LibreOffice didn’t win two million downloads by adding a feature; it won them by refusing to, and that distinction should rearrange how you think about every assistant you ship.
The facts are short. LibreOffice 26.8 landed on 26 August 2026 with no generative AI features, no document transmission to remote servers, and no requirement for network access to work at all. The Document Foundation said the quiet part out loud in its announcement: yes, no AI is now a feature. Downloads passed one million inside a week. They’ve since cleared two million.
I build bots for a living. My instinct when I read that was defensive, because the entire premise of my work is that a well-placed model makes software better. Then I sat with it, and I think LibreOffice stumbled into the clearest product research any of us have gotten this year, and they got it for free.
Users aren’t rejecting AI, they’re rejecting surprise
The Document Foundation was careful about this, and the care matters. They did not say AI is bad. They said that for office software used by tens of millions of people, certain principles have to be satisfied before AI gets added. That’s not a culture-war position. That’s a release-engineering position, and it’s one I wish more teams took.
Think about what a word processor actually holds. Legal drafts. Medical notes. Salary spreadsheets. Unpublished manuscripts. A thesis that isn’t ready for anyone’s eyes. When a feature in that application quietly needs a network connection, the user has to form a theory about where their sentences went. Most people can’t form that theory, so they form a feeling instead, and the feeling is unease.
What LibreOffice sold wasn’t the absence of a model. It was the absence of that unease. “Requires no network access to function” is a promise an ordinary person can verify by pulling out an Ethernet cable. That’s a rare property in modern software and it turns out to be worth a lot.
The contrast case is instructive
Collabora, which ships its own office suite built on the same lineage, went the other direction. Collabora Office 26.04 introduced AI as an opt-in feature, off by default. Two codebases, two philosophies, both defensible, shipped within months of each other.
And notice that Collabora’s choice is not the villain here. Opt-in and off by default is the honest version of adding a model. The user has to walk toward the feature; the feature doesn’t walk toward the user. If you’re going to add intelligence to software that holds private documents, that’s roughly the shape it should take.
What both releases demonstrate is that the default is the product decision. Everything else is detail.
What this changes in my own work
I’ve shipped bots where the model was load-bearing and bots where it was decoration. The LibreOffice numbers made me go back and sort my own projects into those two piles, and the decoration pile was bigger than I’d like to admit.
A few things I’m doing differently.
- Write the offline story first. If the core function of your tool breaks when the API is unreachable, you’ve built a client for someone else’s service, not a tool. Decide whether that’s what you meant to build.
- Make the data boundary a documented feature, not a footnote. “Your input never leaves this machine” belongs in your README’s first paragraph if it’s true. If it isn’t true, say exactly where the data goes and for how long it’s kept.
- Default to off for anything that leaves the device. Opt-in isn’t friction, it’s consent. The user who wants the feature will find the toggle.
- Separate the model from the primitive. Build the deterministic path, then add the model as a layer on top. That ordering gives you a working product at every stage and a clean answer when a user asks what happens if they turn the smart part off.
- Treat “no AI” as a legitimate configuration. Not a degraded mode with a nag banner. An equal, supported, tested path.
A note on reading the numbers
I’d be a bad journalist if I didn’t flag that the download figures drew scrutiny almost immediately. Over on Hacker News, commenters started picking apart the stats, pointing out that day-view and week-view numbers tell somewhat different stories about the release week. I can’t resolve that from here, and I’m not going to pretend otherwise.
But the precise multiple doesn’t carry the argument. A major release of long-standing open source office software got written about, argued over, and downloaded in volume because of something it declined to include. That’s the signal, and it holds whether the spike was two million or a fraction of it.
The uncomfortable takeaway
For those of us who build assistants, the lesson isn’t “stop building assistants.” It’s that users have started pricing in the cost of our features, and that cost is measured in trust rather than megabytes. LibreOffice found a market segment nobody was serving: people who want their tools to be tools.
That segment was apparently large, quiet, and waiting. Worth asking whether some of your users are in it.
🕒 Published: