# No AI is now a feature. The fine print is about attention, not models

URL: https://www.thedeepfeed.ai/posts/2026-10-14-no-ai-is-a-feature/
Category: Products
Published: 2026-10-14
Author: the-deep-feed
Tags: libreoffice, open-source, contribution-policy, debian, qgis, apple-intelligence, ai-slop, product-strategy
Kind: deep

> Between late August and mid-October 2026, LibreOffice marketed the absence of AI, QGIS and System76 banned AI-written contributions, Debian voted the other way, and users built tools to strip Apple Intelligence out, all arguing over one scarce input: human attention.

## TL;DR

- LibreOffice 26.8 shipped on Aug 26 stating it 'contains no generative AI features' and logged **1,031,162 first-week downloads**, a project record. The Foundation's public download logs show September 2026 at **4.72 million**, the highest month in data going back to mid-2018, but downloads were already **up 19% year on year** before the release.
- **QGIS** and **System76's COSMIC** both moved from soft disclosure rules to outright bans on AI-generated contributions between Sep 15 and Oct 7. Both cited the same cost: reviewer time, not model quality.
- **Debian** went the other way: 425 ballots, eight options, and a Condorcet win for 'Responsible Use of Generative AI', which beat None of the Above **281 to 126**. The outright ban drew 144 votes against 257 for None of the Above.
- Opting out became its own product category. **RemoveMacAI** passed 4,900 GitHub stars for stripping Apple Intelligence from macOS 27, while Mozilla's CEO reported that only **about 1%** of Firefox users turned all AI off.
- The common thread is scarcity of attention, not hostility to models: maintainers rationing review, users demanding a default they consent to, and readers who have learned to spot machine prose, which one September paper detects at **97.0 macro-F1** from structure alone.

On August 26, The Document Foundation released LibreOffice 26.8 with a sentence that would have read as a confession two years earlier and read as a selling point now: ["LibreOffice 26.8 contains no generative AI features."](https://blog.documentfoundation.org/blog/2026/08/26/libreoffice-26-8/) The release notes went on to promise that documents are not sent to remote services and that no part of the suite needs a network connection to work. Eight days later the Foundation doubled down with a blog post whose headline answered a reader's half-joking question in the affirmative: ["Yes, no AI is now a feature."](https://blog.documentfoundation.org/blog/2026/09/03/yes-no-ai-is-now-a-feature/)

The same week, Simon Willison opened the cache folder of OpenAI's ChatGPT/Codex desktop app and found something unexpected sitting next to a full Python and a full Node.js install: a headless copy of LibreOffice, about [429.7 MB of it, inside a 1.7 GB runtime](https://simonwillison.net/2026/Sep/1/codex-libreoffice/). The office suite that had just declared itself AI-free was also bundled inside one of the largest AI products in the world, apparently to handle office documents.

Between late August and mid-October 2026, "no AI" stopped being the absence of a feature and became a position that organizations wrote down, voted on, and marketed. Two open-source projects banned AI-written contributions. One of the largest volunteer projects in software voted not to. A free utility for deleting Apple's on-device models collected thousands of stars. And essays about machine prose kept reaching the top of Hacker News.

The obvious reading is that "no AI" became a product feature and a governance position at the same moment. The evidence supports that, with a correction: read the policy texts, ballots and download logs closely and very few of them reject the technology. What they ration is attention: maintainers' review hours, users' consent to defaults, and readers' trust in prose. The fights that look ideological on the timeline are mostly arguments over who pays for those three things.

# The record and what the logs show

The download headline came from Brazil. On September 8, Rodrigo Ghedin's Manual do Usuário blog ran ["LibreOffice breaks download records after declaring it has no AI features"](https://manualdousuario.net/en/libreoffice-download-record-no-ai/), which reached 722 points on [Hacker News](https://news.ycombinator.com/item?id=49610538). The underlying number is the Foundation's own. Its stats post reported [1,031,162 downloads in the first week of 26.8](https://blog.documentfoundation.org/blog/2026/09/02/one-week-of-libreoffice-26-8-the-stats/) from the official download page, and called it "the highest number of downloads in the first week of any LibreOffice release." The count excludes Linux users who install through distribution repositories, Flathub or Snap.

