AI Tutors Don't Know When to Stop Shutting Up

The chatbots don't know when to start chatting with learners.

The AI tutor space is incredibly crowded right now with dozens of different startups trying to find space from one another and keep their heads above some very frothy waves. Unfortunately, many of them can’t differentiate themselves beyond their user interface (is your submit button a rounded rectangle or does it have sharp corners?) and system prompting (is your tutor “patient and firm” or “sensitive and demanding?”). 

I wish them best of luck because none of those differentiators are likely to matter commercially or academically, given the significant ways chatbot tutors differ from human tutors. Here is another challenge to their success.

AI tutors don’t know when to stop shutting up. The chatbots don’t know when to start chatting.

Currently, the way a chatbot knows “it’s my turn to respond” is that a student presses the “please respond now” button.

The student has to decide they are done with their turn in the conversation and tell the chatbot “it’s your turn.” It seems easy for edtech operators to forget (or convince themselves not to remember) that this is not how good tutoring works.

Two of a tutor’s most important decisions are a) when to start talking, b) what to start talking about. It is frequently the case that the tutor will decide to start talking even when the student hasn’t asked. This is especially true because many students won’t ask—sometimes because they don’t know they need help and other times because it’s easier to run out the clock than confront the difficulties of learning. 

For example, in this video, Katrine Bryan decides to intervene in the work of a student named Alena, not because the student asked for help, but because Bryan noticed via our teacher dashboard that she had spent significant time on an incorrect line of reasoning.

For other examples, I asked teachers on Twitter:

If you're tutoring a kid 1:1, under what circumstances do you decide to intervene and say or ask something?

Nobody said “when the student asks.” Instead they described signals like:

If they seem to be hesitating I ask them to talk to me while they work so that I can hear their thinking. @mommy_piso

When I can see they are just guessing at what to do next. Or when I notice their verbal thinking makes sense but the execution of "showing work" is wack a doodle. @jodi_donald

I wait for the point of hesitation, where they begin to realize something isn’t right about their approach. @mathESource

When I see “the deer in the headlights stare”. @mathteacher24

When I can no longer "see the gears turning in their head", which means I think they're stuck in an unhelpful way. @eka_mark_moon

In case it isn’t obvious: chatbots can sense none of those signals.

What about multi-modal inputs of the sort we saw from OpenAI last month? That demo was really neat, but the tutor frequently interjected at inappropriate moments. Will a multi-modal input be able to detect signals from students like “the deer in the headlights stare” or “the gears no longer turning in their head” or “the realization that something isn’t right” or “the point of hesitation.” Will the multi-modal input understand whether a student’s pause means they’re thinking or struggling? Will the multi-modal input understand whether that struggle is productive or not? 

Time will tell, but I suspect that any technology that can sense and educate humans on that level will remake society so completely that improving education will be among the least of our concerns.

Featured Comment

Rob Nelson:

There is plenty of attention paid to OpenAI's "partnerships" as they roll out ChatGPT Edu, the LA Unified School District's announcement in March about its new chatbot, and every single demo of a new generative AI feature. What doesn't get covered? When the contracts don't get renewed because the product isn't so great or doesn't meet any real need. When the contracts don't get signed in the first place because diligent testing reveals the demo does not reflect how the product actually works.

Odds & Ends

¶ In the science-fiction novel The Diamond Age, author Neil Stephenson described “The Young Lady’s Illustrated Primer,” an interactive book that helps students learn asynchronously and remotely. That vision has, of course, enraptured huge numbers of education technologists since its publication. It was interesting, then, to find a decade-old Reddit “Ask Me Anything” thread where a user tells Stephenson, “My ultimate goal in life is to make the Primer real. Anything you want to make sure I get right?” Stephenson’s response goes straight to the heart of education: “Kids need to get answers from humans who love them.” via Andrew Sutherland.

¶ Imagine Learning has a new survey out on teachers and generative AI. I’m not going to link to the answers because the study population isn’t representative of anything more than “teachers from our Facebook group.” But I will call out these questions as interesting:

• How has the use of generative AI tools (such as Chat GPT, Bard, DALL-E, etc.) in the classroom impacted your job? AI has made my job [a lot easier / somewhat easier / neither easier nor harder / somewhat harder / a lot harder].

• Do you think the use of generative AI tools (such as Chat GPT, Bard, DALL-E, etc.) in the classroom can help bring teachers back to the profession? [Yes / No / Not sure]

AI usage is much lower at the end of the 2023-24 school year than I’d expect for a technology I have been told endlessly has obvious and transformative benefits for teachers and learners. At some point usage will creep up, though, especially as AI becomes embedded at different depths into existing products. Our survey instruments should change at that point to perceptions of actual impact on outcomes we care about—student learning and teacher retention chief among them. I’d love for Education Week or any of the other groups who are seriously surveying the space to take up these kinds of questions.

¶ MagicSchool strikes a partnership with Adobe. You can now use Adobe’s image generator inside MagicSchool. It’s wild to me that in their video announcement, they refer to zero (0) classroom use cases and address zero (0) teacher pain points. At a certain point, this becomes crazy-making for me. Are those use cases and pain points just obvious to everyone else? How far down the list of teacher needs do you have to go before you find “on demand clip art generator”? Much, much higher up that list for math teachers FWIW you’d find “generate math diagrams and illustrations from text.” On Twitter, some people told me, “ackshually this is already possible,” so I gave them four challenges along a range of difficulty and some opportunities for extra credit. Will update.