one-on-one tutoring? The answer was always: we can't. Too expensive. Too labor-intensive. A Department of Education-funded randomized trial of more than 18,000 students across 147 schools found that AI-assisted adaptive instruction nearly doubled student growth on standardized tests by the second year of implementation. The market signal is equally clear: the global AI in education market is projected to grow from $7 billion in 2025 to $136 billion by 2035.

If the answer is "not very" and "not deeply," AI will give you exactly what you already have: generic content delivered faster. But if you've done the work, if you've built something with real intellectual weight across both taxonomies, and you understand who's receiving it, then yes, the gap closes. Not perfectly. Not yet. But meaningfully, and at a scale that Bloom couldn't have imagined.

We may be living through the revenge of the liberal arts majors. The people who will get the most from AI in education aren't the ones with the best prompt engineering. They're the ones with the largest vocabulary for describing how humans actually learn, and the stubbornness to build something worth teaching.

The tool was never the variable. The question was always: what did you bring to it?

Roam is the virtual workspace our team relies on to stay connected across time zones. It makes collaboration feel natural with shared spaces, private rooms, and built-in AI tools.

Roam’s focus on human-centered collaboration is why they’re our Premier Partner, supporting our mission to connect the builders and leaders shaping the future of AI.

The AI Collective is a community of volunteers, made for volunteers. All proceeds directly fund future initiatives that benefit this community.

Thank you to the thousands of volunteers around the world who make this work possible. We truly could not do this without you.