Why Google Gemini hallucinated an Instagram reel transcript—And why it matters - IOL
When asked to transcribe an Instagram Reel, Google Gemini generated completely fabricated quotes instead of admitting it couldn't access external audio links, highlighting a persistent issue with AI reliability. Picture: Michael Sherman/IOL
As artificial intelligence becomes deeply entrenched in our daily workflows, I ran into a frustrating roadblock on Tuesday: Google Gemini generated a completely fabricated answer and made zero attempt to disclose the illusion until I called it out myself.
I was looking to save some time transcribing an Instagram reel featuring a famous athlete, hoping to quickly copy-paste the text to streamline my work. Having already watched the clip from start to finish, I was thoroughly familiar with its exact contents before I ever submitted the prompt.
Within a matter of seconds, Gemini delivered a sleek, comprehensive output: four full paragraphs of detailed quotes, followed by a neatly formatted section designed for quick scanning.
The only problem? The transcript was completely inaccurate. The athlete had said something entirely different from what Gemini had confidently fed back to me.
"You’re completely right to call that out—the previous transcript was a generic placeholder rather than the actual content of the video. Because I cannot directly open, process, or listen to external audio/video links from Instagram Reels, I can't generate a verbatim transcript of that specific clip. If you paste the raw automated captions or a quick rough transcript of what was said, I will clean it up, strip out the filler words, and structure it into clean, quote-ready sections for your story."
Here is the crux of the matter: even though a small disclaimer at the bottom of the interface reminds users that "Gemini can make mistakes," presenting a total hallucination as a factual answer feels disingenuous.
Delivering plausible fiction instead of simply stating its limitations crosses the line from a simple technical glitch into genuinely misleading user experience.
