Linus Torvalds Just Handed Every Business Owner a Working AI Policy. Two Days Later, Someone Showed You the Catch.

Last week the creator of Linux settled the "should we use AI or not" argument in about one sentence. This week, someone accidentally demonstrated exactly why that sentence isn't the whole story. Together, the two moments are the clearest guide I've seen to how a normal business should actually think about AI, not from a vendor trying to sell you something, but from the most famously blunt pragmatist in software.

Here's the short version. Adopt AI the way Torvalds says. Then respect the thing the second story proves: the human judgment sitting on top of the AI is the actual product. Remove it and you get something that makes headlines and does nothing.

What Torvalds actually said

Linus Torvalds runs the Linux kernel, the software underneath most of the internet, your phone, and probably your business software's hosting. He is not a hype man. In late 2024 he dismissed most of the AI industry as marketing noise and said he'd simply ignore it.

Last week he changed his tune, and the way he did it is the useful part. Responding to developers who wanted the kernel project to take an anti-AI stance, he said the project isn't in that camp, and "if somebody has issues with that they can do the open-source thing and fork it." AI, he said, is a tool like any other, and a clearly useful one now, whatever you thought a year ago.

But he didn't pretend it's painless. He openly conceded that AI creates real work for maintainers, adds noise, and keeps surfacing embarrassing bugs. His point was that the answer to that isn't to stick your head in the sand and pretend the tool doesn't exist. It's to make the tool help your people rather than hurt them. He also refused to hold AI to a standard of perfection humans don't meet either, since human experts make plenty of mistakes too.

Strip away the kernel-developer context and that's a complete AI policy for a business. We wrote separately about how Torvalds frames AI as a tool rather than an author, which sits alongside this well.

Why this is the policy your firm needs

Most businesses I talk to are stuck between two bad positions. One camp wants to ban AI outright because a junior once pasted something confidential into a chatbot. The other wants to mandate it everywhere because a competitor put "AI-powered" on their website. Both are the head-in-the-sand move in different costumes. One pretends the tool doesn't exist, the other pretends the tool is magic.

Torvalds' stance is the sane middle, and it translates cleanly to any small business. AI isn't banned and isn't mandatory. Nobody's forced to use it, and nobody gets to stop a colleague from using it on ideological grounds. You judge it on one thing: does it produce a better result than not using it? Not whether it's exciting, not whether it's principled. Whether it works. That's a policy you can write on one page and actually enforce.

Then someone said "hold my beer"

Here's where it gets instructive. Two days after Torvalds told critics to go fork the kernel if they didn't like his AI stance, a project popped up that looked like exactly that: a reimplementation of Linux written in Rust, the fashionable systems language. The tech press ran the irresistible headline. Someone forked Linux to spite Linus.

Except almost none of that was true. It isn't a fork. It isn't a response to Torvalds at all. It's a from-scratch rewrite of Linux version 0.11, a prototype kernel from 1991, and it appears to have been generated largely by AI. The consensus among actual kernel people was blunt: nobody is ever going to run anything real on a bot-generated rewrite of a 35-year-old prototype. Impressive-sounding headline. Hollow underneath.

And that is the catch. That project is a perfect specimen of the exact failure mode Torvalds warned about in the same breath he endorsed AI. It's noise that looks like signal. Output that clears the bar of "looks like an achievement" while completely missing the bar of "is a thing you could use."

The part everyone skips: the maintainer

Reread what Torvalds said AI is good for on the kernel. It keeps finding embarrassing bugs. That's the tell. AI is valuable there precisely because a competent human maintainer sits between the AI's output and the codebase, catching what's useful and throwing out what isn't. The AI proposes. The human decides. Take the human out and you don't get the Linux kernel; you get the Rust stunt.

For a small business this matters more than it does for coders, because your dangerous failure mode isn't output that's obviously broken. It's output that's confident, plausible, well-formatted, and wrong. An AI that mangles a reconciliation in a way that looks clean is far more dangerous than one that visibly falls over. The demo is never the deliverable. The review step is where the value actually lives. Getting the processes underneath right before you automate is part of what makes that review step actually work.

What I'd do this quarter

Not a checklist. Four moves.

Write your one-paragraph AI stance. Not banned, not mandated, judged on results. Circulate it. It ends more arguments than a 20-page policy.

Decide what AI drafts versus what AI decides. The line is simple: AI drafts, extracts, and summarises; humans decide and sign. Anything numeric or client-facing gets a named human reviewer before it leaves the building.

Build the review step on purpose. If your AI workflow has no explicit human checkpoint, you've automated the production of confident-looking mistakes. Add the checkpoint before you scale the volume.

Never mistake a demo for a deliverable. The Rust-kernel headline is what an impressive AI demo looks like. Ask of any AI output the same question the kernel developers asked: would I actually run my business on this?

Where OrionX comes in

The pragmatic stance is the easy half. The hard half is building the workflows so AI genuinely helps your team instead of quietly generating output that someone has to catch later. That's the work we do: designing AI into a business with the human judgment step built in, not bolted on, so you get the productivity without inheriting the failure mode.

If your business is stuck between banning AI and mandating it, or you've got tools producing output nobody's actually checking, get in touch with OrionX and we'll help you find the sane middle.