Why I don't use generative LLMs, and nor should you

Created: 2025-11-29

At the moment of writing, I don't have as many sources for these as I would like, but I may add them later. (Feel free to send me some yourself.)

Here is my roughly-written list of reasons why I don't use generative LLMs, and nor should you -

  1. Their excessive use of energy and water, both for training and use.

  2. Their blatant violation of copyright (including freedom-respecting licenses, both permissive and copyleft).

    Whatever the courts end up deciding, my stance is set - to me, it will always be a violation.

  3. The way they DDoS servers in the process of scraping the web, in turn causing the servers to resort to privacy-invasive and user-hostile techniques like CAPTCHAs.
  4. Their furthering the destruction of privacy.

  5. Their destruction of their sources.

    I've read numerous times that LLM output is not suitable for training LLMs. In other words, they rely on human output.

    What text, images, and videos will you train on, when all the sources have been replaced by LLMs?

    What (say) StackOverflow answers will you train on, when nobody asks nor answers questions on StackOverflow anymore?

  6. Finally, I am an activist for anti-monopolist software (commonly called free/libre/open software). Most LLMs in their current form are not anti-monopolist. Therefore, all sensible people who understand why free/libre/open software is important should reject them.

What constitutes an anti-monopolist LLM?

  1. Obviously, the model itself has to be anti-monopolist (free/libre/open) software.
  2. Anybody should be able to train the model independently, which means a list of all data sources should be published at the very least, or (ideally) all data sources should also be anti-monopolist data.
  3. Anyone should be able to do the training, which means it should be possible to do on any kind of hardware. (Which means they would also have to be vastly more power efficient than now.)
  4. They would have to run offline, rather than as centralized services. (This also resolves their privacy issues.)
  5. The weights have to be released, or it's no different from secret sauce binary-only software.
  6. The weights have to be considered derivative works of the training data, and licensed accordingly.

This would kill just about all popular LLM products. Therefore, none of them are anti-monopolist, and by using or funding them, we're falling for their bait.

Conclusion

In closing - if you make LLMs, fund them, or use them…that tells me -

  1. You don't care about the environment
  2. You don't care about community-made websites and resources

  3. You don't care about creative workers, or any workers at all. In other words, you are a useful idiot for the corporate overlords, and possibly a corporate bootlicker too.

    Maybe you think you're immune to the problems, but wait and watch…it's going to get you too. (If we let it.)

  4. You don't care about your privacy, the consequent power you're giving others over yourself, and the consequent erosion of our democracies.
  5. Really, you don't care about who or what you trample over, as long as you can get your fancy little toys, and satisfy peer pressure and FOMO.