Did “no AI” cause LibreOffice 26.8’s download record?

An audit of the claim in Manual do Usuário (8 Sep 2026) that LibreOffice 26.8 “became the app’s most popular update… after the foundation behind it declared that LibreOffice doesn’t come with generative AI features”. The test: is the record surprising once you account for the trend, and if it is, how many other things changed at the same time?

1. What the record actually looks like against the baseline

TDF publishes very few release-week figures. Only two “first week” posts exist for the current versioning era, plus an annual total for 2025. Everything below is built from those three numbers; anything derived from them is marked as such.

LibreOffice 25.2, first week (Feb 2025) 647,961 Average week in 2025 (44.8 M ÷ 52, derived) ≈ 861,700 LibreOffice 26.8, first week (Aug 2026) 1,031,162
Official download-page counts only; distro repositories, Microsoft Store and Mac App Store are excluded from all three figures. Sources: TDF blog, 25.2 stats, 26.8 stats, 2025 annual report.

The framing “most popular update ever” hides the scale. The record week is 1.59× the 25.2 launch week eighteen months earlier, but only about 1.2× an ordinary week in 2025. A launch week that beats a routine week by 20% is what you would expect from any major release; the question is only whether 1.59× over 18 months exceeds the trend.

FigureValueStatus
25.2 first-week downloads (Feb 2025)647,961published
26.8 first-week downloads (Aug 2026)1,031,162published
2025 annual downloads (record year)44,809,742published
2025 update-page visits581,615,673published
Donation transactions, 2024 → 2025104,430 → 140,593 (+35%)published
Social engagement on launch post, 25.2 → 26.811,313 → 50,061 (4.4×)published
Implied annual growth needed for 1.59× in 18 months≈ 36% / yrderived
25.8 first week (Aug 2025), 26.2 first week (Feb 2026)not published
Weekly download series for 2026not published

The two missing launch weeks matter most. If 25.8 and 26.2 sit on a straight line between 648k and 1.03M, 26.8 is trend. If they were flat at ~650k and 26.8 jumped, something specific happened in August 2026. TDF has that data and hasn’t released it; the claim in the article is unfalsifiable until they do.

2. Is 1.59× surprising given the trend?

Independent signals suggest the underlying demand curve was already steep before anyone wrote “no artificial intelligence” in a headline:

Under a plain trend model, 26.8’s week is roughly where a major-release week in late 2026 should land. That leaves little residual for “no AI” to explain — but a trend model assumes nothing else changed, and several things did.

3. What else changed between 25.2 and 26.8

Each of these is a competing explanation for part or all of the +383k. None is exclusive of the others.

Candidate driverWhenMechanism on first-week downloads
New Windows update mechanismQ4 2025, after 25.8Existing users are routed through a new-features page; TDF credits it for the Q4 2025 donation surge. If it also nudges users to fetch the installer from the website on a major release, launch-week counts inflate mechanically. TDF itself says it can also suppress counts (update without re-download) — direction unclear, magnitude unknown.
Microsoft 365 Copilot price increaseJan 2025 onwardPushes cost-sensitive users to free suites regardless of AI stance. Already visible in the March 2025 uptick.
Windows 10 end of support14 Oct 2025Fresh installs and OS migrations generate installer downloads; effect persists into 2026 as machines are replaced.
European digital-sovereignty push / Euro-Office2025–2026Public-sector and media attention on non-US office suites; TDF itself ran a sovereignty campaign in the weeks before 26.8.
26.8’s own features26 Aug 2026Paragraph Composer, OpenType variations, bidi/CJK improvements, multiple page sizes in Impress, Draft view, MS Office compatibility work. A typographically strong release attracts a different but real audience.
Media amplification of “no AI”26 Aug – 3 Sep 2026The Register, GamingOnLinux, OMG! Ubuntu, Hacker News. This is the article’s mechanism — but note it also amplified the release as a whole.
SeasonalityAug vs FebComparing an August launch to a February launch mixes back-to-school/new-hardware timing with the release. Unknown sign without a weekly series.

The one measurement that does isolate the “no AI” story

Social engagement on the launch announcement rose 4.4× (11k → 50k) while downloads rose 1.59×. The “no AI” line clearly worked as a headline — that is what a 4.4× engagement jump looks like. But if attention converted to installs proportionally, downloads should have moved far more than they did. The most defensible reading: the messaging bought a large amount of talk and a modest, un-isolated amount of downloads.

4. Attribution model — split the +383k yourself

The gap between the 25.2 and 26.8 launch weeks is 383,201 downloads. Assign what you think each non-AI driver plausibly contributed; whatever is left is the maximum the “no AI” stance could have caused. Defaults use the 2025 donation growth rate as the trend prior.

35%
40k
60k
trendfeatures / launch bumpexternalleft for “no AI”

5. Scoring the claim against the usual causal tests

Verdict: coincidental until shown otherwise. The record week is roughly what a project growing at 2025’s observed rate would post on a major release in late 2026. The “no AI” framing demonstrably generated attention — it is the only variable with a measured jump out of proportion to trend — but the download figure itself carries no signature that distinguishes it from the baseline plus a normal launch. The article promotes a plausible contributing factor to the sole cause without a counterfactual, a weekly series, or any attempt to net out the trend.

What would change this: TDF publishing first-week figures for 25.8 and 26.2, or weekly totals for Jan–Sep 2026. A flat 25.8/26.2 followed by a jump at 26.8 would be strong evidence for a release-specific cause; a straight line through all four would close the question.

Prepared 8 Sep 2026. Publication figures from The Document Foundation; contextual claims from Computerworld, OMG! Ubuntu, 9to5Linux and The Register. The attribution model is a reasoning aid, not an estimate.