“Adaptive” Learning Technologies: Pedagogy Should Drive Platform

GUEST COLUMN | by Tim Hudson

The world of ed-tech is moving rapidly. As new learning software is created, the word “adaptive” is increasingly being used in claims describing how technologies personalize and individualize learning for each student. As the Curriculum Director for DreamBox Learning, I oversee development of math lessons that are built on an adaptive learning platform. Developers like DreamBox have built these “recommendation engines” with the worthy goals of (1) ensuring the success of every student, (2) enabling each student to learn at her own pace and level of achievable challenge and (3) supporting teachers whose daily challenge is differentiating for an entire classroom of students at different points on their learning paths.

We know too much about human learning to embrace adaptive platforms that ignore pedagogy.

A primary way in which any adaptive platform supports teachers is by making recommendations for students in real time. Teachers are stretched thin and work tirelessly for their students.  It’s asking too much of teachers to require they analyze individual student data and make assignments to every student on a daily or weekly basis. Not only is this continuous data analysis cycle challenging and time-consuming, but it also requires deep curriculum expertise in order to preserve what the authors of the Common Core refer to as “consistent progressions” and “coherent connections” as students learn and develop.

Despite these noble goals and strategic support, technologies claiming to be adaptive have rightly come under scrutiny from educators who understand curriculum and pedagogy and are justifiably skeptical. In August, Audrey Watters described “any company touting ‘adaptive learning’ software” as being influenced heavily – if not entirely – by the behaviorist B.F. Skinner. The Skinner excerpt she references in her post accurately describes the approach used by many adaptive platform developers. I believe it also describes how many educators expect “adaptive” software would be designed. Skinner’s ideas do not match the research about how humans develop cognitively. So Watters is right to be wary.

In October, Dan Meyer posted two criticisms of “adaptive” technologies. In the first, he drew comparisons to Stanley Erlwanger’s research on the failures of Individually Prescribed Instruction (IPI).  Meyer appropriately lamented the poor feedback given to students along with the fact that antiquated prescriptive approaches always fail to develop mathematical intuition and appreciation for the beauty of the subject. In his second post, Meyer referenced quotes about two developers that built adaptive engines around analysis of behavioral data while disregarding the quality of students’ learning experiences. Meyer noted, “I don’t have a lot of hope for a system that sees learning largely as a function of time or time of day, rather than as a function of good instruction and rich tasks.”

Watters and Meyer are just two of many legitimate skeptics of adaptive learning software. Their concerns are valid and must be considered when examining whether new tools are truly helping students learn. Even though “adaptive learning” developers have noble goals, the design of each adaptive platform reveals important pedagogical approaches and assumptions made by the developers. The adaptive platform determines the pedagogy and the way students engage with learning. Not all adaptive platforms are capable of supporting strong pedagogy and rich learning tasks.

Broadly speaking, there are two ways to build an adaptive learning platform. The first is an approach similar to how Netflix makes entertainment recommendations. It begins with existing, static content such as textbook passages, online lectures, and quiz or standardized assessment items. Usually, this content is very narrow in scope in order to isolate specific skills and diagnose errors in using those skills. Next, this content is arranged into a sequence (or “learning map”) based on a best guess of how students should encounter the skills. When students begin lessons, their progress data, behaviors (such as click-rates or login times) and assessment scores are subjected to “learning analytics” to establish a learner profile and position on the lesson map. That profile is used to recommend lectures, choose lesson sequences and report usage insights such as suggested study and practice times. Additional analytics are then applied so that crowd-sourced user behaviors inform and adjust the learning map and sequencing for future students.

While this “behavioral profile” platform design is effective for making entertainment recommendations, it has several weaknesses and limitations when applied directly to learning. First, it replicates many of the mistakes of IPI, most notably the flawed assumption that “learning comes about by the accretion of little bits” (Shepard, 1989). Second, such a platform is completely dependent on a pedagogical model where the teacher (or system) “delivers” content and students become “receivers” of information. The lessons and instruction are static, and students therefore never engage in authentic, independent thinking.  Such a platform may collect mountains of data, but they are not data about students’ understanding and cognitive development; they are data about behaviors and the ability to replicate procedures on shallow assessment items. Third, the “adaptivity” for students not making progress is essentially recommending that they passively receive the same or similar static content again. It seems that for online lectures, the strategy of “pause and rewind” has become the 21st century equivalent of a teacher speaking “slower and louder.”

