Why AI-Heavy Projects Rarely Feature on LinuxLinks

I have reviewed free and open source software for more than two decades. In that time, plenty of development fads have come and gone. The latest is the flood of projects built largely, sometimes almost entirely, with generative AI.

I think AI tools are hugely beneficial to the development process. In capable hands, they can find bugs, explain unfamiliar code and deal with tedious work. But using AI as a tool is not the same as asking it to produce most of a program that its “developer” barely understands.

There is no reliable detector that tells me what percentage of a repository was generated. Some projects openly disclose they are AI-heavy. That’s appreciated. But many projects do not. I judge the software instead. I install it, try its features and look at how it is being developed.

The warning signs tend to appear quickly. Hundreds of trivial commits. Great slabs of repetitive code. Documentation that looks thorough until you try to follow it. Most telling of all is the huge feature list, much of which bears little relation to what the program can actually do.

GitHub statistics are increasingly worthless as a measure of maturity. Thousands of lines of code can now be generated over a weekend. Some repositories look exceptionally busy but contain little more than a thin copy of an established program, given a different interface and a longer list of supposed features.

Then, just as quickly, development stops.

That is hardly surprising. When a project required little effort to generate, walking away from it costs very little. The “developer” loses interest, starts another repository and leaves the old one to rot. Meanwhile, I may have spent hours installing the software, working around faults, taking screenshots and preparing an article. I have better uses for my time.

There are other problems. Generated code may contain poor error handling, unsafe dependencies or unexpected network activity. A “maintainer” who does not properly understand the code is poorly placed to find and fix those faults.

Licensing is not magically settled by adding an MIT or other OSI-approved licence. It may be unclear who owns generated code or whether parts were reproduced from somewhere else. In some jurisdictions, purely AI-generated material may not qualify for copyright protection at all. I’m not a legal expert, though.

None of this amounts to a complete ban on software developed with AI. If a program works well, does something worthwhile and is maintained by somebody who understands it, I will consider it.

But inclusion in a LinuxLinks roundup is a recommendation. I don’t intend padding roundups with AI-slop. If you spot an AI-heavy project already covered on the site, let me know. I will take another look and decide whether it still deserves its place.