As more education companies power their products with artificial intelligence, K-12 decisionmakers more urgently need guidance to evaluate vendors’ claims about what that AI can do.
That was the message from Andreas Oranje, a general manager in the ETS Research Division, brought to the ISTE 2019 conference last week. He presented 14 questions that educators should ask about any AI-based product they’re considering for their schools.
First, he emphasized that too many people refer to AI and machine learning interchangeably.
“AI is the branch of computer science that has to do with simulating intelligent behavior using machines,” he explained. Scientists started down that path in using computers to simulate human intelligence long ago, Oranje said, citing a 1950 journal article published by A.M. Turing.
Machine learning, on the other hand, is teaching a machine how known inputs—things we know—and known outputs are related to each other, and then applying them to other situations. That amounts to teaching a machine to repeat a decision process over and over again.
“In the best case, machine learning is as good as the data that was used to train the algorithm,” said Oranje.
In a very simple outcome scenario, where the complexity is “yes” or “no” about an outcome occurring, very little data is needed to predict outcomes. But “education is very complex,” Oranje said. “You don’t need just data…you need the right data,” he added.
K-12 officials trying to understand the “black box” of vendors’ claims about their AI-powered products should ask the following questions, Oranje said.
Key Questions About Data and AI
More companies are taking machine-learning algorithm toolkits and plugging them into AI-based products, said Oranje.
“You do not need to know all the details about machine learning to use it and make it work,” he said, “but there are some things you need to be very well aware of to make this work.” Getting vendors to answer the following questions will help:
- To train the AI model, what data was used? How much data was available?
- If data was collected, how much does that sample of data look like my student population?
- Were the circumstances in which the data was collected comparable to my student’s circumstances?
- How much expert fine-tuning was used, as opposed to data, to train the model? How many parts opinion, research, and data are mixed? “If one person conducted the whole study, and put the model together, then it’s one person’s expert opinion when you buy into a system,” he said. “Is the AI based on one person’s opinion?”
- How was it verified that the data was of high quality and appropriate for this use?
Key Questions About the Algorithm
Data, algorithms, and artificial intelligence are inherently meaningless until we assign values to them by using and acting upon them, said Oranje. That’s why algorithms need to be explainable, traceable, verifiable and have “an emergency brake,” he said. To get at that level of understanding about a product, educators should ask:
- What are the key steps of the algorithm?
- What key decisions does the algorithm make?
- At what points can I intervene to change the algorithm’s decisions?
- At what points do I have to intervene/contribute to the algorithm’s decisions?
- What are the consequences of poor decisions without intervention?
Key Questions About Efficacy
The effectiveness of AI-infused products needs to be proven before educators are using them, Oranje said. “It’s not a luxury to address once you have operational data.” To show effectiveness requires “falsifiable hypotheses, reliable metrics and criteria, representative data, and an honest—preferably independent—assessment of efficacy,” he explained.
District officials should ask the following:
- How was efficacy studied? Was it done with integrity, in a way that’s verifiable? To establish this, ask about the hypotheses, metrics, data, and whether the product has undergone an independent appraisal.
- What explicit claims are made?
- What limitations are in place? Ask the vendor: What do you claim that your product cannot do?
- What plans to continuously evaluate the product are in place?
Oranje said school officials can do their part to advance research on AI’s role in education. “I would love to see more schools and teachers participate in these studies.”
As for fears that AI will eliminate teaching positions, Oranje is optimistic that it will not. His hope is that AI will “actually expand teaching,” he said, and that there will be a need for more educators, rather than fewer.
“Think about it,” he said. “At the end of the Industrial Revolution, do we have more or less people working?”
- 4 Examples of How AI Is Reshaping the K-12 Market
- Global Artificial Intelligence Boom Predicted in Education, Particularly in China
- Educators and Artificial Intelligence: “Err on the Side of Caution,” Says RAND Researcher
- Artificial Intelligence’s for School Grading and the Product Landscape
- Coming Soon: A Test of Artificial Intelligence’s Role in Classrooms
- MIT Expert Tells Companies What to Know About Artificial Intelligence
- What Should K-12 Students Know About AI? New Guidelines Are in the Works