DreamBox took a different approach and built an intelligent adaptive learning engine.  The intelligent difference lies in the pedagogy. Wiggins and McTighe say it best in Schooling by Design: “An understanding is a learner realization about the power of an idea… Understandings cannot be given; they have to be engineered so that learners see for themselves the power of an idea for making sense of things” (p. 113). At DreamBox, our starting assumption is that students are brilliant. Their critical thinking skills are underestimated if we think understanding can be given through content delivery.

Therefore, when creating lessons on the DreamBox platform, we design digital tools that empower students to have their own realizations. Our lessons adapt in real-time because they aren’t static content; they were built to be interactive and adapt at any moment to any child. Our non-linear sequencing is informed by decades of research about children’s natural development and growth in mathematical reasoning. Our lesson progressions are not crowd-sourced using other students’ preferences and behaviors. Our lessons are written by experienced teachers who decipher students’ thought processes based on their interactions with our digital tools. DreamBox teachers write lessons that respond to different strategies in specific ways, ideally just as a teacher would in person.

Students using DreamBox choose and create their own strategies, and they get feedback specific to their ideas. A simple example is the act of counting by groups. Young children naturally count by ones, but counting by groups is not an immediately intuitive idea. The pedagogical approach of other adaptive platforms would involve a teacher explaining, “You can also count by 2’s, 5’s, or 10’s. Here’s how.” In this content delivery model, the student is receiving someone else’s idea and strategy. As one of my math professors would say, “she’s getting answers to questions she’s never asked.” The platform and pedagogy of DreamBox empower students to develop ideas and skills in a better way.

Adaptive learning software is a powerful partner with teachers and schools to ensure student success. We know too much about human learning to embrace adaptive platforms that ignore pedagogy. Sound pedagogy drives DreamBox’s intelligent, adaptive platform.

Tim Hudson, Ph.D., is Director of Curriculum Design at DreamBox Learning. Prior to joining the company in 2011, he served as the Curriculum Coordinator of K-12 Mathematics for the Parkway School District in St. Louis, MO, facilitating the development of the K-12 math curriculum; the development of district-wide math benchmark assessments; the creation of a math intervention program; and the selection of math-based technology tools and textbook materials. He is a longtime member of the National Council of Teachers in Mathematics (NCTM). Write to: timh@dreambox.com


Wiggins, G. & McTighe, J. (2007). Schooling by design: Mission, action, and achievement. Alexandria, VA: Association for Supervision and CurriculumDevelopment.

Shepard, L. A. (1989, April). Why we need better assessments. Educational Leadership46(7), 4–9.

  • Sanjay M


    Hello Tim, Very thought provoking article. But can you give some more examples around student choosing their own strategy (counting in 2s or 5s is a very elementary example, can you take a high school math example?).

    Also “non linear sequencing” that you talk about – is it the same as skipping of explanations if all are clear and if students are confused then go further in depth of the concept? Something that happens in any adaptive learning.

    • Tim Hudson

      DreamBox is currently a K-5 product. So while I don’t have a high school example, I do have upper elementary examples that can help answer your question. One good example is a lesson about division with remainders. You can play this lesson like you were a student (http://bit.ly/NanvgR) or access the teacher version of the tool (http://bit.ly/WubPtN).

      In this lesson, students are given the task of packing gumballs into bags of varying size. Students are free to use facts and relationships they know to start packing in groups of their own choosing; there is no explanation up front to give students a strategy or procedure to follow. There are a limited number of moves, though, so students need to exercise some efficient and strategic thinking. The interactive tool helps students reflect on their initial strategies and gradually realize better strategies. You could also look at our sample lessons with the open array (distributive property) and fraction multiplication to get a sense for how students are empowered to think independently and have realizations on their own.

      Regarding the non-linear sequencing… Our adaptive platform does allow students to skip content when they have demonstrated proficiency. But in this instance, I was referring to the fact that DreamBox allows students to engage with concepts and ideas that are not ordered in a rigid linear sequence. For example, a particular second grader could be presented with these four lesson choices: place value in the hundreds, representing addition on a number line, fluency with subtraction, and early skip-counting for multiplication. None of these lessons is inherently a pre-requisite for any of the others. Students benefit because they have an element of choice and don’t hit “dead ends” while playing. They are also better able to connect concepts and thereby deepen their number sense.

