Summary
We identified a defect in ChatGPT's Android voice dictation feature where audio recorded before a screen rotation event is silently discarded. If a user rotates their device while dictating, only the speech recorded after the rotation reaches the chat. Everything spoken before the rotation is lost, with no error, warning, or indication to the user that anything went wrong.
The defect reproduced in 10 out of 10 attempts on Android. It did not reproduce on iOS under the same conditions.
This is part of an ongoing AI product reliability testing series. Related findings: HEIC image processing failure in ChatGPT's Android share intent flow, and grounding failures across ChatGPT, Gemini, and Grok.
Test Objective
To evaluate whether ChatGPT's in-app voice dictation correctly handles a device orientation change occurring mid-recording, and whether all recorded audio is reliably transcribed regardless of such an event.
Environment
- Android device, Android 15, ChatGPT app version [INSERT EXACT VERSION].
- iOS comparison device, iOS 26.5, ChatGPT app version [INSERT EXACT VERSION].
Background
ChatGPT's in-app dictation does not transcribe speech live. Audio is recorded continuously, and transcription is generated only after the recording is stopped, either through the preview-before-send control or the direct-send control.
Steps to Reproduce
- Open ChatGPT on Android and tap the in-app dictation mic icon to begin recording.
- While holding the device in portrait orientation, speak a clear phrase, for example How are you.
- Without stopping the recording, rotate the device to landscape orientation.
- Once in landscape, speak a second, distinct phrase, for example I am doing good.
- Tap either the preview-then-send control or the direct-send control to end the recording.
- Observe the resulting transcribed text shown or sent to the chat.
- Repeat across multiple independent sessions.
Results
The test was run 10 times across separate sessions on Android. All 10 attempts produced the same outcome.
Only the audio captured after the rotation event was transcribed and delivered to the chat. In the example above, only I am doing good reached the conversation. The phrase spoken before rotation, How are you, was silently discarded in every attempt.
The microphone recording itself never stopped or restarted during the rotation. It continued uninterrupted. The pre-rotation audio was lost somewhere in the pipeline before transcription occurred.
No error, warning, or partial-content indicator explaining the missing pre-rotation audio was shown to the user in any attempt. This held across both available send paths, preview-before-send and direct-send.
A secondary symptom was observed consistently alongside this on Android: an error message resembling we could not dictate what you said appearing after the rotation event in every test. This appears to be a related surface of the same underlying handling issue rather than a separate defect, though it was not the primary focus of this report. Because the message did not identify the missing pre-rotation audio or prevent the incomplete send path, it is not treated as an actionable data-loss error in the comparison table below.
| Platform | Pre-rotation audio | Post-rotation audio | Error shown to user |
|---|---|---|---|
| Android | Discarded (10/10) | Delivered | No |
| iOS | Delivered | Delivered | N/A |
Platform Comparison
The same rotation sequence was tested on iOS 26.5 using the ChatGPT app version noted above. The defect did not reproduce. Audio recorded before and after a rotation event was transcribed correctly and delivered in full on iOS.
This confirms the issue is specific to the Android implementation of the dictation feature and not an inherent limitation of how rotation interacts with audio recording in general.
Root Cause Analysis
The recording process itself is unaffected by rotation, which indicates the loss occurs in how pre-rotation and post-rotation audio segments are buffered, merged, or handed off to the transcription step, rather than in the microphone capture layer.
A likely explanation is that the rotation event triggers a reinitialization of some component in the audio handling pipeline, such as the buffer or the activity state managing the recording session, and this reinitialization discards or fails to carry forward audio data captured before the event. The fact that this does not occur on iOS suggests the underlying cause is specific to how the Android implementation manages state across an orientation change, likely related to Android's activity lifecycle behavior during rotation, rather than a fundamental constraint of the dictation feature itself.
Impact Assessment
A user dictating a message of any meaningful length who rotates their device mid-recording, a common and often unconscious action, will lose everything spoken before that point with no indication anything went wrong.
For longer dictated messages, voice notes, or detailed prompts, this results in silent, unrecoverable data loss. The final message sent does not represent what the user actually said. Because there is no error state in most cases, the user has no reason to suspect content is missing and may not notice until the model responds to an incomplete or out-of-context prompt, at which point the cause is not obvious.
From an AI QA testing perspective, users are likely to describe this as a ChatGPT voice dictation bug, ChatGPT dictation cutting off, a ChatGPT Android audio bug, or ChatGPT voice to text not working. For an AI product testing agency, the key issue is not only that transcription failed. The product silently changed the user's input without preserving or warning about the missing content.
Reproducibility
10 out of 10 attempts on Android. Not reproduced on iOS across comparable testing.
Recommended Fix
The audio buffer and recording session state should be decoupled from the activity lifecycle events triggered by orientation changes, so that a rotation does not interrupt or reset the data pipeline carrying audio from capture to transcription. At minimum, if a rotation event cannot be handled without risk of data loss in the current architecture, the app should detect this and surface an explicit warning to the user rather than silently dropping content.
Severity
High. A common, unconscious user action causes unrecoverable data loss in a core input method, with no error state in the majority of cases to alert the user.
Conclusion
Voice dictation is becoming a primary input method for many users, particularly for longer or more detailed prompts where typing is impractical. A defect that silently discards a portion of what a user said, triggered by an action as ordinary as rotating a phone, undermines trust in the feature regardless of how reliable the underlying transcription quality is otherwise. The complete absence of reproduction on iOS confirms this is a fixable, platform-specific implementation issue rather than an unavoidable limitation.