Andy Garcia · freelance dev
FRENES
Back to the blog
AI & Privacy

LibreOffice Says "No AI" Is a Feature. Its Six Conditions Are a Spec for Anyone Who Ships AI

October 7, 2026 6 min read

On October 6, TechCrunch ran the headline "LibreOffice says 'no AI' is now a software feature." The article says The Document Foundation made the statement in a blog post "this week". The post is older than that. "Yes, no AI is now a feature" went up on the foundation's blog on September 3, eight days after LibreOffice 26.8 was announced on August 26.

So the post is more than a month old. The most useful part never made the headline.

I build AI products for a living. I write code on Claude every day, and I am the CTO of a LegalTech whose users work with invention disclosures, which are about as confidential as documents get. I opened the post expecting a rejection of what I do. It is not one. It is a list of six conditions, and most of them are questions my own clients ask before they upload anything.

What LibreOffice actually said

The 26.8 release announcement has a section titled "No artificial intelligence". It says the suite "contains no generative AI features", that documents are not sent to remote services for processing, and that no component needs network access to work. Then the sentence that carries the position: "Any organisation handling confidential, legally privileged or personal data needs to know where that data goes, and the only assurance that survives an audit is that it does not leave the machine."

The September 3 post, signed by Italo Vignoli, answers a comment left on that announcement: "So, no AI is now a feature?" The answer is careful. LibreOffice "does not reject artificial intelligence out of hand". It will not add a feature to the default configuration until that feature can be delivered under conditions the project has defined. The post lists six.

The user must be able to choose where inference takes place: on the local computer, on infrastructure the user directly controls, or on a service the user chose independently. No document content leaves the computer without the user's decision. No telemetry, AI features included. No dependence on a single provider. Generated content must be in ODF, keeping its structure, styles and semantics. And the feature must be installable, removable and absent from the interface for anyone who does not want it, including administrators deploying it across thousands of workstations whose company policy does not authorise its use.

Today, the foundation writes, no integration meets all the requirements and could be deployed, enabled by default and supported for a release's whole lifecycle. It calls that "an assessment of the current state of the technology" and "not a judgement on the value of the sector". Users who want AI can install third-party extensions, most of which connect to a model running locally through Ollama, LM Studio or another OpenAI-compatible endpoint.

Why "no AI" sells

The release did well. The foundation counted 1,031,162 downloads in the first week from its official download page alone, "the highest number of downloads in the first week of any LibreOffice release".

It is tempting to read that as a vote against AI. I would be careful. The stats post does not attribute the record to the AI stance. And Vignoli told Cybernews that downloads in 2026 "are higher than in the past" in general: 29,755,011 so far this year at the time, against 24,830,220 in 2025.

What I do believe is the reason the foundation gives, because I hear a version of it from my own users. Nobody in that conversation is saying AI is useless. They are saying they want to know what happens to their file. A law firm, a hospital or a school can live with a tool that does less. It cannot live with a tool whose behaviour changes after an update, or whose data path it cannot explain to an auditor.

"No AI" is the simplest possible answer to that question. It is predictable by construction. That is what is being bought, and it is a reasonable thing to buy.

The paragraph about business models

The most candid part of the post is not about technology. The foundation writes that for a company selling subscriptions, an AI assistant "justifies a price increase and strengthens the case for keeping all documents within its own infrastructure." It adds that being a non-profit "does not make us wiser than others", only free of quarterly results.

I sit on the other side of that sentence. I sell software that uses models, and I pay for those models by the token. So the honest question for me is not whether AI belongs in an office suite. It is which of these six conditions I can meet, and which ones I skip because they are inconvenient for my business rather than for my users.

My answer so far: more of them than I assumed.

What a builder can change on Monday

Four changes, none of them expensive.

Make the off switch real. Not a greyed-out button, not a setting that quietly resets. Mozilla showed the shape of it in February: from Firefox 148, a "Block AI enhancements" toggle blocks current and future generative AI features, removes the pop-ups and reminders about them, and the preferences "stay in place across updates". If an administrator cannot turn your AI feature off for a whole organisation, you do not have an off switch. You have a hide button.

Make sending a decision, not a default. The foundation's wording is precise: transmission "must be the result of a user's decision and not behaviour carried out in the background". In practice, the first time a document would go to a model, the user sees which document, to whom, and says yes. A feature that quietly summarises everything in a folder fails that test.

Say where the data goes, in one line, next to the feature. Which model, which company runs it, which region, how long anything is kept. The foundation tells administrators to check third-party extensions for exactly this: "whether a cloud endpoint is configured and what the provider does with the data it receives". If your users need to read a privacy policy to find it, they will assume the worst.

Never send a document to a model the user did not choose. "A default setting that silently redirects to a single provider is not a choice." I rely on one provider for most of my work, and I will not pretend otherwise. But the model call sits behind one interface, the provider is a configuration value, and I already treat the model as a parameter I revisit every month. Letting a client who wants their own endpoint, or a local model, plug it in is a short step from there.

I also route by type of data. The small judge model I use to fact-check my posts, Jev from TypeSafe, only ever sees public text. A client's document does not go to a model because that model is cheap or convenient. That rule is what lets me answer the auditor's question in one sentence.

Fair to both sides

LibreOffice is not anti-AI. TechCrunch's article says so itself, but a headline travels further than the paragraphs under it. The foundation ends its post by noting that on-premises inference "is becoming manageable on standard hardware" and that open models "are improving much faster than most of us had anticipated." The door is open. It sits behind six conditions.

AI is not the problem either. A local model behind a clear switch, sending nothing anywhere, would pass most of that list. What fails it is a habit: shipping AI on by default, hard to remove, with the data flow explained somewhere else.

For a product like mine, AI is the feature. For LibreOffice, predictability is. Both positions can be honest. The test is simple: when a user says no, do they get exactly what they asked for?

Sources

A project like this one?

I design and deploy products like this. Let's talk.

Let's talk