Tech and Edtech

When I was getting started in education reform, there was Linda Darling-Hammond saying "STOP".

When I was getting started in education technology, there was Larry Cuban saying, "it's not going to be this easy."

In both cases, I was incredulous. We're trying to do something here. Something really important. For kids. How can you be so sure it's not going to work? Are you rooting for it not to work?

Well, 20 years later and NAEP scores haven't budged (as they hadn't for the 20 years before that). To be fair, the odds were with the naysayers. Solving education problems at scale is extremely challenging. The articles I pay the most attention to these days are Dan Meyer poking holes in the AI for education narrative, and Tim Daly reflecting on lessons learned from the ed reform years. Maybe I've just grown old...

Yesterday I was scrolling through my feed and saw Ben Kornell urging us:

It's time for us to eradicate #dyslexia. Period. Full stop. With #AI, we now have the tools to diagnose and treat #dyslexia at 3-4 years old, while learners have maximum neuroplasticity and before they struggle with literacy in K-2nd grade. How do we make this available to all?

I completely agree that it's time for us to eradicate dyslexia (hashtag or not) and although I'm not a scientist, the research I've read seems conclusive that treating dyslexia early is much more effective and efficient than treating it later. I've read that students are still quite responsive to interventions in K-2nd grade, but after that it becomes much harder. Intuitively treating it even earlier makes sense too.

I was surprised, though to hear that we now have the tools to diagnose and treat dyslexia at 3-4 years old (or any age for that matter) and the only challenge is how to make them available at scale. By tools, I inferred a technology that can crank out the same results every time, at scale as opposed to human-mediated know how.

Ben linked to a SmartBrief article. I was curious to read more.

The article started by validating the size and urgency of the problem:

Dyslexia is not just an educational problem but an economic one. Affecting 1 in 5 people, dyslexia is the largest category in special education. Other special education categories, such as attention-deficit/hyperactivity disorder and autism, often come with reading and learning difficulties as well. Schools across the US spend over $120 billion a year on special ed.

That sounds right. Then the article notes that

in the last decade, rigorously designed dyslexia studies show no significant positive effect of intervention on broad reading achievement for at-risk readers after third grade.

Sounds reasonable, if depressing. Then states:

Due to economic and staffing limitations, interventions currently only serve about half of students with a reading difficulty. Can schools do more for this group? The answer is “no” if they keep to the current system, but it’s “yes” if they revamp the system.

Wow. Are half of students with a reading difficulty being well served in schools today? The author doesn't say "well-served", but nonetheless that's really surprising. The stats I recite are that only a third of children are proficient in reading by third grade. My intuition is that a small fraction of children who need interventions to remedy dyslexia in schools today are getting ones that work.

The author gives an example:

Next, let us take another fictional example with Chris, a fifth-grader who has dyslexia, ADHD and autism. These conditions often co-exist. Chris’s school intends to use a more elaborate evaluation to classify her disabilities to provide appropriate services. Now, we see even more friction points. The school has to decide on which test batteries to administer among an alphabet soup of literacy, language and cognitive assessments, and by whom. Since most schools have only a few certified personnel who can administer such assessments, schools have to decide on which students to evaluate and which ones to put on a waiting list. Again, Chris’s team has to interpret the evaluation results one student at a time. But the task is considerably harder because of 20 pages of results to translate into actionable plans. Now multiply these efforts for however many students involved.? A special education teacher I met recently said she quit her job because of this Herculean task.

before going on to cite the potential of AI to better understand the complexity of the human brain, identify language processing issues too ephemeral for humans to notice, and have the capacity to serve all students.

This example, and the near term potential of AI no longer squares with my intuitions. My impression is that in the hands of expert practitioners, we have methodologies to diagnose and treat dyslexia. Frankly, the diagnosis seems like it would be the easier part. The treatment is long and painstaking, and the challenge of maintaining student motivation during the treatment is significant. I don't know of any school systems that are using humans to diagnose and treat dyslexia effectively at any sort of system-wide scale. I don't think the core of the problem is deciding which test to administer, or even how to interpret the results, although tech probably could speed that up. It's how do we fix the problem.

Compare Waymo . Waymo demonstrates one of the most sophisticated and challenging uses of AI in our time, and it's trying to solve the really big problem of 40,000 automobile fatalities a year in the US. That said, there are tens of millions of extremely safe drivers in the US. We do not need AI to teach us how to safely operate a car. In fact, safely operating a car is so easy for so many people, that some people attempt it intoxicated, or while distracted, or while texting, or in a rush, or with a medical impairment, or without sufficient training, or with a blatant disregard for the rules of the road, often with disastrous results.

Waymo set out to build self-driving cars because humans had been successfully driving them for a century. Automotive plants assemble cars with robots because, again, humans have been successfully doing so for a century. That's how industry nearly always applies decision-support technology--to automate that which we already know how to do.

I agree that it's time to eradicate dyslexia, and I would be absolutely thrilled if it proves true that we have the AI-powered technology already in hand to do just that. I just don't think we do, because I don't think we have examples of humans already doing this at any meaningful scale. I think the challenges of motivating a child to overcome a deficit and learn are immense for human teachers and will be even more so for AI ones, and this is going to be a marathon, not a sprint.