- Education has long struggled to help all students achieve concept mastery
- With advances in AI, computer-based tutors could be one of the solutions educators have long sought
- See related article: How AI learning is changing higher education
Benjamin Bloom explained the 2 sigma problem in his seminal 1984 article of the same title. Briefly stated, the problem consists of the following:
- The average student who is taught by a tutor using mastery learning techniques outperforms 98 percent of students taught in a typical classroom.
- Society cannot afford to provide full-time tutors for every student.
- As a result, the majority of students fail to reach their potential due to the way we teach them.
Because there appeared to be no viable path to providing each student with their own personal tutor, Bloom called on educational researchers to “find methods of group instruction as effective as one-on-one tutoring.” Educational technologists weren’t so quick to give up on the idea of providing every student with their own individual tutor, however, and have made laudable progress toward the 2 sigma goal by designing what are called intelligent tutoring systems (ITS). Unfortunately, these systems are both difficult and expensive to design and build, and typically work only in a single domain of knowledge (such as algebra).
The recent development of large language models (LLMs) like ChatGPT has opened entirely new possibilities for the design and implementation of computer-based tutors with the potential to help all students achieve Bloom’s two sigma performance threshold.
While the underlying models like GPT-4 are more expensive to create than an intelligent tutoring system, they can span a wide range of domains of knowledge (from psychology to music history to computer programming to Spanish) once trained. They can also be “programmed” to mimic the interaction patterns of highly effective tutors in a number of ways, including lightweight techniques like prompt engineering and more resource-intensive techniques like fine-tuning.
Humans and computers as tutors
Computer-based tutors have a number of benefits when compared to human tutors. Computer-based tutors are available at any hour, day or night (or early morning). They never run out of patience. They don’t judge you when you don’t understand. The same tutor can work with you across several of your classes. And computer-based tutors are already significantly more affordable than a human tutor–a one-month subscription to ChatGPT ($20) costs less than a one-hour session with a typical human tutor.
Of course, human tutors have a number of benefits when compared to computer-based tutors, which we shouldn’t forget. Human tutors can build relationships of care and trust with students. Their encouragement is more real and meaningful. (It’s unclear how much of the 2 sigma effect that Bloom describes is attributable to this interpersonal aspect of the tutoring relationship – future research will help us unpack this further.) Finally, human tutors may know about local student services that fall outside of the traditional academic scope, such as a food pantry, which computer-based tutors don’t know about.
Human tutors and computer-based tutors share some strengths and some shortcomings. For example, both can be trained to use highly effective tutoring techniques when interacting with learners. On the other hand, both humans and large language models are prone to mistakes. Both will sometimes get things wrong, albeit for different reasons. And both can bring hidden biases into their conversations with learners.
Only time and careful empirical research will reveal whether or not LLMs will provide a foundation educators can build on to help all learners reach their potential. But the future looks incredibly exciting.
- AI hurts students’ writing and communication skills - September 21, 2023
- The case for ChatGPT as the ultimate educator’s toolkit - September 20, 2023
- An alternate view of the enrollment cliff - September 19, 2023