Mastery Learning: Why Bloom's Idea Only Works Now

Anyone can build a chatbot that answers questions. The real innovation at TreadEd is a didactic idea that's almost 60 years old – and why AI finally makes it affordable for everyone.

What else is new, from diagnostic steps to voice mode, is described by Christoph in From Training to the Real Thing. This article is about the didactics behind it. And that's almost 60 years old. It's called mastery learning, and it comes from Benjamin Bloom.

Bloom's starting point was a simple observation. If everyone gets the same instruction in the same amount of time, results spread out like a bell curve. A few do very well, most are average, a few fail. Bloom didn't see this as a law of nature, but as a consequence of instruction. Given the time and support each person needs, the large majority can genuinely reach a goal (Bloom, 1968).

His method is simple. Material is broken into units with clear learning goals. After each unit comes a short test that isn't graded, but informs. Bloom called it formative. Anyone who hasn't yet reached the goal gets targeted corrective instruction, explaining the material differently than the first time instead of simply repeating it. Then a second, equivalent test follows. Anyone who has already reached the goal moves on to enrichment (Guskey, 2007).

Assessment was only ever "mastery" or "not mastery", and the latter always counted as provisional (Bloom, 1968). There was no failing. Only not yet.

Group-based mastery learning was therefore a compromise. A single teacher can't accompany every person individually. So material gets broken into small units, tested after each one, and reworked where needed. A good tutor, by contrast, notices in conversation exactly where something isn't clicking and reacts to it. One-on-one tutoring for everyone was just never affordable.

A course, on the other hand, is produced once and distributed to everyone, in the same order, with the same final exam. What someone can't do often only surfaces at the very end. The gap is identified, but not closed.

Digital learning systems have since automated Bloom's compromise. They break courses into small learning goals, check them with short quizzes, and serve up matching content.

But even here, the frame stays a pre-written course. Content has to exist in advance for every possible gap. And what gets measured is mostly whatever can be evaluated automatically – multiple choice rather than explanation.

Assessment can happen in conversation. Until now, only what could be marked right or wrong could be evaluated automatically. Whether someone could explain, justify, or apply something had to be judged by a human. Today, diagnostics can run like an oral exam, as often as needed. Instead of once at the end of a unit, measurement can happen almost continuously.

Corrective instruction no longer has to be pre-produced. Bloom required that corrections approach the material differently than the first pass. AI can do exactly that for every single person, with a new explanation, an exercise, or a role-play, and with examples from their own everyday life. People still plan how learning happens. AI adapts it to each person.

Early controlled studies support this. A didactically designed AI tutor gave students more than twice the learning gains of active in-person instruction, in less time (Kestin et al., 2025). A chatbot without didactic guardrails, on the other hand, can harm learning (Bastani et al., 2025). What matters is the didactics, not the language model.

Diagnostic steps make this even more precise. They don't teach, they check, in writing or by voice, against clear criteria for what has to be solid. A good impression becomes a precise finding. That creates a loop. It starts with a goal. Diagnostics show where someone stands. Only what shows up as a gap gets reworked, then it's measured again. Once the gap is closed, the next sub-goal follows. The AI companion holds this loop together.

A course runs the same way for everyone. The loop only takes an extra round where measurement shows a gap.

It knows the gaps. The companion knows from conversations what someone can already explain confidently and where it's shaky. Not as a score, but concretely – say, that someone mixes up two terms or is always unsure about a particular question.

It addresses them specifically. Instead of repeating the whole topic, it zeroes in on exactly that spot, with a different explanation than the first time, the way Bloom required for corrective instruction, and with examples from the person's own everyday life.

It proactively leads there. Learners don't have to know themselves what they're missing, or ask the right question. The companion suggests on its own where to go next, and then checks whether it has landed.

Here's what that can look like. Lea works in procurement. Her company is rolling out AI assistants, and Lea is supposed to be able to use them confidently day to day. In conversation, she explains well how to give an assistant clear instructions. Then she mentions, in passing, that the assistant surely knows the current supplier prices. That's the gap. A language model doesn't know anything about that at first. It only reaches a pricing database or ERP system through tools someone has connected, so-called tool calls. The companion picks up on this on its own: "Earlier it sounded like you thought the assistant just knows your prices. Want to take a quick look at how it actually gets that data?" Lea then sketches out, for her own supplier comparison, which tools the assistant would need and where she wants to grant access herself. A check-in question follows. If the answer holds up, they move on.

Passing an exam. Whether it's a final exam in vocational training or a certification, the goal is clearly defined. The exam regulations set out what has to be mastered, and often how it's tested. So assessment happens in the exam format itself, for example as a simulated oral exam. The loop ends once the simulation holds up.

A competency for a business goal. Many learning goals don't need an exam. They follow from what a company wants to achieve. If a company rolls out AI assistants in procurement, as with Lea, teams need to be able to use them safely. That means it has to be clear what they should be able to do afterward – for example, judge where an assistant needs access or approvals, and check its results. And what's out of scope, like programming.

This clarity isn't a minor detail. Only someone who knows where the goal is can measure how far someone still is from it. For L&D, that means competency is no longer counted in hours completed, but in demonstrated ability.

Lower cost. Learning time is work time. Every hour that doesn't go into something already known stays on the job. And training no longer has to be bought as a package – it targets exactly where something is actually missing.

Gaps closed with precision. Measurement doesn't just show that something is missing, but what. Follow-up work happens exactly there, and the next measurement shows whether it was enough. That's the difference between "completed the course" and "can do it now".

Personal context. In a course, everyone gets the same examples. In the loop, each person brings their own context – their everyday work, their cases, their questions. Explanation and practice happen on what's actually occupying them. That makes what's learned easier to transfer.

Courses won't disappear overnight. But they're losing their role as the unit around which learning is organized. The goal is the unit. Content, exercises, and exams become building blocks that get used exactly when they're needed.

For education providers, that means their value lies less in the finished course than in clearly defined goals and good measurement. For L&D, it means training can finally be measured by whether people can do something, not by whether they showed up.

About the author

Daniel Kaulbach is Co-Founder of TreadEd.

TreadEd is a creator tool that makes it possible to design dialogue-based learning paths in which learners are always optimally supported through one-on-one guidance by an AI tutor. Experienced employees fully shape the tutor didactically, passing on their knowledge and experience at scale.