iPhone notification summaries were made with GPT3.5 I believe (maybe even the -turbo version).

It doesn’t use reasoning and so when using very short outputs it can produce wild variations since there are not a lot of previous tokens in order to direct the LLM into the appropriate direction in kv-space and so you’re more at the whims of temperature setting (randomly selecting the next token from a SOFTMAX’d list which was output from the LLM).

You can take those same messages and plug them into a good model and get much higher quality results. But good models are expensive and Apple is, for some reason, going for the budget option.