An exciting new online tutoring platform has launched. It’s called Apangea Learning, and offers AI-assisted tutoring that customises content to the learner. The year? 2004.
Ten years later, the service shut down; the business model didn’t stack up: despite all their innovative work, demand was too low.
A year prior, in 2003, Skype launched. With it came new online tutoring platforms that made private 1:1 tutoring more accessible, more convenient, and cheaper. Yet online tutoring did not take over the world, just as Apangea’s AI learning programme did not. In fact, the % of children receiving tutoring has not changed much since the shift away from every child having a tutor (like the GOAT, Aristotle), to mass education in classrooms. Estimates vary by country, but less 2% of kids in the USA receive high-quality private tutoring.
Twenty years after that news story above, we’re seeing a second tidal wave of AI tutoring startups raising millions of dollars on the premise that LLMs will make high-quality tutoring available to everyone. Hastily-prompted, uncanny human avatars that can teach you Spanish are being launched everywhere I look. Startups like Speak, Praktika, Loora, and many fast-followers have scaled up slickly-edited influencer videos having ‘real conversations’ with AI tutors, and grabbed the early adopter market for the AI-curious app downloader.
The technology is certainly better than what Apangea Learning had to work with. LLMs produce perfect multi-lingual sentences. Voice models from OpenAI, Cartesia, and ElevenLabs are now at near-parity with a human narrator or teacher.
And yet, just as Skype did not unleash major growth in online tutoring, and Covid’s huge upsurge in adoption quickly reverted to the mean as schools re-opened, so will this wave of AI tutors rise and, mostly, fall.
Why? Because AI tutors do not – or perhaps cannot – generate and embed the social conditions for tutoring success.
I’m qualified to comment here. I’m biased in the wrong direction (I work for a tutoring platform… hey, we’re even building an AI tutor!), and I’m a lifelong user of tutors, as a dedicated language learner for academic and work purposes. I’m about as valuable a customer, in LTV terms, as a tutoring platform or AI Tutoring app can get. If I’m a little pessimistic about this space, you should be too.
So, let me qualify what I mean by ‘pessimistic’. I’m a believer in AI for teaching and tutoring, but not in the way the market has been trending, and VC dollars have been flowing.
Tutoring is a high-commitment, high-friction activity. Learning anything in a dedicated way beyond the confines of a classroom is a niche activity. Even at universities, which kids choose and pay to attend, the average students don’t want to learn, and can barely read. I understand why! It’s hard and painful to learn something new, and it’s a tiny minority of people who choose to do hard, painful things, like intense exercise.
When being tutored in a new language, or how to write functions in Javascript, my head would often physically hurt at the effort. I persist – and I think my experience is the typical one – because the person helping us do it is waiting for us, pushing us, and guiding us. We’ve formed a social contract with them that obliges us not to just give up. My strong feeling is having a human who I know, and I know is waiting for me, is the difference between showing up and giving up.
Tutoring in this sense is comparable to having a personal trainer, while non-human-tutored ‘learning’ is comparable to gym memberships. 20% of Americans are members of a gym, but churn is 50% within 6 months, and utilisation (how often they attend) is as low as 30%. In other words, while 1 in 5 people is a gym member at any given time, only 6 people in 100 are showing up.
We know gyms are good for us. We know learning is good for us. So what gives? I see gym attendance as a relationship with one’s own mind: can I be bothered to fit it in? Do I want to feel the pain? If I have my personal trainer waiting for me, motivating me, and shouting at me, I’ll show up, and might push through the pain. The human social contract I have with that instructor is what drives the learner.
Tutoring is very similar: the human expert and guide, often paid upfront, is our social contract-holder. We know they’re waiting, we know we’ve paid them. They can feel disappointment, they can motivate us through a high-friction activity (learning something), and they can text us or call us when we’re late to the next class.
Even when using human tutors, the market is largest in test prep age groups, where fear of failing exams – often felt most keenly by the parents – drives purchasing of tutoring plans. This pressure saw the growth (and post-Covid declines) of platforms like Byju’s and GoStudent, who sold long-term, high-cost tutoring plans to anxious parents alongside other content and products.
