Google’sDecember Pixel Feature Dropwas a huge one, bringing fun new functionality to older Pixels while delivering on several features promised for the Pixel 8 series. ThePixel 8 Pro, in particular, got its fully on-device foundational AI model, and it turned out that was none other thanGemini Nano, the lightweight version of Google’s multimodal AI platform meant to compete with GPT-4. Gemini Nano will be used to power features like an enhanced Smart Reply in Gboard and summaries forGoogle Recorder. That last one has started rolling out today, and it’s a lot more finicky than you might have thought.
After Twitter userAndreas Proschofskypointed out some interesting behavior andMishaal Rahmanbrought it to our attention, we did some testing and discovered that Gemini won’t summarize a transcript if the recording is less than one minute long — instead, you get an error message saying “Transcript is too short.” But your recording can’t be too long, either — if it’s over 15 minutes, the message will say “Transcript is too long.” The app suggests you try a different recording and informs you that “Summarization works best with medium-length transcripts.”
Google’s documentation for the featurepoints out another example of its AI being fussy: when transcribing, Gemini Nano “uses restrictions included in safety features to block harmful content.” Exactly what constitutes harmful content is not disclosed, but Google’sGenerative AI Prohibited Use Policy, which is linked in theToSdisplayed alongside these summaries, highlights “dangerous, illegal, or malicious activities,” as well as “content intended to misinform, misrepresent or mislead,” and “sexually explicit content,” as the three primary types of violations.
Google’s jawdropping Gemini demo video isn’t as impressive as it seems
Truthful misrepresentation at its finest
But when your recording is in that just-right Goldilocks zone of 1–15 minutes and it doesn’t include harmful content, things start to look up for the AI summarization feature. Tapping the Summarize button at the top of the transcript tab for a recording will cause a bottom sheet to appear informing you of the basics of the feature. Tapping OK here yields another bottom sheet, this time informing you that Recorder needs to download a 1GB AI model. Once that’s finished, you’ll see a summary of the recording in the form of a few bullet points just above the transcript.
The summaries aren’t anything groundbreaking — in practice, they’re not much different from the Summary section you see at the top of this article — but having the main takeaways at the top of transcripts will be a boon for students and professionals alike, and being able to do all of this with just your phone is a legitimate quality of life enhancement. But while the one-minute lower threshold makes sense when most such recordings won’t have much to summarize, the 15-minute upper bound is a significant limitation in real-world use cases.
