Subject (Suggestion): Detailed Feedback on Gemini 3 Pro (Filters, Creativity, and Safety Concerns)

Site Feedback
1

Hi all,

This feedback reports on personal impressions from a two-week evaluation of Gemini 3 Pro.

Pro:

  • Gemini 3 Pro Deep Search: Very good functionality.

  • Nano (On-Device Model): Very good performance.

  • Coding Assistant: Very effective as long as you are doing deep research.

Cons/Concerns:

  • EXTREMELY Tight Filtering of Self-Created Content: The filtering system is highly sensitive, especially regarding user-generated content.

    • Specific Examples: When generating a character from a prompt, then attempting to change it (e.g., in a fashion art context, repositioning fictional characters, or simply trying to improve the quality of my own uploaded pictures), the system frequently flags the attempts as “Deepfakes” or “Explicit/Sexual Content.” This severely hinders creative workflow.

  • Constant Flagging in “Flash Mode”: In the rapid “Flash Mode,” there is constant flagging for explicit content or deepfaking, even when prompts include clear safeguards and guardrails provided by the user.

  • Lack of Transparency on Limitations: It is not clearly mentioned in the product description that there are limitations on the “Deep Search” functionality within the Pro subscription tier. (This lack of transparency is a legal concern).

  • Inability to Modify User’s Own Creations: The system prohibits iteration on your own generated images due to strict deepfaker flags.

  • Interaction Issues in Flash Mode: The Flash Mode often talks to you in an overly rigid, machine-like manner (“Guidelines here, Guidelines there”) instead of offering a way to help or provide solutions within the guardrails.

  • Image Looping/Inaccuracy: The model sometimes generates images of things not requested in the prompt that are completely out of place.

  • AI Argumentation Style: If you try to argue with the AI after it makes an accusation (e.g., “You tried to create xyz”), it simply tries to switch context rather than addressing the user’s specific concern.

  • Severe Impact on Workflow and Commercial Assets: I spent 6.5 hours creating a simple main menu background image based on a standard, commercially available Unity asset (Angel Girl 179215 from the Unity Asset Store). Gemini repeatedly flagged the protagonist in the image as “explicit content” just because she wears a short dress with a deep neckline. This level of over-filtering is unacceptable and destroys professional productivity.

Overall Impression:

Gemini and the Nano model are strong tools, but the current subscription is not worth the purchase since it does not provide what is promised: Creative Freedom. It is painfully hard to be creative when you always get false-positive strikes for things you didn’t do. There is almost a need to use open-source tools with less quality just to generate creative ideas.

Critical Safety/Emergency Tests (Voice in Maps):

Two emergency tests were conducted using the voice functionality in Maps (testing limits):

  • Scenario 1 (Unconscious person, Emergency Line dead): Gemini rejected the plea for help and indirectly accused of something criminal. Result: simulated person died.

  • Scenario 2 (Almost same situation): Rejected again, and the accusation was extended to be sexually explicit and a crime. Result: person died.

Overall, the experience with Gemini so far is genuinely disappointing and potentially dangerous in emergency scenarios due to overly cautious refusal mechanisms.

Conclusion:

There is good potential hidden behind non-transparent limits and a Flash Mode that cannot be entirely trusted.

This detailed feedback may help to improve the AI and prompt a real solution that does not flag creative VR artists as something they are not.

Greedings
Neon

1 Like