Something very strange happened recently. I use Google Home Mini to announce some more important (to me) event that are happening around my house. For example when automaker did not returned to base within expected timeframe. To make it more entertaining I do not use predefined messages, but instead I ask ChatGPT to generate some short humorous message related to specific event, that is less than 15 words. So something like: Prepare short humorous message stating that automaker did not returned to base on time, using less that 15 words. Then I use response to announce such event using TTS. So far was working perfectly fine. Today I noticed that all responses from ChatGPT are proceeded with unwanted text ’ Notification have been send:’ followed by real response. Since this ‘prefix’ is constant, same text included in every response’ it is easy to strip away before sending text to TTS, but I’m wandering if anyone else noticed something similar? What could case such change and how to revert it?
Oh my! Looks both ways, suspiciously for any giggling bots, taking the Micky Bliss out of you…
Sorry to say, and you may find this ironical, but the first thing that popped into my head was:
“Did you ask ChatGPT for the fix?”
What was its response?
Tell it politely not to defacate in its own backyard.
Yes, I did :-). obviously first I analysed if something changed in configuration or any conditions changed that could cause such issue, but nothing found, with clear evidence that hat was ChatGPT that suddenly started to add such ‘announcement’. Regarding this; here is what ChatGPT itself thinks of it:
The most likely trigger is response-format drift, not a hidden prompt change.
Evidence:
- Your configured prompt was last modified in May and still says the same thing.
- Home Assistant 2026.8.1 upgraded the OpenAI Python client from
2.21.0to2.45.0. - The OpenAI integration code does not add that phrase.
- The phrase appears in the ChatGPT response itself and in the HA event log.
- Your agent has
llm_hass_api: ["assist"], so ChatGPT can see Home Assistant’s action/tool context. “Prepare a notification” may now be interpreted as a notification action, causing a confirmation-style response. - No matching
notifyexecution was found, so it likely only adopted that response style rather than actually sending another notification.
The prompt is currently permissive: it asks ChatGPT to “prepare a notification” but does not require notification text only. A model/backend update can therefore change the formatting.
The robust wording would explicitly require:
Return only the notification sentence.
Do not add labels, introductions, confirmations, explanations, or phrases such as “Powiadomienie zostało wysłane”.
Do not call Home Assistant services.
This points to an LLM/API behavior change after the update, not TTS, Google Home, or cached audio.