Yeah this is a huge point… it’s not congruent with Bunpros branding
Some people view products not having AI as a feature. I’d reckon adding AI is probably a bad business decision in general.
If there was an AI “Writing Dojo” then I’d want it to be in a separate bubble like the Kaijugation game is. Anything AI needs to have a very clear and explicit purpose that can only be completed inside Bunpro itself (and only with AI) rather than just ChatGPT strapped into Bunpro. (And low-key I would like conjugation drills to be added to the core SRS.)
I don’t understand why absolutely everything has to have an “Ai powered tool” of some sorts stuffed into it. I know we’ve probably lost that war, but we’ll keep complaining.
At this point, yes.
As others have said, there’s nothing stopping you from using it on the side to augment Bunpro. I do it all the time. When I come across a sentence I don’t quite get, or I want to understand the nuance between two similar words—just copy/paste into ChatGPT and ask away (but take the answers with a grain of salt).
I think there’s a valid use case for AI on Bunpro but not necessarily direct integration—as a tool for the Bunpro team to use behind the scenes (if they’re not already). I.e., use it to generate a bunch of candidates for example sentences that can then be reviewed by the (human) content team and accepted/rejected/amended as opposed to having to generate all the content from scratch.
Also keep in mind that tokens are getting expensive now, so if they did want to integrate it directly it would probably get pricey.
Just thought I would chime in to say we do not have any plans for “AI powered” features inside of Bunpro. We did include it as a question in the survey (which has sparked discussion) but that was largely to help gather sentiment/user insights, not as foreshadowing 
While we play around with AI behind the scenes from time to time (we would be remise not to at least explore it), our experience has largely been that it isn’t trustworthy enough within the context of a learning application for us to feel comfortable using the output directly.
Even for our own use cases where we find ways to save time spent on menial/repetitious tasks, we always end up checking and more often than not fixing/improving the output before any user ever sees it.
We aren’t strictly anti AI. We have always said users should find and use the tools they feel help them the most in their learning journey, even if that isn’t Bunpro. There are so many AI “tools” out there nowadays that I am sure people who really want to use AI can find something that suits them.
At the end of the day, while we sometimes still make mistakes, we are human after all, we feel it is very important that both we the team, as well as you the user, can trust the content on Bunpro.
I agree that AI has a place in language learning, and I use it personally. But I don’t think it needs to be injected into every single app/website/tool under the sun. Nothing is stopping people from going to one of the big sites and using it there. Just be sure to double check stuff, and ask for reference materials from known sources.
But why not just use AI yourself, do you need a wrapper?
Just a note here to avoid confusion related to my post on flashcard scheduling algorithms, since it was referenced here. We should really be mindful about what being “AI powered” really means.
The concept of “AI” is being shoved everywhere as marketing for any business to present itself as a top notch brand. But what “AI” means in a broader sense is not that well defined. Technically a simple comparison statement could already be defined as AI, and machine learning has existed since the 50s, being in constant development. But in the public eye that concept has been evolving to always denote the latest trend, or what is supposed to be the cutting edge technology.
The focus for research and development has been narrowing from Machine Learning, to Deep Learning, to specifically Transformer networks and from that getting into this boom of LLMs. And it’s now when the “AI” buzzword is being overused, because an LLM is the product that an end user interacts with directly, instead of being just a core technology invisible to the user.
That doesn’t mean that AI hasn’t been here for a long time before LLMs. Things like the recommendations we get on streaming services have always been machine learning. Now phones are said to have “AI cameras”, but computational photography has been on every smartphone since more than 10 years ago. The only difference is having now that crappy “AI” label to redefine what we already had.
So what I’m trying to say, is to ask people not to throw everything into the same sack. Not everything “AI” is about generative content for end users, processed at huge datacenters managed by big corpos, consuming natural resources and with large bills to pay.
I’d never use a Bunpro where LLMs are the focus, because I don’t want doubtful generative content in a platform that shines precisely because of its human curated content. I would however love a Bunpro that implements technology advancements wisely. Replacing their simple and completely deterministic SRS with an algorithm like FSRS where memory decay is probabilistic (that would be called AI a few years ago) is a step forward. And so would be if something more sophisticated falling into Deep Learning territory was implemented, it doesn’t mean that there will be a ChatGPT scheduling our reviews. That’s why benchmarks like the one I shared exist, to put in perspective the value of these advancements.
TL;DR: hell no to chatbots on Bunpro. Yes to better SRS.
If Bunpro incorporated LLMs I would cancel my subscription and never look back.
Out of politeness I’ll just leave it at that.
Just curious, how do you decide whether a usecase fits the use of AI?
I think it could help with analyzing what questions users get wrong and trying to pick bunpro questions that focus specifically on weak areas for the user.
I constantly run into recall problems with grammar that has the same English meaning, or that sound similar in Japanese, or that I have trouble conjugating the verb
つかわざる
つかざる
I think it could help with generating mnemonics, as long as the community could edit and vote on which ones work best.
ために 🆚 ように mnemonic
ために 
“TAME-ni = I feed the lion TO TAME it.” You’re the lion tamer: you personally hold the whip and steak, every move deliberate, all control in your hands. Feeding, cracking the whip — actions you do, on purpose, for a goal.
Phonetic lock: tame is literally the English word “tame” — a tamer’s job is 100% intentional action.
