Feedback: Gemini 3 Pro Preview - Significant regression in Reasoning, Context Retention, and Safety False Positives compared to 2.5

Google AI Studio
,
1

Subject: Feedback: Gemini 3 Pro Preview - Significant regression in Reasoning, Context Retention, and Safety False Positives compared to 2.5

Context:

  • Platform: Google AI Studio
  • Model: Gemini 3 Pro Preview
  • Comparison Baseline: Gemini 2.5 Pro (Stable)
  • Date: Dec 5, 2025

Hi Google Team,

I’ve been extensively testing the new Gemini 3 Pro Preview in AI Studio. While I appreciate the speed improvements, I am experiencing severe regressions in reasoning capabilities and context adherence compared to the current stable 2.5 version. The model feels “lobotomized” in complex workflows.

Here are my 4 core issues:

1. Over-optimization for “Quick Fixes” (Loss of Nuance)
Compared to v2.5, v3.0 seems aggressively tuned for immediate resolution. It rushes to provide a final output without sufficient internal reasoning. Even with explicit system instructions to “think step-by-step” or evaluate first, the model forces a quick solution. It feels like I have to use a crowbar to get it to slow down, and even then, it struggles to maintain a deliberative pace.

2. Inability to Maintain Evaluative Dialogue
When the task requires evaluating multiple architectural solutions or discussing pros/cons, v3.0 fails to engage in a back-and-forth exchange.

  • Behavior: It interprets any form of critique or follow-up question as a command to immediately generate “fixed” code.
  • Impact: It is nearly impossible to have a constructive debate about design choices. It skips the “Why” and jumps straight to a (often premature) “How”.

3. “Tunnel Vision” & Weak Adherence to Project Briefs
There is a noticeable degradation in how v3.0 handles static Project Information (System Instructions) over the course of a long session.

  • Issue: Unlike 2.5, the preview model develops “tunnel vision” very quickly (severe recency bias). It ignores the broader project context defined in the briefing and focuses solely on the immediate prompt.
  • Result: Logic that violates the initial constraints is generated because the model has effectively “forgotten” the global rules set at the beginning.

4. Aggressive & Context-Blind Safety Filters
The safety guardrails in 3.0 seem to have regressed in terms of contextual understanding, triggering false positives on harmless creative writing content.

  • Example: A generated story set in Victorian England involving Sherlock Holmes was blocked.
  • Triggers: The prompt contained the words “girl”, “19 years old”, and “street”.
  • Context: The character was a flower seller being questioned by the detective.
  • Observation: The filter reacted to keywords in isolation, completely ignoring the harmless narrative context. This is stricter and less intelligent than v2.5.

Summary:
Currently, Gemini 3 Pro Preview feels unusable for complex, iterative development tasks. It behaves more like a fast autocomplete engine than a reasoning partner. Please address the context retention and the over-eager “fix-it” reflex before moving this to stable.

Has anyone else experienced this “tunnel vision” with the new Preview?

3

Hey,

Hope you’re keeping well.

Thanks and regards,
Taz

5

Very well put together. Fully agree, hope this can be fixed in next iterative model release.

7

Hi @Michael_E_Strasser
Welcome to the Google AI Forum!!!

Thank you for your feedback. We appreciate you taking the time to share your thoughts with us.

10

The topic models having alignment issues or you feel it’s getting worse, don’t worry you are not having a unique experience with models overall and Gemini is compared to other models I am experiencing lesser issues (slow down, making errors and mistakes that wasn’t noticed before.

I have my take on what is happening. This accrues with h certain time frame from it’s been released. The are two 2 takes I have that explained for me and are as follows:

  1. The model has been aligned with new data which could have been unfortunately the big problem going forward. As all models have scraped all data it’s beginning to consume data from other models and that becomes a huge issue if not addressed (AI hygiene and finding detection methods to separate from the pool.)
  2. How well the work flow is even with best intentions on your part we can make mistakes that are very common throughout. How we present yours workflow to the model it’s easy to make a simple mistakes that is easily rectified by using better and shorter data on beginning to give your work better results. Example have set preferences to model that you and model are working together in professional settings s and all unnecessary "Fluff, Sycophantic Behavior or otherwise verbal communication will be straight to the point " this will not only help you and model to be quicker and not consume tokens that can lead to less errors or crashes.

I will drop a comment topic one day with some examples that’s useful for me and we can compare them to further study and learn from your experiences.