SideKick's Knowledge Base is written in Japanese. When I started thinking about providing English versions, I ran into one uncomfortable problem: I don't have enough confidence in my own English reading ability to judge accurately whether a generated article is good enough to publish.
What the problem was
Having an AI generate English articles isn't the hard part. The hard part comes after. Whether the generated English actually matches the meaning of the Japanese original, whether any information is missing, whether the phrasing sounds unnatural — judging all of that by reading it myself, as a non-native English speaker, leaves me uncertain about accuracy.
At the same time, trusting the AI's assessment unconditionally and auto-publishing goes against SideKick's principle that the final call is always made by a human.
What I changed
So instead of having the AI directly decide pass or fail, I designed it to write a report in Japanese — describing what it checked and what it found questionable. The actual pass/fail decision is then determined mechanically by the program reading that report.
I also anticipated cases where the AI might rewrite text too aggressively while trying to fix issues, so I added a step that calculates the similarity between the text before and after any corrections. If the text changes too much, the automatic correction is stopped and the decision is handed back to a human.
Why I went with this approach
Rather than taking the AI's output at face value, the idea is to keep a record of what the AI actually checked, in a form a human can read, and leave the final pass/fail calculation to reproducible code. That way, even though I can't read the English myself, I can read the Japanese report and decide whether the article is ready to publish.
This is also one concrete expression of something SideKick has consistently valued: the AI is a collaborator, and the final judgment belongs to the human.
What still remains
This mechanism hasn't been used on any actual Knowledge Base articles yet. The next step is to run it on one real article and see how it behaves in practice. I want to be straightforward that this is still a prototype.