Larry Berger (CEO of the EdTech firm Amplify) describes his work developing personalised learning systems during the 2000s and 2010s as rooted in an ‘engineering’ model of learning. 

Although Berger now sees this as highly flawed, he reckons it remains the dominant mindset among many EdTech developers and personalised learning innovators. 

As Berger puts it, this ‘engineering’ approach involves the following logic

  • Start with a map of all the things that kids need to learn. 
  • Then measure the kids so that you can place each kid on the map in just the spot where they know everything behind them, and in front of them is what they should learn next. 
  • Then assemble a vast library of learning objects and ask an algorithm to sort through it to find the optimal learning object for each kid at that particular moment. 
  • Then make each kid use the learning object. 
  • Then measure the kids again. 
  • If they have learned what you wanted them to learn, move them to the next place on the map. If they didn’t learn it, try something simpler. 
  • If the map, the assessments, and the library were used by millions of kids, then the algorithms will get smarter and smarter, and make better, more personalized choices about which things to put in front of which kids. 

Berger is now quick to distance himself from this procedural approach to learning. As he puts it, “I spent a decade believing in this model—the map, the measure, and the library, all powered by big data algorithms. Here’s the problem: the map doesn’t exist, the measurement is impossible, and we have, collectively, built only 5% of the library”.

Berger points to a number of fundamental shortcomings to this way of approaching the design of education technologies.

For example, he concludes that while some progress on developing a ‘map’ has been made in areas such as early reading and elementary school mathematics, most other forms of learning remain stubbornly un-mappable.

Similarly, there is no such sequential inventory of everything that someone needs to learn in terms of developing writing skills, reading comprehension, let alone more complex areas of mathematical reasoning. Indeed, as concepts, arguments and understandings get more complex, there is certainly no straightforward map for learning in the natural sciences or social studies. 

Berger also notes that even if an element of learning is ‘mappable’, there are no assessments that can accurately gauge an individual’s learning performance – let alone to assess with any precision what should then be learnt next.

Moreover, most learning objects cannot be understood in isolation, but are iterative. As Berger puts it, “most of the available learning objects are in books that only work if you have read the previous page. And they aren’t indexed in ways that algorithms understand”.

The final point of contention for Berger is perhaps the most insurmountable – the problem of human nature. As he concedes:

“Finally, as if it were not enough of a problem that this is a system whose parts don’t exist, there’s a more fundamental breakdown: Just because the algorithms want a kid to learn the next thing doesn’t mean that a real kid actually wants to learn that thing”.