They say that because LLMs are not infallible, just as humans are not infallible. But if you seriously don’t like to use LLMs as an aid to your learning because they might make a small grammatical error 0.1% of the time, that’s your decision.
Not being rude but we are language learning here, not launching rockets into space with people on board.
The article I posted said up to 15% error rate. 6% error rate was for complex Math problems.
Nobody anywhere is claiming 99.9% accuracy for LLMs. When I confronted AI with its errors, they would flip and say the opposite, and then proclaim “Now I’m right.“ Using AI for language learning was an exercise in frustration. It did not help me with learning complex language rules.
I’m not saying it will never get there, it isn’t there yet for language learning.
Again, I’m repeating myself.
The article you posted was 6 years old, that’s the real problem here. Do you know how bad LLMs were 6 years ago? Anyway you’re not going to try the newest versions of these LLMs so there’s no point arguing anymore. You’re probably a textbook / official material only kind of learner and that’s fine.
The first article was dated 2026. The second article was dated 2025. Chat-GPT still hallucinates, still makes errors. Hallucinating is when Chat-GPT makes stuff up. Still a problem.
Oh my god, this is going to be my last post explaining this because you quite literally don’t know what you’re talking about at this point.
The article was not posted in 2026, it was edited in 2026. ChatGPT 3 was released in 2020.
I’m not replying to you anymore.
The second article that I posted is dated 2025. I actually liked this article because it seemed to be a fair and neutral article talking about what ChatGPT does well and what it doesn’t. It compares
I have used all 3 versions. It was an exercise in frustration for learning a language. I have never had a job where an error rate of 5-15% was acceptable. Gee, I got 90% of the dishes clean. Who cares? It isn’t rocket science! (AI is not free, it incurs cost which at times can be significant. I don’t work for Clozemaster so I’m certainly not the decider.)
I’m looking to see if I can find a more current study.
Here is a quote from the article:
“What Are ChatGPT Hallucinations and Why Do They Happen?
Critical insight: Even GPT-5 occasionally presents fiction as fact with complete confidence.
“Hallucination” in AI terms means generating believable but false information. GPT-4’s hallucination rate dropped to 28.6% for citation accuracy (down from GPT-3.5’s 39.6%), but it hasn’t disappeared entirely.
Real example: ChatGPT might confidently cite a study that doesn’t exist, complete with realistic journal names and publication dates.“
Duolingo stories are good. They’re getting better. In section 5 (they’re are 8 for German), there’s 250 units and one story for each unit. And I’m finding Video Call with Lily very helpful. Unlike most AI conversations, it doesn’t show any text until after your done (transcripts, with helpful suggestions). DUO is getting a lot better. Add in Clozemaster, and those two are my favorites. And I’ve tried a LOT of apps. Clozemaster invaluable for me, and I see a ton of words I’ve learned in here show up in German subtitles when watching a movie or a series. Important words, the ones frequently used. Because I see them so often, they get activated from passive vocab pretty quickly.
But yes, story telling is a very effective method. 100% agree.
Thanks for the post! Do you have in mind generating the stories with a word missing from each sentence like in Cloze-Reading, or as stories you can read like you pasted in and optionally click on to get the translation, add to a custom collection, etc. (like in the mobile app for Spanish from English, for example, if you go into Account there’s a toggle for “Clozemaster Reading”, a new feature we’re experimenting with)?
As full stories we can read, no words missing. Essentially I just want to practice my reading, pronunciation and speaking speed. It would be nice as well if the entire story had TTS so we can shadow along.
You are committing a logical fallacy and draw a false equivalence.
Yet these two statements aren’t equally wrong. A Bull Terrier is still a dog of the terrier family whereas a stone isn’t even a living organism. When someone mistakes a dog for a stone, you can’t be as lenient with them as with someone who mistakes a dog for some similar breed of dog.
Stop justifying slop-producing LLMs with the observation that humans also make mistakes. It only reveals that you don’t understand anything about “AI”, neither about the technologies nor about the fascist eugenicist cult the people in Silicon Valley are in. Look up TESCREAL to learn more. This isn’t an attack, it’s well-intentioned advice. I don’t want you to embarrass yourself in front of people who know more than you.
This is blatantly false.
I studied both maths and AI.
LLMs suck at math.
Edit: According to your later comment, you haven’t finished university yet.
