On 18 April 2026, Gemini users on Android found the app had lost features they had been using for weeks. This was not an uninstall or an account problem: the app had rolled itself backwards.

What disappeared

The fault arrived with version 17.14 of the Google app, on both the stable and beta channels, and it took several specific things with it.

The ability to upload NotebookLM notebooks from the plus menu in the prompt box. The temporary chat button in the top-right corner. And the option to import memory into Gemini, from the profile menu.

The interface rolled back too: the prompt box returned from the list design to side-by-side pills, the black background reverted to the earlier grey, the suggestion pills lost their outline, the fullscreen glow effect vanished, and the Android overlay went back to its previous design with limited access to the Tools menu.

What you could do about it

Very little. Because this was a client-side regression combined with the configuration the server hands down, no setting and no reinstall would undo it. The only workable route was the web version at gemini.google.com, where the features were still there, and waiting for Google to fix the rollout.

It was not an isolated case

That April was a bad month for the app. Six days later, on 24 April, another bug reverted cars running Gemini for Android Auto to the older Google Assistant after updating. June brought a third that blocked hands-free voice calls, both in Android Auto and on the phone itself, which Google eventually fixed. In August, a regression causing crashes was resolved in version 1.2026.3170610, and the app's average rating climbed by more than a full point, from 3.12 to 4.20, as soon as it was.

What this really says

The lesson is not about Gemini. It is about what it means for a tool to live entirely on somebody else's server: features can disappear without you updating anything, without warning, and with no option to fall back to the previous version. Installed software can be frozen on a version that works. A service cannot.

It is the same reason running language models on your own machine pays off for a lot of people, at least for the things they use daily. There is a hardware cost worth calculating before romanticising it: the VRAM calculator says which card each model size needs, and the inference speed estimate says how fast it will actually answer.