On Tue, 28 Jul 2026 at 12:43, Ian Jackson <[email protected]>
wrote:
> Aigars Mahinovs writes ("Re: Ballot option: Call the Free Software
> community to develop Free AI."):
> > This analogy fails completely. Racist material is racist because of its
> > **content**. Not just because it was written by racist people. You
> refuse to
> > even look at LLM-generated content.
>
> As I have said repeatedly, I regard LLM output as harmful, in general.
> This is not just my weird opinion. There are a number of studies that
> show extremely alarming, and surprisingly rapid, cognitive degradation
> and psychological harm in LLM users. We don't really know yet whether
> that damage is reversible.
>
If you share those studies, we might look at them. But I can not support
such findings with my personal or professional experience. *Some* people
are showing cognitive degradation and harm from something as simple as
social media usage. In the past there were studies showing
irreparable psychological harm of video games. This is why I view such
claims and even studies with a heavy dose of scepticism.
> Obviously not all LLM output is equally harmful but I very strongly
> wish to minimise my exposure. Given that almost all uses of LLMs are
> unethical because the LLMs themselves are produced by such highly
> unethical means, I feel this is a very reasonable position.
I see your position on the ethics point and I understand it. I disagree,
but this is an easy opinion-based position to have. And I expect it will
evolve over time as more "ethical" LLMs get developed and become more and
more usable. And by evolve I do mean changes in both directions - people
against LLMs on principle softening their stance where "good" LLMs are
concerned and also people using LLMs raising their standards on which LLMs
to use as more "good" alternatives appear and become practicable.
We already can see that stratification today. Few people would consider it
ethical to use Grok (with who it is enritching and also how they generate
energy for their data centers), but on the other end of the spectrum
running some open-weights LLM on my own, already existing and bought,
hardware powered by a local solar farm creates far fewer negative
externalities. Models trained in sustainable datacenters, like EU model
SOOFI already exist as well
https://huggingface.co/spaces/Soofi-Project/Pretraining-Tech-Report
Naturally there is difference between "resource usage" ethics and "they
stole all our copyrights" ethics. I expect that progress on one of them
will be easier than on the other. And someone would have to come up with a
very convincing argument why it is perfectly ethical for a human to go to a
library, read all the coding books, read all open source code on Github and
then start writing code under their own copyright, but a software model
doing the exactly same action with exactly the same inputs and outputs is
not ethical. I have not seen any convincing argumentation put forward on
this question. "but its a human and that makes it different" is not
convincing and not really even an argument to me, it's just a statement and
an assumption.
If there are other branches of the "ethical" evaluation of LLMs that I am
missing, please let me know. I am always happy to learn.
--
Best regards,
Aigars Mahinovs