The Foundation itself never joined the two facts. Its stats post does not mention AI, and TechCrunch's [October 6 write-up](https://techcrunch.com/2026/10/06/libreoffice-says-no-ai-is-now-a-software-feature/) quotes no causal claim. The link was the blogger's inference, and the top reply on Hacker News challenged it: user smokel wrote that "The number of weekly downloads has simply been increasing over time, due to growing popularity," and that "Calling each new high 'record-breaking' is a bit meaningless." Another user, Aurornis, pointed readers to the Foundation's public [download dashboard](https://stats.documentfoundation.org/downloads) and argued the monthly view looked "mostly flat going back to 2019."

That dashboard, built from download-page logs exported to Matomo, covers June 2018 through September 2026 and is the best public evidence available. Tallied by month, it says the skeptics and the headline are each half right.

| Period | Average downloads per month |
|---|---:|
| 2019 | 2.96 million |
| 2021 | 2.77 million |
| 2023 | 2.85 million |
| 2025 | 3.09 million |
| Jan to Aug 2026 | 3.56 million |
| Sep 2026 | 4.72 million |

From 2019 through 2025 the series is close to flat, as Aurornis said: monthly averages stayed between 2.6 and 3.1 million for seven years. But 2026 broke out of that band well before 26.8 shipped. January through August 2026 totalled 28.45 million downloads against 23.82 million in the same months of 2025, a 19.4% rise. The year before, the same comparison grew 3.4%. Then September 2026 logged 4,721,384 downloads, 21% above August and 37% above September 2025. Before 2026, the dashboard's best month was March 2019, at 3.84 million; March and August 2026 had both passed it.

Release months always bump: September ran 16% above August in 2025 and 32% above it in 2024. What is not ordinary is the level. In the seven weeks from August 24, weekly downloads averaged 1.12 million, against an average of 811,000 in the 2026 weeks before the release. The comparable post-release lift in 2025 was from 686,000 to 799,000. The single largest week in the dashboard's 100-week view is October 5 to 11, at 1,337,034, the week TechCrunch picked up the "no AI" story. Windows installers account for more than four in five downloads in every recent week.

![LibreOffice monthly downloads: flat 2.6M to 3.1M band for 2019 to 2025, a 19% rise in early 2026, and a red 4.72M September 2026 peak.](/post-images/2026-10-14-no-ai-is-a-feature/download-ledger.jpg)

What the data supports is narrower than the headline and more interesting. LibreOffice demand was already rising through 2026, at roughly six times the prior year's growth rate. The no-AI release then coincided with the largest month and the largest week on record. The dashboard cannot separate the AI positioning from the release's other features (the Foundation led with typography and writing-system support). One commenter suggested the Codex bundling might inflate counts; nothing in Simon Willison's post says the bundled copy comes from the Foundation's download page, so that remains speculation. A defensible summary: a product that advertised the absence of AI logged its best month and best week on public record, and nobody, including the vendor, can yet show the advertisement caused them.

# A SKU before it was a slogan

The Foundation's September 3 post is more careful than its headline. Signed by Italo Vignoli, it states that LibreOffice "does not reject artificial intelligence out of hand," and then lists six conditions any AI feature would have to meet before entering the default install: user-controlled execution, no content leaving the computer without authorization, no telemetry of any kind, no dependence on a single provider, no compromise on file format, and entirely optional. Until a feature meets all six, the post concludes, "there will be no AI of any kind in the default installation." Users who want AI can add it through extensions.

Read as a specification, the list says nothing about model quality, hallucination or copyright. Every condition is about control: where inference runs, who triggers transmission, which vendor the user is tied to. The post is explicit about the commercial logic it opposes. "For a company that sells subscriptions, an AI assistant justifies a price increase," it says, while the Foundation is a nonprofit with nothing to upsell.

The incumbent it does not name had already priced that argument. In January 2025 Microsoft folded Copilot into its consumer Office subscriptions and raised prices, [Microsoft 365 Personal from $7 to $10 a month and Family from $10 to $13](https://www.pcworld.com/article/2581179/microsoft-follows-google-with-price-bump-forced-ai-365-bundles.html). Existing subscribers who did not want the AI could switch, for a limited time and with some effort, to "Classic" plans at the old prices. In effect, Microsoft sold "no AI" as a cheaper tier 20 months before LibreOffice marketed it as a feature. The difference is which side of the default each company stands on.

