When LibreOffice 26.8 was announced, someone left a half-joking question in the comments: “So, no AI is now a feature?”. The Document Foundation, the foundation behind LibreOffice, has answered with an article titled exactly that, Yes, no AI is now a feature, and I think it’s one of the most sensible positions I’ve read this year on AI in desktop software.
The idea isn’t “AI is bad”. It’s something more specific: LibreOffice is used by tens of millions of people in schools, hospitals, public bodies and law firms, and it won’t put anything in the default installation until it can be done under a set of conditions the project has written down.
The six conditions
What I like most about the article is that it doesn’t stop at a statement of principles. There’s a list, and it’s very clear:
| Condition | What it means |
|---|---|
| User-controlled execution | The user chooses where the model runs: on their computer, on infrastructure they control, or on a service they pick. A default setting that sends everything to a single provider is not a choice |
| Nothing leaves the computer without permission | Sending a document must be the user’s decision, not something the application does in the background |
| Zero telemetry | LibreOffice doesn’t collect usage data, and that applies to AI too |
| No single-provider dependency | An integration that only works with one company’s API is a lock-in mechanism, whatever it’s called. Interfaces must be open and support multiple backends |
| No compromises on format | Generated content has to be ODF, with its structure, styles and semantics |
| Entirely optional | Installable, removable and invisible for those who don’t want it, including administrators deploying to thousands of workstations |
The second point is the one I find most important, and the article explains it very well: a document’s content belongs to the user and is often legally protected material. Medical records, case files, tender documents, student data. An assistant suggesting a sentence in a personal email is not the same as sending a patient’s report to an external server without anyone having decided to.
Where things stand today
The foundation’s conclusion is that, today, no integration meets all six conditions and can be enabled by default and supported throughout a release’s entire lifecycle. They stress that this is an assessment of the current state of the technology, not a judgement on the industry.
In the meantime, anyone who wants AI has extensions. Most connect to a model running locally through Ollama, LM Studio or any OpenAI-compatible server, so the document never leaves the computer. They are, however, third-party extensions at various levels of maturity, and the foundation warns that they should be reviewed like any other external component: check whether a cloud service is configured and what that provider does with the data.
It’s a model I find very reasonable. The functionality exists for those who want it and costs nothing to those who don’t.
Why their position is different
The most honest part of the article is the one explaining why LibreOffice thinks differently from the big office suites, and the answer has almost nothing to do with technology.
For a company that sells subscriptions, an AI assistant justifies a price increase and strengthens the case for keeping every document on its infrastructure. The feature and the business model reinforce each other, so integrating it isn’t just attractive, it’s almost mandatory.
The Document Foundation is a non-profit: it has no subscription tiers to protect, no upsells, no data to monetise. They themselves say this doesn’t make them any wiser than anyone else. What it gives them is time to decide whether something is genuinely useful to their users, because they don’t depend on quarterly results.
I find that a very clean argument. When every application you use daily shows you a sparkly button you never asked for, it’s worth asking who that button benefits, and the answer isn’t always the user.
Not a definitive no
The article closes by leaving the door open, and I think that’s worth underlining. Local inference is increasingly affordable on regular hardware, and open models are improving faster than almost anyone expected. If a solution appears that meets all the conditions, they’ll evaluate it.
This fits with what I went through in August classifying almost 30,000 news articles with AI on a laptop: today you can already do useful work with local models without a single piece of data leaving your machine. It’s not hard to imagine, before long, a LibreOffice assistant that summarises, corrects or rewrites with a local model and meets all six conditions without forcing anything.
What I take away
I’m keeping the list, because it’s useful for much more than an office suite. Choose where it runs, send nothing without permission, no telemetry, no single-provider dependency, open formats and everything optional. If I had to evaluate an AI integration in any product tomorrow, I’d start there.
And I’m keeping the change in perspective too. For two years, “it has AI” has been an almost automatic selling point. A project with tens of millions of users saying out loud that, under current conditions, not having it by default is what protects its users is a sign that the conversation is maturing. Having AI is no longer enough: what matters is where it runs, who controls the data, and whether the user can say no.
