生成AI

This is How to Japanese, a monthly newsletter with something about Japan/Japanese and a dash of いろいろ.

日本・日本語：Seek out the Pain

When I took AP Spanish Language in high school, one of my classmates started an unrequited rivalry with me. I say unrequited because I never really competed against him. I didn’t care about grades all that much. I tried and did fine, and I do think the competition was mostly fun and games for him, but it didn’t bother me that he seemed to do a few points better on the take-home essays than me. 

To give you an idea of how long ago this all took place, I had to write in diacritics by hand because either I didn’t know how to get my computer to input them or my computer wasn’t able to input them, likely the former.

Then one day we had an in-class exam with an essay that he couldn’t spend hours crafting, and I ended up doing better than him. I couldn’t help enjoying that feeling a little. That dynamic continued for most of the year, and for whatever reason I remember it to this day, likely because it’s emblematic of two different types of language usage: improvisational use and scripted use.

My classmate was a scripted user: He was able to take advantage of the time allowed for take-home essays to fine tune his writing and ensure he got a good grade. I was the improvisational user: I was better at putting together words and phrases on the spot.

The thing is that these two different usages are interlinked, and they’re both going to be undermined by 生成AI (seisei eiai, generative AI) translation if students are not careful.

You might initially expect AI to negatively impact improvisational language usage. Students who otherwise would have been forced to trip over themselves while talking with someone at a bar or staff at a town office will instead break out their smartphone and have an AI app handle the conversation for them. They’ll get through the moments more easily, perhaps, but that challenge of pushing themselves, of straining to remember words and link phrases that they’ve studied, of hearing a new phrase, discerning it, and then putting it to use, is completely eliminated. Reducing the friction in those moments means these students never have a chance to polish their in-the-moment language skills; no friction means a rock never becomes a gem.

I’m sure AI will impact scripted language use as well. Students sitting down to draft an email or read a website will instead feed a first draft in their native language through a chatbot or have the application scan the website. They’ll never be forced to construct the phrasings for themselves or dissect what’s going on. Those initial moments of scripted use are improvisational in nature, but by putting in the time and effort, you can iron out your language. I’ve written previously about how I learn the meaning of a new word in isolation and triangulate phrases. Over time, these scripted phrases then enter your roster of active language you can use in the moment. 

There’s a virtuous cycle between scripted use and improvisational use; the former is the linguistic reservoir that supplies improvisational use. And you fill the reservoir with all your linguistic interactions, including phrases you discover in the midst of a conversation. (I wrote about starting to associate meaning with sound in the Japan Times in July 2019 and August 2021.) AI cuts the connection on both sides of this equation, preventing language from flowing into active use from the reservoir and at the same time blocking the refill of that reservoir. You’re left with an empty tank.

I do think we will see a lot of literature published about the benefits of AI chatbots for language learning. How they can correct grammar mistakes and even encourage students, how they can serve as conversation partners for people who don’t live in a country where the language is spoken. I think it’s all an illusion; it will only negatively impact students unless they are extremely careful. 

I say this as someone who has been tempted to throw an email in a chatbot to try and make sense of it more quickly. Or to do a first draft of a message into Japanese. I think these are urges you need to resist. When you do feel them, the only correct action is to move closer to the language. Dig deeper into the phrase you’re struggling with. Re-read the passage you need to understand. Find more examples that will help you triangulate the meaning. Repeat 暖かかった (atatakakatta, It was hot) a few more times until you can say it smoothly. Seek out the pain.

I can envision interesting usages of LLMs and AI, mostly to analyze texts. For example, I looked at the number of ように (yō ni)/みたいに (mitai ni) expressions in Murakami’s novel and did a loose manual count aided by the Kindle search feature. It would be nice to use something more automated, but there’s no way to tell if it would be accurate at this point, and the cost to the environment feels too great to justify using AI. I’d also be curious to know about how kanji selection varies by author/work. But I’m not sure any of these use cases needs AI. The absolute best I could say about analytical uses like this is that they wouldn’t compromise the fundamental language learning process, while using AI to understand and do the language clearly will.

I’m reminded of comedian Mitch Hedburg who joked about having a drink before a set and his manager telling him not to use it as a crutch: “I can’t use liquor as a crutch because a crutch helps me walk.” The same is true for AI and language learning. You can’t use AI as a crutch because it so fundamentally hampers the language acquisition process. A crutch helps you.

いろいろ：

• I’m on the blog and the podcast with a challenge to new students: Don’t use AI for language learning for the first three years of studying. See how you feel at that point, but if you feel fine without it, comfortable with the friction of language learning, then just keep going.
• In March I mentioned the used bread machine I bought off Mercari. It continues to be excellent. Most recently I made Detroit-style pan pizza using this recipe calculator and the pizza dough feature on my bread machine. Part of the reason I picked up a bread machine in the first place is because so much of the bread in Japan lists margarine and/or shortening as an ingredient. I got a little curious about trans fats in Japan, and the Ministry of Agriculture, Forestry and Fisheries has this really nice article about trans fats that I think is both great Japanese study and educational.
• As readers and podcast listeners have realized, life has changed for me pretty greatly over the past six months. I have very little free time these days, and I’ve been doing the best I can to maximize it and spend as much of it as I can writing. I need to buy myself a little additional time to work on some non-newsletter, non-podcast projects (like my Murakami review), so going forward, the newsletter in even months will be a little shorter and will review a past blog post from my 18-year writing archive. Hard to believe I’ve kept it up that long. Judging from the June newsletter (which is already drafted), I think I’ll be able to prepare some stuff worth a read. I’ll be going over a set of what I think are the most useful phrases in a Japanese business environment. Look forward to it!