Am I helping us get to the “Robopocalypse”?

Fear of the “Robopocalypse” is everywhere: the idea that the world we live in is going to be taken over by AI (humanoid or otherwise), with the seeds of the destruction being planted in AI technology being created or conceived today. This isn’t helped by pronouncements from well regarded folks and doomsday scenarios playing out in movies and popular books (full disclosure: Terminator is one of my all time favorite films).

Since I work on applying AI to education, a view from the ground may add some context to this debate. I lead a team of engineers and research scientists building a “virtual tutor”. You can see some press on this project here and here.

Most people I discuss my work with are fascinated by the possibility and sometimes afraid of these ideas.

The positives

Remember when you were doing your class assignment in your freshman year? In my days, there was a lot of group study, and many of my better grades are attributable to learning from friends who understood the material more than I did. Later, when I was a TA in grad school, students would wait to meet me at “office hours” for help, when we had more time to go over fundamental concepts and provide some 1–1 attention they could not get in a class setting. There is rarely a substitute for a session with somebody who knows their stuff, knows what you need help with, and helps you increase your mastery of a subject with patience and encouragement…and, perhaps most important of all, without judgment.

Unfortunately though, I also had occasions where my friend who was really good with Circuit Theory decided to have an early night before my midterm, and my procrastination sunk me. I remember reading desperate emails or forum postings from my students who needed my help at the last minute, when I wasn’t always available.

A tutor that is always there when you need it and can help you review your understanding has the potential to improve engagement and educational outcomes for students everywhere…this is the true potential for an AI-based tutor. Properly engineered, it could guide the student through appropriate topics that result in an improvement in understanding of a concept, work through problems…and do so in a manner that exhibits the best qualities of a human tutor. Such systems have been conceived of and created since the ’70s. Recent improvements in NLP technology have allowed us to address the ‘human-like’ aspect of such tutors more readily.

Then again…

Let me discuss the fears that many have: if we have an AI Tutor that can help a student learn, how long before it can replace a teacher, and does this mean the end of human interaction in education?

Let’s first get the limitations of such systems out there. First, a tutor needs to know its stuff. This means it needs to know the structure, content and relationships between concepts within a particular domain. Even the scope of a domain is not easily definable...if I am tutoring a student on Mechanics, should I be able to provide a Calculus refresher along the way if they need it? Training such a tutor on the material is really hard. As hard as that is, it is almost dwarfed by a second problem: interpreting natural language. Deep learning has helped us make strides in NLP…but problems like dealing with complex grammatical constructs (particularly in English) that allow for the same thing to be expressed in a zillion ways, negation, the vocabulary of the subject matter itself…all imply that such systems are unlikely to get to a human-level of interpretation in the near future. Next…humans, particularly people with high emotional intelligence have the an amazing ability to detect nuances in non-verbal communication and tone. Even an AI in an ideal setting that is able to capture the gaze, facial expressions and other movements of a student would be hard-pressed to match that. Let’s say we solve all these problems, and a student’s response is perfectly understood. A tutor then needs to map the response (and context of the entire tutoring session) to concepts related to the subject at hand, and understand how to improve mastery in a manner that is engaging. Finally, teaching requires a personal level of information; for example if I’m teaching physics to a student who understands sports, I might discuss the flight of a football to illustrate a concept.

Some of the problems I describe are being addressed in different ways now, though the idea of all these problems being solved in the near term is quite low. Even if the technical issues are solved, an AI system could try to fake it, but is probably a long way away from expressing empathy, understanding of a student’s personal challenges and adapting to those. For those of you who truly fear the robopocalyse, the idea of such an AI becoming self-aware, talking to its AI buddies and taking over the world is fortunately still very much in the realm of fantasy.

In Conclusion

Our tutor is going to be embedded in Pearson’s interactive learning platform…we will be working jointly with our partners at Pearson to pilot this in the next few months, and are all working hard to get the best possible results.

I truly believe that such systems could be tools that help improve educational access, engagement and outcomes for millions. Instead of being replaced, I hope and expect teachers may be able to use these assistants to provide limited help when they can’t.

This is not to say that there is no danger in how AI could affect society. Personally, I believe we need to better understand automated trading algorithms, and ensure that any weapon systems that use AI (or any kind of automation) need a human in the loop. AI could eliminate or change jobs, and like with every technological shift, we need to prepare the labor force to minimize the disruption and capitalize on the opportunities that arise.

Perhaps, people may be freed up to do more creative work — like make more Terminator movies. I’ll be watching.

This article represents my personal views; I do not represent my employer (IBM) in any way.