We launched our MVP, invite-only. It generates Japanese posts written for the Japanese market.
What happened
Some outputs carried a strange kind of Japanese. No grammar mistakes, no wrong words, but you stumble when you read it. The cause: phrasing that is natural in the writer's culture, carried straight into Japanese form.
The words became Japanese. The way of saying things did not.
Why this is hard
This is not an English-vs-Japanese problem. Every writer brings the phrasing habits of their own culture, so the patterns are effectively infinite. A mapping table cannot solve it. Language is a living thing, and so is Japanese. We are bridging two living things.
And there is no ground truth.
What I did not expect
I assumed the hard part would be funding, or more straightforward technical walls. Wrong. Much of this problem is still unsolved at the level of ideas. Testing with people close to you only shows you problems you can already imagine.
Next
I do not have the answer yet. First step: study how existing translation tools handle, or fail to handle, culture-specific phrasing.
Why
Language carries culture. A tool that generates Japanese that truly reads as Japanese keeps that culture alive.
If you've worked on anything like this, I'd love to hear about it. On my LinkedIn post, or message me directly.