ように 
“YO-ni = I keep the egg warm so a chick will hatch.” You can warm the egg, but you cannot hatch it. The hatching happens on its own — you just create conditions and hope. That’s exactly ように: the outcome is a state that arrives by itself (見える, わかる, 話せる).
Phonetic lock: yō → the yolk in the egg. You never control the yolk — it becomes a chick (or doesn’t) by itself.
Key Takeaway
Steak in your hand = action you perform → ために
Yolk in the shell = outcome you wait for → ように
Test any sentence: “Is the verb a steak or a yolk?” 見える? You can’t “do” being-visible — it’s a yolk. → ように
I use AI quite often to compare and contrast nuance differences between sentences. Bunpro doesn’t do much active testing on nuance differences. The newly revised hint system helps but I’ve been stuck at the same levels for more than a year and don’t feel like I’m making much progress. I’m mostly still at an n3 level reading comprehension without extra effort, but can understand real world texts if I look up a lot of words. And go through several rounds of AI grammar explanations for difficult sentences constructions and colloquial language usage.
I’ve noticed that I tend to rely on English keyword phrases to help remember the grammar. It starts to feel like taking a shortcut just to get through my reviews because I get tired and review piles are not super fun. Some kind of active learning system would be really helpful. A Socratic Japanese sentence selector mechanism that uses your complete answer history to help you improve the fastest.
Adding AI for voicing the vocab sentences I think was a good move. Hearing the sentence helps me to understand more. I wish they’d also use it to read the Japanese hints, too.
I’m using jitaku, migaku, bunpro, satori reader, Manabi flashcards, and migii because bunpro is geared toward a specific, narrow type of learning.
AI is getting better, I but I know when it goes off the rails when explaining grammar. And I always double check by pasting in entries from bunkei jiten, and Dictionary of Japanese grammar (mono kakido app). The last bump from Anthropic, along with about 2 pages of agent harness have made the answers a lot better.
If there’s a way to help people learn Japanese faster and become proficient without them hitting long stretches of feeling stuck on an ability plateau, I would support that.
One other side note: renshuu.org has a lot of features that I wish were in bunpro. For example, you can study cards that have reading, listening, J->En, En->J at the same time. Each word is selectable and can be added to your learning queue.
This is more in the vain of what I’m thinking when I talk about adding AI to Bunpro, not a crazy change but just small ways to help improve the experience. Everyone learns a little differently and I think having more tools and ways to help is awesome.
I really hope that the Bunpro devs can get the api up and running soon. That would let people who want to use AI with Bunpro do so in their own enviornment while also not affecting the main Bunpro site.
I use AI on the side for practice, but agree that it doesn’t seem wise to integrate it into the site.
Having spent 30 years professionally in the Fortune 100 pharma/biotech space in technology with a significant amount of it based in AI ever since grad school (HMM, SVM, BPNN, all the way up to the current LLM fad), it isn’t exactly clear to me what your specific use case is for AI in Bunpro. I think the majority of folks have stated their opinion and skepticism. I am not going to pile on but I just want to make a point with regards to healthcare. AI is primarily relegated to decision support systems where it is a tool for clinicians but it is not making decisions. Most AI device approvals by the FDA have been made in the space of imaging for radio and RTM for cardio for pattern recognition. And the primary workhorse for both of these uses are CNNs and not LLMs. For what it is worth, I have since retired since 2024, but I still keep an eye on the pharma space for investment purposes.
People here arent even referring to the actual point you were trying to make. Some people just hear “AI” and, in their endless wisdom, immediately form their opinion. People here do realize that any advanced non-deterministic or reasoning algorithm is technically “AI,” right?
Instead, we get the same dozen’th discussion about the same things, and the same polarized or obviously biased positions feeling the inexplicable need to digress into their own ideology or opinions while the people actually in favor of the idea need to walk around eggshells. Because of this, 90% of people have not gone into even the slighest bit of detail of the question regarding AI since that would expose the actual point.
As for your question: Yes, of course AI could help a site that heavily relies on SRS by a landslide considering SRS is literally a type of reasoning algorithm in itself. It could make the experience of every user here better. But considering they definitely know this and could have done this ages ago, I doubt the bunpro Team will do anything. I mean, at some point they kind of have to but that might as well take another 10 years.
But I don’t blame them since running high-quality AI’s is pretty expensive, and no one likes AI slop.
maybe they could use ai for practicing conversations. i use chatgpt to help me with grammar so i can make more complex sentences, ai can be a good thing if you use it correctly.
It would also let companies like Google take their data and repackage it, intentionally or “AI unintentionally” into their own translate application which has started to shift toward Duolingo like features.
The issue I have is that the post is about introducing a technology, not about introducing a feature.
There’s a few ideas listed, which may or may not be nice, but the post reads as permeating a website which is based on very tightly authored content with GenAI fluff. So it’s no surprise people don’t want this irrespective of their prejudices on AI.
Now, I’m off to write a post about integrating Blockchain into Bunpro…
Diamond hand on the new Bun coin 
There are many domain specific definitions using the same words, AI can mean an npc reacting to its environment or it can mean neural networks, in this context its without a doubt referring to neural networks.
Your point on cost is very valid, I will not want my Bunpro subscription to go up, so they would have to make tiers (meaning they might actually spend a lot of time on something only 1% of the userbase will be using).
