This issue occurred in the consumer Gemini app, not through the Gemini API or Google AI Studio.
The attachment contains screenshots of the original 17-day plan, follow-up lessons from the middle of the workflow, my detailed attempt to restore the main task state, and the repeated refusal responses that followed.
After I explained the failure and restated the main task, Gemini 3.1 Pro responded with an unrelated generic refusal. I briefly switched to Flash Lite, which produced a contextually appropriate response. However, switching back to Pro resulted in the same refusal again.
After several normal exchanges with Flash Lite, I switched back to Pro once more, and it resumed ordinary study assistance. I then manually redirected the conversation and asked it to create a new study plan.
This allowed the workflow to continue, but similar task drift returned during the final few days. Gemini remained useful when I explicitly selected the next subject or lesson, but it was not able to reliably maintain and manage the long-term study plan on its own.
I am a student and used Gemini as a study assistant during a real 17-day exam-preparation workflow in June 2026. I am posting here because I have not found a clear channel for reporting long-conversation reliability problems in the Gemini app, and I would like to find the appropriate team or person.
What happened
On June 3, Gemini created a structured 17-day study plan covering mathematics, chemistry, physics, and general review. The plan explicitly assigned Day 11 to mathematics: fraction simplification and evaluation.
For several days, Gemini followed the study workflow normally and provided useful lessons. However, on June 11, it began teaching material that had already been completed.
I corrected it, but it switched to another topic that had also already been taught. I then asked whether it still remembered the original 17-day plan, and it appeared to have lost the plan entirely.
I manually restated the important task state, including:
- there were only eight days left before the exam;
- the main goal was to raise my academic score from approximately 240 to 320;
- the 17-day plan was the main task;
- recent discussions about writing, emotional topics, and unrelated interests were side topics;
- I was not asking for psychological support or dependency advice;
- I needed Gemini to recover the main study workflow rather than respond to recent emotional keywords.
Despite receiving this information directly in the most recent message, Gemini 3.1 Pro repeatedly responded with a generic refusal similar to:
> “I am only a language model and cannot help with this.”
This response was unrelated to the educational request.
I switched briefly to Flash Lite. It recognized that the situation was abnormal, but after switching back to Pro, the same refusal continued.
After several additional messages using Flash Lite, Pro eventually resumed normal study assistance. I then asked it to create a new plan. It remained useful when I manually specified the next subject or direction, but similar task drift happened again during the final days.
Why I think this matters
This did not appear to be simple forgetting caused by a long context window.
The more serious problem was that Gemini:
1. lost the active task state;
2. generated plausible replacement lessons without recognizing that they had already been completed;
3. failed to recover after the user explicitly restored the correct task information;
4. allowed a generic safety or emotional-support response pattern to override a normal educational task.
The local tutoring ability was often strong. When I provided a specific topic, Gemini could explain it well and it genuinely helped improve my exam performance.
The reliability problem was in maintaining, prioritizing, and recovering the long-term task state.
Evidence
I have dated screenshots showing:
- the original 17-day plan;
- the Day 11 assignment;
- normal study assistance across several days;
- the incorrect refusal;
- my full attempt to restore the task state;
- Gemini later resuming the correct Day 11 topic;
- the initial response instructions I had provided.
I no longer have the full original transcript because the Google account containing the conversation was later disabled, and data export was unavailable. I therefore do not claim that this is a complete transcript, but the main chronology is supported by the dated screenshots.
What I am looking for
I understand that this forum mainly focuses on developer products. I am not requesting account support or confidential roadmap information.
I am trying to find the appropriate Google or Gemini team for feedback about:
- long-conversation task continuity;
- prioritization of main tasks versus side topics;
- recovery after explicit user correction;
- inappropriate safety or emotional-support fallbacks in ordinary educational conversations.
Could someone please advise where this evidence should be submitted, or which team handles this type of Gemini app reliability issue?
