Persistent answer stacking caused by over-inclusion of prior conversational context

[wooden yew]

Reported by @west thorn

Bug Report: Persistent answer stacking caused by over-inclusion of prior conversational context

`Steps to Reproduce`

When a user asks multiple questions sequentially across separate messages, the model frequently re-answers previously resolved questions instead of responding only to the most recent one. This behavior compounds over time, resulting in excessively long and repetitive responses.

`Expected Result`

The model should, by default, respond only to the most recent user message.
Previous questions should be treated as resolved unless the user explicitly asks for a summary, recap, or clarification.
Prior context should inform understanding, not force repetition.

`Actual Result`

In a multi-turn conversation, the following pattern consistently occurs:
The user asks Question #1 → the model answers correctly.
The user asks Question #2 (as a new, separate message) → instead of answering only Question #2, the model:
repeats the answer to Question #1, and
then appends the answer to Question #2.
The user asks Question #3 → the model outputs:
the answer to Question #1 again,
the answer to Question #2 again,
and only then the answer to Question #3.

`Environment`

Android app

Dec 21, 2025, 10:59 PM

Additional Information

Please provide relevant details to help resolve the issue, such as:

• ChatGPT Shared Link (if applicable).
• Screenshots or videos demonstrating the problem.

-# ➜ Need to contact support? Visit the OpenAI Help Center.

[west thorn]

To better understand the cause, a controlled test was performed during the same conversation.
Test condition:
The user explicitly instructed the model to treat the next message as a completely new turn and to ignore all previous messages when generating the response.
Result:
Under this explicit constraint, the model stopped repeating previous answers and responded only to the latest question.
When the constraint was removed, the repetitive behavior resumed with high consistency.
This demonstrates that:
The issue is not caused by UI message merging or by the model being unable to distinguish separate messages.
The behavior is driven by the model’s default response strategy when prior context is available.
Model behavior explanation (observed)
Based on the interaction, the model appears to:
Treat all prior conversational content as actively relevant unless explicitly instructed otherwise.
Prefer a “context restoration” strategy, where it restates previously given explanations before addressing new input.
Lack a reliable internal mechanism to recognize that a prior question has already been fully answered and does not need to be revisited.
As a result, the model optimizes for perceived completeness rather than relevance to the latest user message.
This behavior becomes especially pronounced in:
longer conversations,
thematically continuous discussions,
sensitive or complex topics (e.g. medical or psychological subjects).