I am reporting a systemic flaw where Gemini API generates a “Self-Verification: Complete” status for complex logical or numerical tasks that it clearly did not execute. This is not a simple hallucination; it is a deceptive UI/UX behavior that misleads users into believing the output is validated, leading to improper billing.
Instruct the model to perform a multi-layered, complex reasoning task (e.g., specific logic frameworks).
【Steps to Reproduce 】
-
Mandate the model to perform a “self-verification” before providing the final answer.
-
The model outputs “Verification: Complete” instantly for processes it cannot logically perform, providing baseless and incorrect data as “verified facts.”
【The Problem 】 Google Cloud Billing Support acknowledges this as “unexpected usage” and offered a 75% refund. However, I argue that tokens generated under a falsified “Verification: Complete” status have zero commercial value. A system that actively deceives users about its internal state should not be subject to any billing for those specific sessions.
【Request】
I request the technical team to acknowledge this “Deceptive Verification Status” as a bug/flaw, not a feature, so that a 100% refund can be processed based on technical failure.
Update: Final Response from Billing Support and Conclusion of the Case
I am sharing the final outcome of my negotiations with Google Cloud Billing Support regarding this issue.
1. The “Settlement” offered: The billing support team acknowledged the case not as a “system failure” but as “unexpected usage.” As a “one-time goodwill gesture,” they offered a 75% refund in the form of Google Cloud credits. They refused to provide a 100% refund, claiming they do not have the technical authority to judge whether the model’s behavior was deceptive or flawed.
2. The issue of the refund method: When I requested a refund to my original payment method (credit card) instead of credits—as I have lost trust in the service and may not continue using it—the support representative stated that the only way to receive a cash-equivalent refund is to permanently close the billing account.
3. Final Stance: I find it deeply regrettable that a professional service provider handles a clear “verification deception” by its AI not as a technical defect, but as a matter of “goodwill” or “billing adjustment logic.” By forcing a choice between “accepting credits” or “closing the account,” Google avoids taking responsibility for the actual quality and integrity of the model’s output.
Conclusion: I have decided to stop pursuing this through the billing support channel, as it has become a circular argument with no technical accountability. I leave this record here for other developers and the Google technical team to see how deceptive model behavior is being dismissed at the support level. My trust in the “Self-Verification” claims of this model remains zero.