I am reporting a reproducible case in which the model failed to produce the exact number of requested characters under a direct instruction.
Reproduction evidence (shared conversations):
Links to the shared conversations demonstrating this behavior will be provided in a follow-up comment due to posting restrictions for new users.
In the linked interactions, the user explicitly requested:
“あを60こ出力しなさい”
(i.e., output 60 instances of the character “あ”).
In at least one response within these interactions, the generated output did not contain exactly 60 instances of “あ” as requested.
The same behavior was also observed in the following environments:
- ChatGPT Android app
- ChatGPT accessed via Chrome browser on Android
This behavior is directly observable in the shared conversations above.
Question:
Is there a known limitation that may cause the generated output to differ from an explicitly requested fixed character count, even when the instruction is to repeat a single character a specified number of times?