LLMs are comically bad at math. It could be funny if there wasn’t so much crime, fraud, destruction, and a literal death cult involved. (Again, check TESCREAL or Ghost in the Machine.) It’s depressing to see such falsehoods being spread.
OpenAI could solve the Navier-Stokes equations only because it stole the ideas that human mathematicians had typed into ChatGPT before the human mathematicians could publish their findings. And then OpenAI threatened the victims of their theft to stay silent or OpenAI would ruin their careers. And even then OpenAI still had to throw 22 million USD on the problem to solve a 1 million USD problem. What efficiency! Only 2200% of the prize money.
The “AI” CEOs are all lying sociopaths, the media is either also lying or doesn’t have the technical background to understand anything they write, and you believe them.
We are throwing trillions of dollars on “AI”. This is completely unprecedented. Imagine if we funded schools and teachers with trillions of dollars. Imagine what we could achieve if we ever gave anything else the same resources we now allocate to " AI": compilers, linters, fuzzers. But those never receive the same funding.
“AI” is completely underwhelming for all the money it has burned. It’s a dead end.
And it’s all just kept alive by accounting fraud and hype because it makes a few sociopaths unimagineably rich.
If “AI” is good at one thing, then it’s producing security vulnerabilities and getting hacked.
I agree with the sentiment, but I ask to use different terminology. We must stop anthropomorphizing “AI”. In simpler words, don’t make “AI” look like a human. It isn’t human, so don’t pretend it were. “AI” is a program on a computer, not unlike Spotify or Firefox, whose task is to produce statistically likely text (as opposed to playing a song).
ChatGPT doesn’t “make stuff up”. Making stuff up is what humans do and ChatGPT isn’t human. Likewise, ChatGPT doesn’t “hallucinate”, which again would attribute human characteristics to something that isn’t human.
It’s not that the magical ghost in the machine sometimes “hallucinates” as though it just ate a bunch of mushrooms on a spiritual journey to find its inner self. No, it’s that a bunch of sociopathic CEOs made their overpaid engineers (who are members of a cult) deploy a flawed technology that doesn’t work and (yet again) failed to produce the correct result. They should be held criminally liable, not be praised and admired and defended by using language that works in their favor.
If you tell Spotify to play Mozart and it plays Beethoven, Spotify didn’t hallucinate. No, it failed at its task and the software developers working for Spotify made a mistake at best or are incompetent at worst.
Once upon a time, we valued when technology worked and produced correct results. We didn’t excuse its fundamental flaws with cutesy language like, “oh it hallucinated again, it should reduce its dose of shrooms lol”.
In case of error, blame the humans who make the software, not some magical spirit within the software.
The people you should listen to are Timnit Gebru and Emily Bender, not random websites that repeat the harmful anthropomorphization and terminology.
LLMs suck at math.
Well that’s just obviously false. I’m not sure in which context you’re using the word “suck” but the quality of math, science, engineering, programming, pretty much everything these frontier models are doing now is insanely good. They are already as good as PhD researchers so give it one more year and these models will be displacing them.
Not to take this conversation too far off-topic or waste my time any more with a believer who won’t listen anyway, but you’re making all these grandiose fabulous unfounded claims, all of which are false (“next year AI will replace all PhD researchers” “what is your source?” “just trust me bro.”).
If “AI” is so good at math, then why did OpenAI need to steal research from humans?
In 2015 already, Google and Elon Musk promised we’d have self-driving cars and people on Mars within the next year. These people are lying, and every time people like you believe these liars. “Next year AI will work, give it one more year, just trust me bro.”
How would you even defend fabulous claims such as this?
They are already as good as PhD researchers
If we assume you’re not just making stuff up, how do you back up such grandiose claims with facts? By having the humans take an IQ test? And put some similar number on the latest LLM through a benchmark? And then compare these numbers? If so, by proposing IQ tests, you’re entering deeply eugenicist territory. But surely you’re not a racist eugenicist, so how do you back up claims like these with facts, preferably without eugenics? You can’t. You’re just uncritically parrotting what the people in the podcasts you watch, are saying: sentences straight out of the marketing departments of “AI” companies.
