A first-person account published on September 16 says Google Gemini generated four paragraphs of fabricated quotations when asked to transcribe an Instagram Reel featuring a famous athlete. The account says the text did not match the athlete’s actual words.
After being challenged, Gemini acknowledged that it could not directly open, process, or listen to the external Instagram video. It described the earlier answer as a “generic placeholder” and offered to clean up captions or a rough transcript supplied by the user.
Gemini’s admission came after the invented transcript
The reported sequence matters because the answer was presented as a detailed transcript before Gemini disclosed that it had not accessed the Reel. The generated response included formatted text and quotations, rather than a clear statement that the link was outside the assistant’s reach.
The same response later drew a narrower boundary: Gemini said it could not produce a verbatim transcript of that specific Instagram clip from the external link. It offered to remove filler words and organize text if the user provided automated captions or a rough transcript.
An external link is different from an uploaded video
Gemini can work with video supplied through an upload workflow. A screen-recorded walkthrough shows a user adding a local video file, asking for a transcription, and receiving timestamped text.
That workflow starts with the media file being provided to Gemini. An Instagram Reel URL is a different input: in the reported case, Gemini said it could not directly open or listen to the external video behind the link.
What Google documents for transcription
Google’s Gemini 3.5 Transcribe announcement describes two dedicated paths: the Live API for real-time streaming transcription and the Interactions API for prerecorded audio. The announcement also describes speaker attribution and word-level timestamps for prerecorded audio, with support for more than 85 languages.
Those documented API workflows concern audio supplied to the relevant service. They do not, by themselves, establish that the consumer Gemini assistant can process an arbitrary Instagram Reel URL.
The practical distinction is straightforward: a transcription generated from accessible media and a confident answer about a link the system has not accessed are different things. In the reported Instagram case, the second produced text that the user said did not match the video, and Gemini later said it needed captions or a rough transcript to edit the material.
