On Fri, 2026-09-04 at 08:35 -0300, juan via agora-discussion wrote:
> Ok. So I reverse my position and say we should regulate. I'd say that the
> regulation should probably say something to the effect that no automated
> messages should be sent to the lists.

"Automated" is too strong a restriction: automated messages that are
written by (human-written) programs, rather than LLMs or other AIs,
have been commonly used in Agora's past and have been unproblematic.
(For example, there have been instances of writing programs to send
emails at specific times in order to win timing scam wars without
needing to be awake, and cases where offices were automated using a
program that would automatically send the report once the events had
been updated by hand.)

I suspect that, if we end up with a consensus to ban LLM-authored
content from the lists, the restriction should say that directly (e.g.
"do not post passages of text to public or discussion fora that were
primarily authored by a large language model"). Such a ban would also
include copied-and-pasted LLM output, which I think is probably
desirable (I see "a person copies and pastes LLM output to the mailnig
list" as being more problematic than either a person using an LLM
without posting the output to the lists, or an LLM playing without
explicit human approval, because at least the "LLM plays autonomously"
situation teaches us something about what the weaknesses of LLMs are
and how to combat them, whereas the human-copies-and-pastes scenario
doesn't even do that).

As for the topic more generally, I think my position is close to "I
strongly dislike LLMs but am not going to pretend that they don't
exist; I am not going to use them myself but acknowledge that other
people use them; and because people are going to use them, it can be
important to understand what their strengths and weaknesses are,
especially for the purpose of being able to demonstrate the weaknesses
to people who might consider using them".

There's also something of an economic paradox based on how LLM use is
generally charged for: most people use LLMs via a subscription in which
they pay a fixed amount per month for a certain amount of LLM usage.
These subscriptions are priced on the basis that most people won't use
the entirety of their subscription, and in fact the less of a
subscription is used, the more money the LLM company makes. As such, I
would prefer people to not purchase subscriptions to LLMs, but for
people who have done so, using the subscription as much as possible
reduces the profits that the LLM companies make. I think it is possible
that this phenomenon outweighs the environmental damage of the LLM
queries, although I am not sure. This means that, if someone is
planning to use the spare capacity of an LLM subscription they would
otherwise be wasting on Agora, that is to me not obviously a net
negative in terms of externalities.

I definitely do sympathise with the point of view of "I hate reading
LLM output so much that I don't want to see it in any circumstances",
though, even if I don't share it myself (to me, seeing obviously
incorrect or useless LLM output is a good thing, especially if it is
paired with an explanation of how it is incorrect, because it helps to
demonstrate why such tools should not be used).

-- 
ais523

Reply via email to