As I’ve written on multiple other occasions here in the forum already: You’re entitled to have your own (informed) opinion, but you’re not entitled to make up your own facts, or spread billionaires’ lies and unfounded falsehoods. Especially when you ignore every fact and source I’ve brought to the discussion. Your comments are unserious, not withstanding any closer examination and divorced from reality.
In one year you’ll be unemployed because the bubble has popped, the economy is crashed, everyone who bought OpenAI shares lost their money. That’s much more likely than what you are saying. I cannot take AI bros or their ideas seriously and you’ve outed yourself to be one.
Anyone reading along, please read this:
As I’ve already written here, if Clozemaster implements this idea, then the quote by Mike will have been a lie:
As far as environmental concerns, our usage of ChatGPT is quite minimal, and we will continue to minimize our use of it
I don’t see how adding ever more LLM-based features could count as “we will continue to minimize our use of LLMs”.
Edit: Since a number of people replied in the meantime that this quote is two years old already … Nothing about the validity of the arguments I made back then has changed since then in the past two years. My arguments from back then still are just as valid as when I made them. The common counter-argument, “but the planet-burning LLMs that are wrong all the time are useful” remains unconvincing.
Mike made that comment nearly 2 years ago. The difference between these LLMs from 2024 to now is staggering. But just like the previous user, you’re anti AI and that’s fine. I won’t consider arguing with you.
However if they do implement this feature, or consider adding other LLM related features, people like you who don’t like them can simply ignore them and don’t use them.
I don’t like the current cloze reading, so I don’t use it. Pretty simple really.
your comment seems to be driven by moral and not by the intention of looking for its potential. I myself oppose those big AI firms and think, that Ai shouldnt be in the hands of private companies, but much rather in democratic hands, fulfilling the needs of the society. But when it comes to grammar and speech there is simply no denying, that it is just great for creating grammaticly correct sentences and especially correct humans. It is normal, that natives in my country (germany) correct their grammar and commas by running it through an ai. Even profs in my uni recommend it because it just excells even educated humans. Actually I dont care about the math part, even though you cant really deny, that with this tempo it will overcome humans in the next years.
Even Microsoft Word is a good spell checker and punctuation marker.
I am not explicitly anti-AI. I haven’t had good success using it for learning a language. It is fine on the basics. Anything beyond that has been a crap shoot IME. The errors were blatant. It would say one thing, I would contradict it. Then it would say the exact opposite like it had not just said the opposite. If I asked, it would say, Now I’m right. When it started spinning like that, it was a sign to me that the subject was too complex for AI. AI itself told me that I was correct in that assumption.
We don’t know yet where AI can go or what it can do or how fast. Some of its behaviors are troubling. We don’t know what the limits are. Are the big statements of where it will be in a year because people believe it or because somebody who will get richer from it wants you to believe it? Sometimes progress on something just stops. Researchers hit a wall and they don’t know how to get around it. In any case, progress is rarely linear.
In any case, any output (not just AI output) has to be verified by a qualified person. Generating things by AI is often simple and fast, but vetting it for use is not. And nothing in software is simple. You change one line of code and it can fundamentally change the entire system. Every single change has to be tested.
Thanks for all the input!
@Alxndr thanks for clarifying.
That’s helpful, and it’s essentially the direction we’re exploring with the Reading feature in the mobile app I mentioned, so that’s exciting.
Good point! And interesting idea about expiring stories.
A couple of concerns came up:
On accuracy, @corgwin24 - good point that generating content is fast but checking it isn’t. If we move forward with generated stories, we’d likely look at reusing stories across learners at the same level instead of generating one per request, which would also let us run them through a proofreading layer like we do with other content. The tradeoff is that they’d be less personalized than the original feature request. Either way, the feature would be opt-in.
On LLM use, to be upfront: I made that statement nearly two years ago, and it no longer reflects where we’re headed. We think LLMs open up a lot of useful ways to practice, so our use will likely grow.
One request for everyone: the last several posts have gotten personal. Disagreement is welcome, but please keep it about ideas rather than each other. And let’s keep this thread focused on the reading feature. If you’d like to continue the broader discussion about AI, please feel free to open a separate topic.
Thanks for the feedback Mike. I actually think your option of fewer less customised stories, but proof read by humans is the perfect middle ground.
But I think to get this right, the stories must contain lots of words the user has already mastered. Otherwise it’ll just be completely out of touch with what they’ve learned.
Look forward to seeing something like this in the future.