Critical Input Injection Error – Phantom Transcript Text Without User Action
Summary:
A spontaneous input injection occurred while using ChatGPT on iOS (iPadOS), resulting in the assistant appending duplicate NSFW transcript warnings—“This transcript contains material not suitable for use by minors below 18 years of age”—without any user action or input. This is the second documented incident of this type.
No dictation, no typing, no voice prompt, and no interaction occurred at the time of insertion. This constitutes a critical security and trust violation, as it fabricates user-originated content in the transcript.
Details:
• Device: Apple iPad running latest iPadOS
• App: ChatGPT official iOS app
• Model: GPT-4.5 (confirmed active)
• Timestamp: May 4, 2025, ~evening EST
• User Action: None—hands were off the device. No speech, no typed input, no keyboard focus.
• **Injected Text**: “This transcript contains material not suitable for use by minors below 18 years of age. This transcript contains material not suitable for use by minors below 18 years of age.”
• **Behavior**: The assistant injected the above line into the active transcript without prompt. There was no ambient sound that could have triggered this. It appeared as if the user had said or typed it, but they had not.
Why This Is Critical:
• Trust Violation: The assistant falsely logged input as user-generated.
• Transcript Integrity Risk: This corrupts conversational records and could create legal, ethical, or moderation conflicts.
• Security Risk: Any system that can silently inject unprompted content can also be exploited for disinformation, impersonation, or malicious automation.
• Liability Risk: Falsely injecting sensitive or NSFW warnings into a transcript—especially without user intent—could have reputational or compliance implications.
Requested Action:
1. Acknowledge this is a confirmed bug class and assign it a tracking ID.
2. Investigate transcript injection mechanisms at the speech-to-text and input buffering layer.
3. Implement trace logging for spontaneous input injections for audit purposes.
4. Issue a patch or rollback for any recent update that may have introduced this behavior.
5. Restore pre-send transcription editing (previously functional) so users can verify input before it’s submitted.
User Note:
This bug has already occurred multiple times, and this report documents the most recent case with clear memory. I am an experienced power user and I can confirm with 100% certainty there was no interaction when the phantom input appeared :
Title : subject-critical-input-injection-error-phantom-message-appeared-without-user-action/1250963
