From our other sites
The story
The government has decided on a framework requiring AI operators to disclose information about the training data used for generative AI. The goal is to let copyright holders and AI users check whether their works were included in the training data upon request, but on 5ch, opinions split sharply over how to interpret Article 30-4 (“information analysis”) of the Copyright Act.
The “training is basically legal” camp and the “if similarity shows up, it could be illegal” camp traded blows citing past court rulings, and the thread turned into a marathon argument.
Operators will post information about their generative AI’s training process, the types of training data used, and how it was collected on their websites. If a copyright holder or AI user requests it, and certain conditions are met, operators will disclose whether copyrighted works are included in the training data — creating a system that lets people check for possible copyright infringement.
Source: news.yahoo.co.jp / Original article here
What people said
If the copyright holder decides it's bad, it becomes illegal.
Training isn't unconditionally allowed no matter what.
That first line is way too out there lol
No, I'm dead serious.
Even training becomes illegal if the copyright holder judges they've suffered a disadvantage from it.
Like, say, if it was used in a way tied to content revenue.
Revenue is about the output, not the training though, isn't it?
Is asking not allowed or something?
Probably because they don't actually know.
"Illegal if the copyright holder says so" — what a broken trump card.
At that rate you could just shut the courts down.
It IS a trump card.
1. Why is it "basically legal"? Japan's Copyright Act (Article 30-4) has a special exception: if the use is for "information analysis," you can use data without the copyright holder's permission. Since an AI reading huge amounts of data and analyzing features and patterns (training) is considered to fall under this "information analysis," no permission is required in principle.
2. Cases where it's NOT clearly legal (risk of being illegal). The following cases may violate the law.
Using it for human enjoyment instead of the AI: copying a work not to have the AI draw something, but purely so a human can enjoy the copy.
Training on pirated material: collecting and training on data known to have been illegally uploaded, like from piracy sites.
Cases that seriously harm the rights holder's interests: deliberately targeting and training on only a specific artist's works in a way that takes away that artist's work.
What's with the sudden essay-length lecture right after "it IS a trump card"? If that logic actually worked in practice, nobody would be fighting about this.
Exactly.
This is how believers twist logic among themselves until even the generation itself becomes "okay."
So let's remember: training is fine except when it falls outside "information analysis" — that part's still a no-go.
"Let's remember" — who do you think you are, playing teacher? Where's your source for this "anything outside information analysis is a no-go" ruling?
That's exactly the kind of argument that doesn't fly in the world of content rights holders.
For example, there are actual court cases where someone happened to draw something extremely similar, it was judged highly similar, and they lost in court.
So just chanting that phrase makes style reproduction invincible? Then what exactly are all these people fighting about?
All they'll be able to do is whine "but the AI said so!"
Pulling from actual court cases is credible enough for me.
What do you mean "what do you mean" — it's just having the AI produce its sources, that's all.
As long as this applies, normal use doesn't become illegal.
It's rough how you're treating that one phrase like a magic spell.
You're just declaring victory without even defining what "normal" means.
There are cases where something ended up too similar and got ruled out on similarity grounds, so it's not always the case.
Unfortunately.
Under Article 30-4 of the Copyright Act, it's legal if the use is judged to be "not intended to enjoy the thoughts or feelings expressed in the copyrighted work."
As long as no one deliberately extracts something hidden in the model on purpose, nothing gets infringed.
You're saying it won't be judged that way, then turn around and say "if it's judged legal, it's legal"?
Maybe start by rethinking your premise.
Like in the article at Re: #1, there's a whole trend toward protecting and disclosing intellectual property — if you can't even consider why that's happening and just keep chanting "legal, legal," you're starting to sound like a kid throwing a tantrum.
Aren't you just treating Article 30-4 like a magic spell?
If "stuff hidden in there is safe" were really the logic, there'd be no reason for this disclosure push to even come up.
Not sure what you're misunderstanding, but still.
That's why this guy just keeps asking Chappy (5ch slang for ChatGPT) and can't develop the argument any further than that, so he just keeps repeating the same line over and over — ironically, he's starting to sound like an AI himself.
He thinks just saying this one thing is enough, but he's got the premise wrong from the start… it's honestly kind of sad.
Yeah, classic true believer.
Article 30-4 in the first place only lifts copyright restrictions in specific situations — everything else still falls under normal copyright law.
So not all training automatically becomes legal.
He went quiet the moment someone pushed back — guess that's just how this guy operates.
What, is there some "lift restrictions" button you press? Article 30-4 is just a narrow exception — it's not a blanket pardon that makes all training legal.
Background and Key Points
Japan’s Copyright Act was revised in 2018 to add Article 30-4, an unusually permissive “information analysis” exception: using a copyrighted work without the rights holder’s permission is allowed if the use is not meant for a human to enjoy the work’s expressive content, but rather to extract data, patterns, or statistics from it — the theory AI training falls under. This is broader than US fair use case law and was written before generative AI existed, which is why Japan became known internationally as unusually AI-training-friendly. The new disclosure framework being reported here does not rewrite that law; it’s a transparency requirement layered on top, letting rights holders and users ask AI operators what went into a model and how it was collected.
The thread’s actual fault line is not “is AI training legal” in the abstract — both camps cite Article 30-4 — but whether a copyright holder’s own judgment that they’ve been harmed (especially via output similarity) can retroactively make the training itself illegal, or whether 30-4’s exceptions (piracy-sourced data, deliberately targeting one artist, non-analytical reuse) are the only carve-outs that matter. Posters swap fragments of real court rulings on similarity without agreeing on what those rulings actually settled.
What the thread never separates out is that training-stage legality (Article 30-4) and generation/output-stage infringement (ordinary reproduction and adaptation rights) are assessed independently under Japanese law — a training process can be lawful while a specific output still infringes, and vice versa. The disclosure rule addresses the former, not the latter, and by itself doesn’t determine anyone’s liability.
*This article is excerpted and summarized from the 5ch (Nnews VIP) thread “[Breaking] Government decides on law/guidelines pushing disclosure of AI-generated training data — is this the end for AI illustrations and videos?.”
Leave a Reply