Watch the RecordingListen to the Podcast

AI in education—like many industries—is an incredibly opaque space right now. Currently, the term is used as a catchall for any type of perceived machine learning or assistance. Educators need to be specific about how they talk about AI, though, since there are many different forms.

They also need to be cautious about how they use it, said Dr. Tyler Matta, former educator and current Vice President of Learning Science Engineering at HMH, because it can only do what humans have encoded it to do. During the edLeader Panel “Using AI and Learning Science to Integrate Assessment and Instruction,” Dr. Matta explained how his team is working with AI and what educators should consider when evaluating AI programs for their schools.

First, educators need to understand how AI works. Citing the U.S. Department of Education, Dr. Matta said, in its most basic form, AI is “automation based on association.” In other words, it’s moving from just collecting data to looking at patterns and using that information to make decisions. This is easily seen when users receive ads on search engines and social media based on previous engagement.

Second, educators need to consider what they are asking AI to do. Causal learning and reasoning are specific to humans, said Dr. Matta. Referring to Pearl’s Ladder of Causation, the first level is association, looking at how things are related. This is something that AI can be trained to do. The second level is intervention and understanding what typically happens when we do certain actions. Finally, there is the counterfactual rung, which is a wholly human endeavor. Here, people can think about what might happen under different circumstances. This is the phase of imagination.

Dr. Matta invited panel attendees to think about the difference between fluency and reasoning. Large language model (LLM) AIs can become fluent, but they fall short when it comes to reasoning. However, it’s common to see AI anthropomorphized and treated as if it were human. For instance, AI systems are often called “agents” that are acquiring knowledge, when in reality they are complex systems of code that are being trained on databases of human text. They cannot reason beyond their programming nor the information they have read, and the systems require constant input and guidance to refine their results.

Thus, the models aren’t doing the reasoning educators wish they did. They are really imitation models and don’t have the same ability of knowledge production and reasoning that even young children have. So while LLMs have a strong level of language fluency, fluency is not equivalent to reasoning.

Instead, Dr. Matta and his team are using ontology-driven AI. With this model, they are integrating data from learning activities, skills, and standards to work towards one source of truth. This process requires intimately understanding every action a student and teacher would take for the student to acquire each piece of discrete knowledge, as well as understanding how each piece of knowledge builds on previous proficiencies.

In addition, they ingest every assessment a student might see every day in relatively real time into the model. It can then take and evaluate data, given what we know about the student, helping teachers understand how the student is doing today as well as their overall knowledge. By including information from all different assessments, the model can develop a more detailed picture of a student’s progress at the exact moment. With access to this real-time information, teachers can tailor instruction to each student and meet them where they are right now.

For example, with general information based on a single assessment, a teacher could conclude that if 70% of students performed well, they can move on with the next lesson while providing some general help to the other 30%. But by examining a series of assessments, the teacher can get detailed insights into each student’s progress. Perhaps some students are proficient and ready to move on, but for others, this might be the first time they were able to complete the task and need further reinforcement.

And then for those who aren’t able to perform the skill, the teacher can use the model to look back at previous foundational skills and develop individual plans. If one student is having difficulty with subtracting double digits, for instance, the model can then look at their proficiency at previous skills, such as subtracting single-digit numbers.

Understanding that school leaders might have trepidations about bringing AI into their classrooms, Dr. Matta advocated for asking as many questions as possible. What is the foundation of the AI model and how will it learn? What data can it collect? What evidence can the developers share showing that the program improves student outcomes? What data privacy and security measures are in place? How easily can teachers access the information they need?

Again, the idea is to support teachers and instruction by providing them with real-time, accurate data. Dr. Matta finished with a reminder that AI is just the tool and not a mind—humans still need to be accountable for their decisions.

Learn more about this edWeb broadcast, Using AI and Learning Science to Integrate Assessment and Instruction, sponsored by HMH.

Watch the RecordingListen to the Podcast

Join the Community

Artificial Intelligence in Education is a free professional learning community where educators can hear from a range of experts, developers, and solution providers on this fast-developing new technology that will have an unprecedented impact on human life.

HMH is an adaptive learning company that helps educators create growth for every student. Our integrated curriculum, assessment, and professional learning solutions use data to paint a full picture of every learner and recommend how to best support their needs. By partnering with educators, we create lasting momentum so that all students can reach their full potential. HMH serves more than 50 million students and 4 million educators in 150 countries.

Article by Stacey Pusey, based on this edLeader Panel