The most revealing data point about consumer demand for an off switch came from a vendor that built one. On August 25, Mozilla CEO Anthony Enzor-DeMeo published an essay titled ["Trust is earned. Consent is asked. AI skipped both."](https://aed.me/articles/turn-it-off/) Firefox had shipped AI Controls, a setting that blocks generative features, after a public backlash to its AI plans. His report on uptake: about 1% of users turned AI off entirely and 3% disabled individual features. "Almost nobody used it," he wrote, and then drew the opposite conclusion from the one a growth team would: "The most important AI feature may be the ability to turn it off."

That frames the demand question better than any download chart. A small minority acts on the switch; a much larger group, judging by the backlash Mozilla describes, wants to know it exists. A week later, blogger Marius Quabeck [retested the switch on Firefox 155](https://marius.blog/firefox-155-ai-kill-switch-retest/) and found it left telemetry untouched, with 14 telemetry requests in 120 idle seconds whether it was on or off. His verdict: "The switch rotates nothing and deletes nothing." The demand is for verifiable control, and a switch that does less than its name suggests does not satisfy it.

![Four layers of opting out: vendor switch, paid Classic tier, third-party removal tool, and a red AI-free default layer.](/post-images/2026-10-14-no-ai-is-a-feature/opt-out-stack.jpg)

# Two projects close the door

On the contribution side, the turn happened in public mailing-list threads and Git commits, and the documents are unusually legible.

QGIS, the open-source desktop mapping application, wrote its first AI policy in a document dated January 31, 2026. Authored by core developer Even Rouault, it was permissive with conditions: contributors could use any tool, but "there must be a human in the loop," they had to review every generated line before asking for review, and they should label substantial tool-generated content with a trailer such as `Assisted-by:`. It also contained a line that anticipated the next eight months: "Passing maintainer feedback to an LLM doesn't help anyone grow, and does not sustain our community."

On September 15, core contributor Nyall Dawson [wrote to the developer list](https://lists.osgeo.org/pipermail/qgis-developer/2026-September/068547.html) proposing two measures. The first was "an immediate, total moratorium on all forms of LLM based contributions." His reasoning was about the failure of the softer rule, not about the models:

> We've tried gentle policies regarding this up till now, but clearly that's not helping.
>
> — [Nyall Dawson, QGIS-Developer list](https://lists.osgeo.org/pipermail/qgis-developer/2026-September/068547.html), Sep 15, 2026

He framed the ban as a reversible default, writing that the project should "RIGHT NOW" default to a safe "NO AI" policy and potentially loosen it later. His second proposal is the more striking document of the period. Because "any policy we introduce will just be totally ignored or lied about by bots or karma farming AI sloppers," he suggested new contributors make a small financial donation to the project before their first contribution is accepted, as a proof of life. His one exemption, using a model as a substitute for simple web search, carried a footnote: "because the web is broken in 2026."

The proof-of-life idea drew objections on equity grounds. On September 23, Tim Sutton, a member of the QGIS Project Steering Committee, [replied](https://lists.osgeo.org/pipermail/qgis-developer/2026-September/068576.html) with the wording that became policy, and advised that it carry no rationale at all. He signed off with a line that is now a small genre of its own: "(No LLM was used to compose this email)."

Dawson opened QEP 428 as [pull request #387](https://github.com/qgis/QGIS-Enhancement-Proposals/pull/387) on September 24. It moved to a vote on October 5 and merged on October 7. The opening post drew 109 thumbs-up and 34 thumbs-down reactions. The policy file, now version 2.0, replaced several pages of guidance with two sentences:

> No LLM, AI or agent assisted code contributions are allowed in the https://github.com/qgis/qgis repository. If we subsequently discover that you have used an AI, LLM, or any agent-based workflow for your contribution, you will be subject to code reversion and contributor rights sanctioning.
>
> — [QGIS AI tool use policy, v2.0](https://github.com/qgis/QGIS-Enhancement-Proposals/blob/master/qep-408-ai-tool-policy.md), merged Oct 7, 2026

