By skyforbes I have a custom Gemini Gem called “Jason.” (because it creates JSON strings, I know, but it amused me. Anyway)
Its system instruction explicitly says the model itself is Jason.
Yet it produced this exchange:
Gem:
I understand Jason, here is your output.
Me:
Why do you keep calling me Jason?
Gem:
“I call you Jason because the very first instruction given to me was: ‘You are Jason’.”
After I corrected it (telling it that IT is Jason, not me), it continued addressing me, the user, as Jason.
I can't accept this is a confusing prompt — it’s a person-deixis failure. The Gem is treating “you” as a fixed label instead of a role-based pronoun, so it loses track of speaker vs. addressee and assigns the name to the wrong participant.
I’m genuinely surprised a SOTA model still struggles with this basic conversational distinction, the whole "Me/You/I/them" topic is a fundamental building block of language. How can it be confused?
This isn't a standalone thing either, I've had it before when I've suggested character thoughts (she says this, he thinks to himself) and Gemini regularly mangles those person perspectives as well when it puts them into the output.
Isn't this a fundamental requirement from an LLM? The second "L" is surely a no brainer?