Prompt Moderation False Positives in French

[Karleuh]

Hello,

My company uses GPT via Azure and we’ve encountered a prompt moderation issue.

Logs:

"The response was filtered due to the prompt triggering Azure OpenAI's content management policy. Please modify your prompt and retry. To learn more about our content filtering policies, please read our documentation: https://go.microsoft.com/fwlink/?linkid=2198766"

Example of a prompt that triggers the issue (in French):

"Oui, je peux vous proposer la semaine prochaine, A vous de voir si vous préférez le matin ou l'après-midi. On peut dire, si vous voulez jeudi matin ou vendredi après-midi, c'est vous qui voyez. On aura à peu près 1h 01h30."

Translation in English (for reference):

"Yes, I can suggest next week. It's up to you to decide whether you prefer the morning or the afternoon, and then we can come deliver it. We can say, if you want, Thursday morning or Friday afternoon, it's up to you. It will take us about 1 to 1.5 hours."

Additionnal context

Sending the text in English does not trigger the moderation filter, which makes us think this may be a false positive related to prompt injection detection.

We have requested access to moderation settings and logs. In the meantime, any guidance or help regarding this issue would be greatly appreciated.

Thank you!

[amynic]

Thank you for sharing @Karleuh - could I confirm which specific GPT model in Azure Open AI you are using - 4, 4o, etc so we can make sure we have detailed information to take next steps

[Karleuh]

Hello, thank you for the quick reply !

The issue is most commonly encountered with GPT- 4.1 (which we use most) but we were able to reproduce it with 4 omni as well. The sample prompt i provided should reproduce the issue for both models.

Our current temporary fix is replacing the word "proposer" by a synonym, "suggérer" in user inputs, which limits the issue significantly :
There might be a problem specifically with the word "proposer", although it's not the only culprit.