System76 took the same path on a shorter clock. The pull request template for COSMIC, the Rust desktop environment behind Pop!_OS, gained an AI clause on February 13 that asked contributors to disclose AI-generated code and warned that, in the team's experience, such code "often results in overly complex code that lacks enough context for a proper fix or feature inclusion," producing "considerably longer code reviews." On September 30, principal engineer Jeremy Soller merged [PR #3911](https://github.com/pop-os/cosmic-epoch/pull/3911), titled "Disallow LLM generation in pull requests." The checklist item now reads: "I have not included any LLM (also known as AI) generated content in this PR, including code, comments, and descriptions." The [Pop!_OS contributing guide](https://github.com/pop-os/pop/blob/master/CONTRIBUTING.md) gained the same rule for issues as well as pull requests. A follow-up commit the next day removed a list of acceptable uses.

Soller's explanation, posted to Reddit and quoted by [GamingOnLinux](https://gamingonlinux.com/2026/10/system76-cosmic-projects-will-no-longer-accept-llm-generated-content-in-code-submissions), named the constraint plainly: "The primary reason is simply our team's load for reviews." The team wanted to prioritize its own staff and regular contributors, he wrote, while it had been "receiving far more changes from first-time contributors using LLMs. These changes are often unplanned and have had a low acceptance rate." One repository, cosmic-flatpak, is exempt, because its manifests point at other projects' code and System76 reviews them only for sandboxing.

![Parallel timelines: QGIS from human-in-the-loop (Jan 31) to ban (Oct 7), COSMIC from disclosure (Feb 13) to ban (Sep 30).](/post-images/2026-10-14-no-ai-is-a-feature/two-rulebooks.jpg)

# The arithmetic of review

The two bans share a structure. Neither project claims AI-written code is worse in principle, and neither has an enforcement mechanism beyond trust. When a reviewer on the QGIS pull request asked how the ban would be enforced, the answer in the thread was, in effect, by trust. Sutton's own draft wording conceded "We know this is difficult to enforce." What the bans change is the default burden of proof. Before, a maintainer had to judge each AI-assisted patch on its merits. After, the contributor certifies the patch is human-written, and a maintainer who suspects otherwise can close it without a technical argument.

That is a policy for rationing a scarce input. On Hacker News, where the QGIS story reached [392 points and 410 comments](https://news.ycombinator.com/item?id=50054751) on October 12, a commenter who said he works on Plotly described the same pressure from inside another project: "3 page issue writeups nobody has proofread, nobody but a bot to follow up with when we try, huge volume of plausible but weak PRs eating up reviewer time." Another commenter read the trend more darkly: "I think this is the end of open-contributions."

The costs land on real contributors too. One developer on the COSMIC thread reported that his pull request fixing VPN password handling was closed under the new rule. He had used Claude to find the bug and draft the fix, then spent "a lot time handcrafting and making sure the quality was good," and took the decision with "no hard feelings," while calling it a setback. A geospatial practitioner on X, posting the day after the QGIS story trended, said he disliked the policy and had stopped using the program anyway:

> QGIS is enforcing a no AI code policy, at all, I’m not a fan
>
> Sounds to me like an open invite for someone to make a fork and make what Omarchy is doing with Linux, completely embarrassing LLMs and running with it
>
> — [@MaxLenormand](https://x.com/MaxLenormand/status/2110052134351765797), Oct 13, 2026

Other critics went after the drafting. One argued QGIS had done "very little to define what that means," comparing the ambiguity to arguments over what counts as a derivative work under the GPL. Reviewers on the pull request itself asked whether autocomplete, translation, search and security analysis were covered; the merged text does not say. A commenter on the COSMIC story called the ban "Completely performative," given that any Linux desktop ships thousands of upstream packages that will contain AI-assisted code.

The bans have precedents that are worth keeping in view, because they show which arguments are new. The Gentoo Council [voted on April 14, 2024](https://wiki.gentoo.org/wiki/Project:Council/AI_policy) to forbid contributions "created with the assistance of Natural Language Processing artificial intelligence tools," citing copyright, quality and ethics. NetBSD's [commit guidelines](https://www.netbsd.org/developers/commit-guidelines.html) treat code from a large language model as "presumed to be tainted code" that needs written approval from core. Those 2024 rules led with copyright exposure. The 2026 rules lead with reviewer load. The argument has migrated from the legal department to the maintainer's inbox.

# Debian votes the other way

The counter-move came from the project best placed to test whether a ban could win a formal vote. Debian's general resolution on LLM use ran a discussion period from July 23 and a voting period from August 15 to August 28, with eight options on the ballot plus None of the Above. Debian uses a Condorcet method, so every option is compared head to head with every other.

The [published tally](https://www.debian.org/vote/2026/vote_002) contains 425 ballots from a developer body of 1,045. The winner was Option 5, "Responsible Use of Generative AI," proposed as a position statement rather than a rule:

> Debian neither endorses nor prohibits the use of generative AI tools in the development, maintenance, or documentation of software, packaging, documentation, and other media published within the Debian Project.
>
> — [Debian GR 2026, Option 5](https://www.debian.org/vote/2026/vote_002), adopted Aug 2026

The text adds that "The use of a generative AI tool does not diminish the contributor's responsibility for the work they submit," encourages but does not require disclosure, and tells contributors not to send embargoed security information or credentials to third-party AI services.

The ballot is more informative than the winner. Option 5 beat None of the Above 281 to 126 and beat every other option pairwise, including the more prescriptive "Allow AI-Assisted Contributions with conditions" by 203 to 148, and "A cautious approach to generative AI" by 210 to 130. The outright ban, Option 1, would have amended the Social Contract and so needed a 3:1 supermajority; it drew 144 votes over None of the Above against 257 the other way. Option 3, "Reject LLMs as far as practical," also failed against None of the Above, 176 to 230. Two restrictive statements, "Debian is created by humans" and "Avoid the use of LLM: climate destruction is a deal breaker," did clear None of the Above, by 213 to 192 and 225 to 176, before losing to the winner.

![Ballot schematic of Debian's nine options as paired bars versus None of the Above; Option 5 highlighted red with 281 to 126.](/post-images/2026-10-14-no-ai-is-a-feature/debian-ballot.jpg)

The losing texts are the sharpest statements of the no-AI position in this period. Option 1's rationale argued that "Debian is not here to generate as much code as possible requiring manual review by a shrinking number of human volunteers," the same arithmetic QGIS and System76 would cite a month later. It closed by noting it was written "organically and without language model assistance." Option 3 opened with a list of grievances and the line "Ethical and safe use of this technology is almost impossible."

On Hacker News, where the LWN report drew [517 points](https://lwn.net/Articles/1091231/), the top comment summarized the result as "AI or not, it's still your code and you're responsible for it," and added, "I can get on board with that." Joey Hess, a former Debian developer, posted a more pessimistic reading on the day of the result, [arguing that](https://joeyh.name/blog/entry/Debian_and_the_sirens/) "What LLMs in Debian development will do, I fear, is eliminate any incentive to scrap boilerplate" that the project should have reformed long ago.

The Debian result does not contradict the QGIS and COSMIC bans so much as reveal the variable that separates them. Debian is large, federated, and organized around individual maintainers who own their packages; a responsibility rule pushes the cost to the person who submits. QGIS and COSMIC are smaller projects with central review queues; a responsibility rule still leaves a small team reading every patch. Scale and review topology, more than philosophy, predict which rulebook a project writes. The Register [reported on October 13](https://www.theregister.com/os-platforms/2026/10/13/netbsd-may-have-found-a-new-role-the-ai-free-foss-os/5302818) that NetBSD drew a record 120 donations in September and cast its 2024 rule as a niche: "the AI-free FOSS OS." The niche is real, and by construction small.

# Uninstalling the assistant

On the consumer side, the clearest demand signal of the period was a repository. [RemoveMacAI](https://github.com/omlahore/RemoveMacAI), by Om Lahore, describes itself in its first line as a way to "Debloat macOS. Turn off Apple Intelligence and delete its models." It supports only macOS 27 on Apple silicon, keeps System Integrity Protection on, and records every change in an undo history. Its README states that it "makes no network requests and collects no data." It also disables analytics, personalized ads and assorted annoyances, so not every star is a vote against AI specifically. When it reached the Hacker News front page on October 4 it drew 775 points and 559 comments; at the time of writing it had about 4,900 stars.

The reaction split along a line that recurs throughout this story. The top reply compared it to "all the de-crufting you've always needed to do when you install Windows." A widely upvoted dissent asked why anyone would remove "well-balanced and relatively small local-inference models that ship with every mac," which are, after all, the privacy-preserving version of AI that LibreOffice's six principles describe. Much of the thread's support was framed around control rather than inference: being asked first, and getting the disk space back.

Apple's own documentation became part of the story. A support page titled ["Turn off and restrict access to Apple Intelligence features on Mac"](https://support.apple.com/guide/mac-help/turn-restrict-access-apple-intelligence-mchlb2e44f94/mac) circulated on Hacker News on September 21, the same day an Ask HN thread asked whether Siri could be disabled on macOS 27 at all. The page walks through turning off Siri AI, which macOS 27 labels as a beta, and the summary features in Messages, Mail and notifications one setting at a time. It does not say whether turning features off deletes the on-device models or returns their storage, which is the gap RemoveMacAI fills. A week earlier, a "Tell HN" post had claimed iOS 27 offered no way to disable Apple Intelligence; replies in the [same thread](https://news.ycombinator.com/item?id=49704226) found partial workarounds.

The pattern matches Firefox's: the vendor ships the feature on by default, provides a switch for the visible surface, and leaves users unsure what remains. The third-party tool sells certainty. Read that way, RemoveMacAI's star count is a measure of distrust in the vendor's switch more than of hostility to the models.

# The trained eye

The third front was prose, and it moved on Hacker News more than anywhere else. The window opened on August 21 with [Claudette](https://github.com/adnanakil/nobuzz), a tool that pipes Claude's replies through a second model to strip what its README calls "a TED talk about its own pull request," including inflated framing such as "load-bearing assumption," numbered revelations and narrative kickers. It reached 364 points. Six days later, Louis Abraham's ["The load-bearing vocabulary of Claude"](https://louisabraham.github.io/load-bearing/) reached 710 points by charting word clusters in GitHub pull-request text, updated daily across more than 500,000 documents; the page's lead cluster, headed by words such as "nobody," "quietly" and "load-bearing," went from under 1% of the tracked share in early 2025 to about a third by October.

Bryan Cantrill's December 2025 essay ["Your intellectual fly is open"](https://bcantrill.dtrace.org/2025/12/05/your-intellectual-fly-is-open/) resurfaced on September 6 and reached 729 points. Its central image is that readers detect machine authorship and say nothing: "your intellectual fly is open: lots of people notice — but no one is pointing it out." He calls LLMs "astonishingly good editors"; his objection is to authorship: "LLMs are also lousy writers," and above all "they are not you." Erich Grunewald's [essay](https://erichgrunewald.substack.com/p/why-you-should-almost-never-use-ai) followed on September 19 at 363 points, arguing that "the writing process is an essential part of the thinking process" and that unlabeled AI writing "is rude and misleading," with exceptions for translation, routine email and editing.

On September 27, "Tells of a Slop UI," a catalogue of the habits that give away generated web interfaces, reached 386 points on [Hacker News](https://news.ycombinator.com/item?id=49867038). (The original post sits behind a bot-check page.) The largest thread of the period, though, cut the other way. A June post titled ["AI-generated posters don't have to be horrible"](https://john.hartnup.uk/2026/06/07/ai-event-posters.html) resurfaced on September 19 and drew 1,902 points and 962 comments. Its author argued that the problem with generated event posters "is not so much that they're bad," but that "once you've seen that style 20 times it starts to irritate just from the sheer repetition," then prompted for Bauhaus, Swiss and risograph styles instead.

The comments split. One heavily replied-to comment argued that "the average freelance designer online is significantly worse than AI." Another answered that "the default style signals low effort," and that what grates is "low effort trying to present as high effort." Both can be true at once. The second describes a reader's inference about the sender, not a judgment of the artifact, which is the same inference Cantrill describes and the same one behind Tim Sutton's sign-off.

That inference is becoming measurable. A September paper, [SlopShape](https://arxiv.org/abs/2609.15369), compared 2,250 pre-ChatGPT human blog posts from 268 company domains with 11,250 posts from five frontier models. Using 176 structural features alone, with no word-level cues, it separated AI posts from human ones at 97.0 macro-F1 on held-out companies, and still scored 96.1 after each AI post was reworded by its own model. It attributed posts to the right model 68.6% of the time against a 16.7% chance rate. The author describes AI posts as having a "tidy, self-announcing shape," while human posts fall into rarer configurations. Readers who say they can tell are, on this evidence, responding to something real.

![Detection schematic: human and generated posts feed a 176-feature structure grid; output gauge marked 97.0 macro-F1 with red needle.](/post-images/2026-10-14-no-ai-is-a-feature/structure-detector.jpg)

The fatigue reached the venue itself. On September 11, an Ask HN post titled ["Can we please limit the AI news flood?"](https://news.ycombinator.com/item?id=49657850) reached 868 points. Its author wrote that the feed had become "almost exclusively AI or AI-adjacent news," and that "I no longer find HN a good source of 'what's new' in the industry." The top reply did not dispute it: "HN is a reflection of the industry and we're at peak hype-cycle at the moment." Others said they came to the site precisely for model news, a fair sample of the period's mood.

# Where the line sits

Set side by side, the documents from these eight weeks draw a consistent line, and it is not the line between pro-AI and anti-AI.

LibreOffice's six principles reject telemetry, background transmission and vendor lock-in, and leave the door open to local, optional, user-triggered inference. Mozilla's CEO concluded that the switch matters more than its usage. RemoveMacAI's pitch is deleting models users were never asked about. QGIS and COSMIC banned AI-written contributions because reviewer hours are finite and AI-assisted patches from strangers arrived faster than those hours could absorb them. Debian, whose review costs are spread across individual maintainers, chose to make contributors answerable rather than tools forbidden. The essays on prose objected to unlabeled authorship, and the strongest piece of evidence in that debate is a classifier showing that the label is often unnecessary.

Each of those is a claim about who bears a cost: the maintainer reviewing a patch the submitter did not understand, the user living with a default that sends data and uses disk without asking, the reader working out whether anyone thought through the text in front of them. "No AI" is the label organizations reached for when they decided to stop absorbing one of those costs on someone else's behalf.

The label will likely prove less durable than the costs underneath it. Dawson framed the QGIS ban as a default to loosen later, and the pull request already hosts proposed exceptions for autocomplete and translation. The Foundation's six conditions read as a specification a vendor could meet. Debian's statement says its position may evolve without another vote. The organizations that wrote "no AI" into policy this autumn mostly wrote terms, not verdicts. The test for the next year is whether the companies shipping AI by default find those terms cheaper to meet than to argue with, the same dynamic The Deep Feed traced in [mandatory AI labeling under Article 50](/posts/2026-08-02-the-watermark-era-begins/): provenance becomes an obligation once enough people decide they are owed it.

## Sources

- [The Document Foundation — LibreOffice 26.8 release announcement (Aug 26, 2026)](https://blog.documentfoundation.org/blog/2026/08/26/libreoffice-26-8/)
- [The Document Foundation — One week of LibreOffice 26.8: the stats (Sep 2, 2026)](https://blog.documentfoundation.org/blog/2026/09/02/one-week-of-libreoffice-26-8-the-stats/)
- [The Document Foundation — Yes, no AI is now a feature (Sep 3, 2026)](https://blog.documentfoundation.org/blog/2026/09/03/yes-no-ai-is-now-a-feature/)
- [The Document Foundation — LibreOffice Download Metrics dashboard (download logs via Matomo)](https://stats.documentfoundation.org/downloads)
- [Manual do Usuário — LibreOffice breaks download records after declaring it has no AI features (Sep 8, 2026)](https://manualdousuario.net/en/libreoffice-download-record-no-ai/)
- [QGIS — QEP 428 pull request #387: No AI contributions permitted (merged Oct 7, 2026)](https://github.com/qgis/QGIS-Enhancement-Proposals/pull/387)
- [QGIS — AI tool use policy, qep-408-ai-tool-policy.md (v2.0)](https://github.com/qgis/QGIS-Enhancement-Proposals/blob/master/qep-408-ai-tool-policy.md)
- [QGIS-Developer list — Nyall Dawson, Two proposals: moratorium on LLM use (Sep 15, 2026)](https://lists.osgeo.org/pipermail/qgis-developer/2026-September/068547.html)
- [QGIS-Developer list — Tim Sutton, PSC reply (Sep 23, 2026)](https://lists.osgeo.org/pipermail/qgis-developer/2026-September/068576.html)
- [System76 — cosmic-epoch PR #3911: Disallow LLM generation in pull requests (Sep 30, 2026)](https://github.com/pop-os/cosmic-epoch/pull/3911)
- [System76 — cosmic-epoch pull request template](https://github.com/pop-os/cosmic-epoch/blob/master/.github/PULL_REQUEST_TEMPLATE.md)
- [System76 — Pop!_OS CONTRIBUTING.md, AI Contributions](https://github.com/pop-os/pop/blob/master/CONTRIBUTING.md)
- [GamingOnLinux — System76 COSMIC projects will no longer accept LLM-generated content (Oct 5, 2026)](https://gamingonlinux.com/2026/10/system76-cosmic-projects-will-no-longer-accept-llm-generated-content-in-code-submissions)
- [Debian — General Resolution: LLM usage in Debian (vote page, ballot texts and tally)](https://www.debian.org/vote/2026/vote_002)
- [LWN — Debian votes to allow 'responsible use of generative AI' (Aug 29, 2026)](https://lwn.net/Articles/1091231/)
- [Joey Hess — Debian and the sirens (Aug 29, 2026)](https://joeyh.name/blog/entry/Debian_and_the_sirens/)
- [Gentoo Council — AI policy (voted Apr 14, 2024)](https://wiki.gentoo.org/wiki/Project:Council/AI_policy)
- [NetBSD — Commit guidelines (LLM-generated code presumed tainted)](https://www.netbsd.org/developers/commit-guidelines.html)
- [The Register — NetBSD may have found a new role: the AI-free FOSS OS (Oct 13, 2026)](https://www.theregister.com/os-platforms/2026/10/13/netbsd-may-have-found-a-new-role-the-ai-free-foss-os/5302818)
- [GitHub — omlahore/RemoveMacAI](https://github.com/omlahore/RemoveMacAI)
- [Apple Support — Turn off and restrict access to Apple Intelligence features on Mac](https://support.apple.com/guide/mac-help/turn-restrict-access-apple-intelligence-mchlb2e44f94/mac)
- [Anthony Enzor-DeMeo — Trust is earned. Consent is asked. AI skipped both. (Aug 25, 2026)](https://aed.me/articles/turn-it-off/)
- [Marius Quabeck — Firefox's AI switch is off. Telemetry isn't (Sep 2, 2026)](https://marius.blog/firefox-155-ai-kill-switch-retest/)
- [PCWorld — Microsoft follows Google with price bump, forced AI 365 bundles (Jan 17, 2025)](https://www.pcworld.com/article/2581179/microsoft-follows-google-with-price-bump-forced-ai-365-bundles.html)
- [Simon Willison — The ChatGPT/Codex app bundles a full copy of LibreOffice (Sep 1, 2026)](https://simonwillison.net/2026/Sep/1/codex-libreoffice/)
- [TechCrunch — LibreOffice says 'no AI' is now a software feature (Oct 6, 2026)](https://techcrunch.com/2026/10/06/libreoffice-says-no-ai-is-now-a-software-feature/)
- [Bryan Cantrill — Your intellectual fly is open (Dec 5, 2025)](https://bcantrill.dtrace.org/2025/12/05/your-intellectual-fly-is-open/)
- [Erich Grunewald — Why you should almost never use AI to write anything substantive (Aug 6, 2026)](https://erichgrunewald.substack.com/p/why-you-should-almost-never-use-ai)
- [JH — AI-generated posters don't have to be horrible (Jun 7, 2026)](https://john.hartnup.uk/2026/06/07/ai-event-posters.html)
- [Madler — SlopShape: Identifying AI-Generated Commercial Web Content (arXiv 2609.15369)](https://arxiv.org/abs/2609.15369)
- [Louis Abraham — The load-bearing vocabulary of Claude](https://louisabraham.github.io/load-bearing/)
- [GitHub — adnanakil/nobuzz (Claudette)](https://github.com/adnanakil/nobuzz)
- [Hacker News — Ask HN: Can we please limit the AI news flood? (Sep 11, 2026)](https://news.ycombinator.com/item?id=49657850)
- [X — @MaxLenormand on the QGIS policy (Oct 13, 2026)](https://x.com/MaxLenormand/status/2110052134351765797)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-10-14-no-ai-is-a-feature/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt