LibreOffice and DuckDuckGo are now advertising the absence of AI, and people are noticing. Their marketing is timely as people reach AI saturation, but we’re not getting too excited yet.
DuckDuckGo still has AI features—it just doesn’t show them in “no AI” mode. And while LibreOffice’s downloads recently hit a milestone, they may largely be attributed to rising interest in alternatives to US software, not lack of AI.
This also doesn’t mean they stopped accepting AI-based contributions to their codebase, raising concerns over the long-term quality of the software.
No AI is becoming a selling point for developers too
Large open source projects, the foundational infrastructure of modern software, are now voting on whether to accept LLM-based contributions or not.
Many projects have formally forbidden AI entirely. Influential open-source compilers, code hosting platforms, language communities and development environments, as well as command-line tools have voted to ban AI-generated contributions or close their bug bounties as AI slop submissions increased. Some went as far as to say that one AI-suggested line among a hundred hand-written ones disqualifies a patch. The Zig language community lead, Loris Cro, calls their practice “contributor poker” — the idea that you’re not betting on a patch, but on a person who could become a long term contributor. AI breaks that bet, which is why it’s out.
The fiercest discussions took place in the Linux kernel, which powers thousands of enterprise applications and servers. Torvalds, its benevolent dictator, explicitly refused to make Linux anti-AI. In July 2026 he told objecting maintainers to fork the project or walk away if they don’t like it. “Linux is not one of those anti-AI projects.”
This followed months of turbulence. In March 2026, Google shipped an AI patch-reviewer, and some maintainers were outraged to find themselves reviewing the AI’s reviews instead of the code, or having to redo work they’d already checked by hand. By April 2026, a formal policy emerged for Linux 7.0: AI-assisted code is allowed, “AI slop” is banned, and AI tools can’t use the legally-binding “Signed-off-by” tag. There’s now an “Assisted-by” tag, and full liability for bugs and security flaws still rests with a human. Some contributors quit over this, and Copilot and Claude Code started landing real kernel fixes by May.
“No AI” is no longer fringe sentiment. Beyond ethical concerns, people have every right to believe that AI assistance can make the codebase worse, not better, because of how it quietly tricks the brains of the humans writing and reviewing it.
Cognitive surrender, or how using LLMs lowers the threshold for scrutiny
Before we started too Inconsistent, we also had to use LLMs but quickly put them down. We half-joked about writing a post called, “I stopped using LLMs to code because they were making me stupid.” We never wrote it, but many other developers have since come out with similar posts about how they’re second-guessing knowledge they spent years developing.
Researchers defined cognitive surrender as the wholesale offloading of thinking to LLMs without oversight or verification. This “uncritical abdication of reasoning” is common when an LLM’s output is “delivered fluently, confidently, or with minimal friction,” which is always. Across thousands of trials, people incorporated AI-generated outputs into their decision-making a whopping 70 percent of the time. We’re (unfortunately) wired to treat fluent, confident outputs as authoritative. We see something confident-sounding, think it must be true, and ignore signals that would otherwise make us stop and think.
We’ve seen this happen with hundreds of developers. As generating plausible-looking code became trivial, attention moved to code review. With bigger and more frequent change requests, people told themselves they’ll review everything carefully, just like we tell ourselves we’ll definitely go to the gym in January. Instead, everyone became too tired, too busy, and aware that if they said no too much, they would become the bottleneck, which is synonymous with “problem”.
The results were predictable. Microsoft teams in charge of email servers declared themselves overwhelmed and delayed product updates. Worse, Meta laid off people and “hired” agents instead, leading to a 40 percent increase in major technical and security incidents in just four months, while human employee time spent resolving those problems increased by as much as 70 percent. By July 2026, Mark Zuckerberg finally conceded that agentic AI didn’t “accelerate them as expected”.
What happens five years after the AI wrote it?
Previously, when reviewing a change request, it (mostly) meant that a human understood the domain and could explain what their code did. Nowadays nobody, not even the person who shipped it, might fully know what’s in the codebase or their changes. If the human reviewing an AI patch is also leaning on AI to explain it, the chain of understanding is broken further still. The moment you surrender your understanding of a codebase, you can no longer be its trusted, reliable steward.
Solutions like “signed off by” or “assisted by” don’t fix the problems we care about. A human can own a bug fix written by an LLM without ever having understood the code well enough. “If you don’t like it, walk away” also doesn’t provide reassurance about whether the codebase will be safer or more maintainable five years from now. Similarly, human “liability” for AI-assisted code means little when software licenses provide the code “as is” to users, meaning “use at your own risk”.
We work with people who want software that is correct, secure and performant. In our world, software is a long-term maintenance game, and a codebase no one really created or understands anymore is a sunk cost. That’s why we don’t use these technologies and still architect and code like we did before they came along. Ask us to change something to your system in five years, and we’ll know exactly where to look because we authored it.
Research also gave us the illusion of superiority bias, yet many people still convince themselves it’s fine, because they are not like everyone else. Just this one time, just this one more task. What they’re really saying is: five years on, it will be someone else’s problem.