In both cases, and in the case of Novakid, where I work, and other tutoring platforms like iTalki and Preply, the presence of a human tutor is the driving factor in learner retention. In A/B tests that we’ve run, the content of a class has a negligible impact on a family’s decision to purchase a tutoring plan, while the teacher – their personality, their charisma, their fit to the learner – drives the decision. Anecdotally, based on the hundreds of reviews that flow into our Trustpilot and Google reviews sites weekly, parents cite the teacher’s skill and charisma as a critical driver of the child’s engagement.
So, why is Novakid building an AI Tutor? If I’m so pessimistic, what’s the point? And won’t improving models change all this? What about the amazing new emotive voices, or emotive video personalities, being released by companies like Anam and HeyGen?
Tutoring is a perfect use case for LLMs (when combined with a database) from a pedagogical perspective. Keeping a learner in their ‘zone of proximal development’, a state of ‘comfortable but challenging’ learning, is an intense task for a human brain, as it requires continually remembering what a learner knows, what is next for them, and producing optimal questions or content to match. Textbooks are good at drawing a ‘line of best fit’ for all learners, but are by their nature unable to optimise to each learner’s real abilities. Human tutors struggle, too – I certainly did as a teacher. LLMs connected to a server find this trivial, though! They can produce the perfect next question or sentence for your knowledge.
The problem isn’t the AI! The problem is us humans. We are – present company excluded – weak, feckless, unmotivated animals in constant pursuit of the lowest effort path to our next dopamine hit. There is a reason TikTok is more successful than iTalki. For the tiny number of us who decide to proactively learn something, the thing that gets us out of bed and into the learning flow is the human, social contract we hold with our tutor. It’s the embarrassment of not showing up for them, the fear of losing the money we spent hiring them, and the enjoyment of chatting to them as we form a bond.
I am of course aware that some people are forming bonds with AI Avatars. We’ve heard about the AI girlfriends from Replika and Character.AI, the AI therapeutic friends in Gen-Z apps, and so on. But those are all successful because they too hit the low-effort pleasure zones of the brain. They ask how you’re feeling, why you’re sad, whether you’d like to have a ‘romantic’ conversation, and so on. They’re not asking you to conjugate a verb or decline an irregular noun.
And what of Duolingo? It’s a huge, fast-growing app with metrics that look like Tinder’s. What did they get so right? They too learned to hit the brain’s pleasure zones. With pings and zings and fire emojis, they feed you 5-10 snappy little questions a day, starting easy and getting a little difficult, never asking you to do something too tricky like “speak in the language for 20 minutes” like an AI or human tutor (or real life language use) would.
And that is why the new breed of “talk to a human” AI tutoring apps will not reach far beyond their current early adopters, and will not cross the chasm, unless they transform themselves into a Duolingo-like experience that only very rarely forces the learner into the high-friction feeling of a long conversation. It’s possible that, with some advancements in AI voice and video, they are able to crack the code of driving a genuine social contract with their learner. Or maybe, both things happen. But the former, speaking as a product person, is a lot more easily done, and with shareholder pressure will come the transformation of many AI Tutors into Duolingo lookalikes with streaks and 2-minute daily conversations, rather than anything resembling a “tutor”.
And what of Novakid? Let me pitch our approach briefly. Our army of highly-trained, professional teachers deliver millions of live lessons, 1:1, every year. It’s remarkable how good they are at their job, judging by parent feedback. Kids love their teachers. Our goal with an AI Tutor is not to replace these social experiences, but to augment them: gap-filling on weak areas between classes, practising certain key vocabulary, using words aloud in a short, fun conversation.
Will it transform our business? Likely not. Just as AI voice calls will not transform Duolingo’s, but will act as a nice feature for the niche of users who are more committed, so will our AI tutor expand the speed of learning for the most dedicated, engaged learners, on a foundation of progress and engagement built by human tutoring.
Do you disagree? Write to me. For the sake of reverting humanity’s gradual slide into a cesspool of TikTok-watching zombification, I’d love to be proven wrong.