  • Rich


    Thoughtful – great job

  • geonz (@geonz)


    I have been frustrated for decades at the rote procedural stuff that claims to be “individualized.” I couldn’t believe people couldn’t even reasonably illustrate basic arithmetic operations. Students got token exposure to “manipulatives” and “concrete experiences,” and then… crunch, numbers, crunch!!!
    I love a lot of Dreambox’s approaches, especially the cool fraction sections where they are subtracting where you’d need “borrowing.” With our older students, they have no floggin’ idea what is going on, and so they painstakingly but temporarily memorize how to make everything an improper fraction, etc…. while in Dreambox you click on a picture of “one” and it turns into 6/6 or 5/5 and you can visually subtract (“borrow,” schmorrow, it’s not as if you’re giving it back…).
    The next stage of development I’d like to see is strengthening the bridge between that visual-concrete stuff and the language, both verbal and mathematical, used to describe it. The nastiest thing that could happen would be to let the sales-rep attitude take over, and stop looking for the students who are still struggling, and looking for possible gaps and flaws.
    (I also work with older students who struggle — on our diagnostic, 1 of the 35 could answer what 4 1/2 x 2 was… though *most* of them showed work… and over half of them thought 1/8 was the same as or greater than 0.8 . I wish Dreambox were packaged in such a way that we could use it… but it’s inspired me to develop things on my own…)

    • Tim Hudson (@DocHudsonMath)

      Thanks for the kind words about our approach and manipulatives. Rote procedural digital learning is neither pedagogically sound, truly individualized, nor effective for sense-making and transfer. We focus significant energy into making sure students connect concrete interactions with the more abstract representations of mathematics. Using language and visuals together is critical to making that happen. When older students can’t multiply 4-1/2 by 2 as you described, we know they haven’t had enough experiences thinking about numbers, playing with numbers, and truly becoming fluent. It’s great to hear that you’re developing some of your own things. If you have a chance, I’d be interested to hear how DreamBox could be packaged in a way you could use it.

  • Phil McRae


    Rebirth of the Teaching Machine through the Seduction of Data Analytics: This Time It’s Personal http://bit.ly/1231loJ Adaptive Learning Systems, DreamBox Learning Inc., B.F. Skinner – A critique of the movement

    • @DocHudsonMath

      Phil – Thanks again for your thoughtful piece and sharing your concerns about adaptive learning technologies. I wrote this particular article in November of 2012 to clarify the misconceptions about Skinnerism and our technology. I agree with you that Skinner’s behaviorist approaches to learning and educational technology aren’t effective, and many edtech products are built upon his theories. But DreamBox isn’t one of them. I’ll share some brief responses to points you made in your article that I’d be willing to also post to your site in order to contribute to the dialogue there.

      Earlier this year a thoughtful secondary math teacher named Brandon blogged to raise many of the same points you raise. For a thorough explanation of how DreamBox content, pedagogy, and technology are different, read my exchange with Brandon in the comments section on his blog here: http://bit.ly/YEj5qs. I had the pleasure of meeting Brandon at a math conference earlier this year, and he generously gave me some of his time to show him more about how DreamBox is not the rebirth of Skinner’s teaching machine. Our lessons are neither linear nor behaviorist, and we don’t frontload explicit instruction (a huge flaw of CAI and similar approaches).

      We don’t sell any student data. It belongs to the schools, and we protect it. We also promise that students aren’t passive consumers when they use DreamBox. They’re engaged in thinking mathematically and independent problem solving.

      We don’t expect students to use DreamBox for hours on end. Two or three 30-minute sessions a week is sufficient – like piano lessons. Our DreamBox whitepapers describe how our technology is designed according to principles of learning – we don’t suggest how to design a school or make staffing decisions. We partner with schools all across the US and Canada, and they make local decisions about when and how students use DreamBox.

      I couldn’t agree with you more that it’s essential for students to be in classrooms and rich social learning communities where they have great conversations about mathematics problems with their peers and teachers. To my knowledge, no school has chosen to eliminate math classes because students are using DreamBox. And I wouldn’t advocate that approach.

      Lastly, I agree with you that learning and student circumstances are complex, and technology can’t solve every problem. What technology can do is support teachers with actionable data for their analysis and planning. The relationships students develop with caring professional educators in great schools will always be the cornerstone of learning and growth. We’re not attempting to displace the human dimension of learning. We’re complementing it.

      Thank you for acknowledging in your piece that emerging technologies have a place in educational transformation. We have yet to reach all learners, and there is always room to improve schools, classrooms, and technology.

  • vlagood


    Not all Adaptive Learning Technologies are created equal. That is true. Most of them are based on certain pedagogical models. Pedagogical models are not equal too. Why have you selected the weakest models (Skinner’s one, Individually Prescribed Instruction (IPI), or even Netflix approach) for comparison with your DreamBox? There are much stronger models for comparison.

Leave a Comment

%d bloggers like this: