Fin occasionally inserts English words or foreign characters into Hebrew replies. How can this be prevented?

We see this issue a lot with teams running Fin in non-English markets, particularly Hebrew, Arabic, and Thai. 

Some things to try:

• If your articles were machine-translated from English they might still contain some English brand names, technical terms, or stray phrases.  The LLM can interpret that as a signal that it is ok to mix languages in its responses.
• If you have content already in Hebrew then turning off real-time translation will help - this prevents it from attempting to do its own translation (https://www.intercom.com/help/en/articles/8322387-set-up-fin-ai-agent-s-multilingual-support)
• Not quite applicable to the problem you are seeing but you can set a language override i.e. on the contact you can explicitly say what language to reply in and avoid it trying to detect (https://www.intercom.com/help/en/articles/9423842-enforce-the-language-of-the-messenger-fin-ai-agent-and-workflows-for-specific-users)
• In the Guidance name the specific things you do not want it to do; explicitly forbidding things is stronger (e.g. “Do not insert English words, transliterations, or any non-Hebrew characters (including Latin or CJK characters”)

Hi ​@בן גטניו and ​@Jon Vaughan , Seán here from the Fin technical support team 👋

I see ​@בן גטניו  already has a conversation already made with our support team and a ticket is with our engineers, I will update here when there is a fix for this!

I’ve just seen this for the first time with some of our replies as well. Our replies are in English but some random arabic words have been inserted.

Currently experiencing the same with Swedish text where EN words are being inserted into the chat because we have RTT enabled. 1) RTT needs to respect the user’s language and stay within.  2) Fin needs to link up with the Glossary feature to align with product terminology in foreign languages. 

Thanks ​@Peter Nixon great to hear. ​@Saad, ​@בן גטניו , ​@Jon Vaughan if there are new examples of any language please let me know here and I can raise it with our team. Thanks for your patience on this one 🙏

Thanks ​@Sean Meade Forum Support Lead this still continues to happen to us for SV language where RTT picks up from the EN HC some bolded words in English and inserts them into a Swedish translated chat. I had to resort to a workaround by adding a disclaimer into fin’s intro that some words may appear in English. Happy to explain more in a DM if needed.

Hey ​@Saad i’ve created a conversation for our Support team to investigate this in more detail. Thanks for flagging. 

It's concerning that clear language guidance didn't reliably prevent the issue. How does your team now identify incorrect or unexpected responses across live conversations, rather than relying only on customers to report them?

​@Parvesh 

Honestly, we don’t currently have a reliable way to proactively detect these issues across live conversations. We still occasionally come across replies from Fin containing unexpected English words or Chinese/Japanese-looking characters.

That said, we’ve noticed this happening much less often with the latest version of Fin. It seems to have improved, though the issue hasn’t disappeared entirely.

​@בן גטניו 

Thanks for being candid about this. I think the gap between noticing that an issue has become less frequent and actually having visibility into individual conversations is important.

This is one of the problems I've been working on with ChatShield. It helps teams review conversations across their Intercom workspace, including both Fin and human support agents, and flags potentially risky responses for human review.

It doesn't replace Fin's safeguards or fix the underlying generation issue, but it can provide an additional layer of conversation monitoring.

ChatShield is available on the Intercom App Store, and there's a 7-day free trial if your team would find it useful to test.