[Air-L] Discussion of ChatGPT and other AIs and their uses

[Charles Melvin Ess]

Thanks to Peter for bringing this up - obviously critical discussion 
that has exploded in the literatures I'm still (somewhat) familiar with, 
especially over the last 5 years or so.

I've responded to Peter privately, partly because I wanted to send him a 
copy of a paper that is about to be published.  But here some of the 
additional response may be of more shared use or interest.

I and a close relative have been active users of ChatGPT over the past 
couple of years - primarily in the domains of scripting and elementary 
programming. At one point, for example, I was tinkering with C programs 
running under 2.11 BSD on a PiDP-11 emulator / replica I built as part 
of my exploring all of this. My relative comes to all of this with 40+ 
years of professional experience, including decades as a systems 
administrator / security expert.

The short story is that the results for us have been decidedly mixed. 
As a start: as Peter rightly notes, everything depends on asking the 
right question / the question properly. Failure to do so can lead down 
some long, very frustrating, and ultimately useless, if not destructive 
rabbit holes. (Though the same can also often result from the right 
questions ...)

It is also clear after a point what the underlying nature of these 
devices is - namely, sophisticated inference engines, totally devoid of 
any sort of semantic understanding, etc., that generate what I often 
call SWAGs - scientific (sorta) wild-ass guesses.  (Less kindly, they 
are "stochastic parrots" that only can regurgitate the material they 
have been "trained" upon - so Bender et al, 2022.)

It further becomes clear that we are often its guinea pigs / trainers: 
our responses to its suggestions - most especially then they go wrong as 
they so often do - are then taken up as part of its training data and, 
hopefully, helps improve it in some way or another for use further down 
the road.

Last but not least - and no surprise: it is "trained" along many of the 
sophisticated lines of the various algorithms and related techniques to 
keep our "attention" that are familiar from SoMe.  The most obvious 
examples are the occasional choices offered between two different 
"personalities." What my relative aptly calls the butt-kissing algorithm 
is also on strong display, along with other simple tricks - "you're so 
close, now," etc.
(In other words: follow the money ...)

I continue to use it - as informed and guided by this sort of awareness 
of its strengths and limits: most pointedly, not as a replacement for my 
own thinking / muddling through complex projects - including research 
and writing - but as a very limited augmentation.

This is, as those who know me might suspect, further grounded in a 
fairly extensive philosophical framework and (relational / feminist / 
critical posthumanist ...) anthropology, one rooted (as a start - only 
as a start) in the Socratic dialogues as well as Aristotle: specifically 
Plato's The Phaedrus and the warnings there towards the end of the 
dialogue regarding the risks posed by the then-new media technology of 
writing.
Contra a tendency in recent media and communication studies (from ca. 
2012 forward, based on what I've managed to find) to interpret the story 
of Theuth and Ammon's warnings as a "media panic" - I rather show in a 
forthcoming paper via an ancient interpretive framework (ring 
composition / inclusio narratives) that this story serves first of all 
as a pedagogical device that helps bring home several of the lessons 
that unfold for the young Phaedrus cross the course of the dialogue.

As a start: that "bare letters" alone only provide us with a _semblance_ 
of wisdom and knowledge, and thereby present the risk of what Shannon 
Vallor has aptly characterized as "deskilling," our failing to acquire 
the practices and abilities ("virtues") necessary for acquiring such 
knowledge and wisdom on our own - most centrally, the capacity of 
phronesis  as a form of self-correcting, specifically ethical judgment.
What the moral panic reading tends to further miss is the subsequent 
discussion of "the gardens of letters" - the positive, proper uses of 
writing (including the central art of writing well vs. writing 
disgracefully which is thematic to the dialogue from the outset) as a 
complementary media technology requisite to the philosophical pursuit of 
knowledge and the capacity for good judgment as essential elements to 
good lives of flourishing.

Such a complementary approach, last but not least, can be traced through 
early critiques of classic AI, including Joseph Weizenbaum and 
especially Hubert Dreyfus (1973) - who identifies Plato's Euthyphro as 
the beginning of the story of AI as it starts out with a request for a 
rule-based approach to ethics, which turns out to be futile: phronesis, 
as not reducible to rules, is required instead - including by AI 
programmers themselves, as Dreyfus further illustrates.

In more technical terms: there arises here what becomes called the frame 
problem. The upshot is that such judgment is not computationally 
tractable - hence Weizenbaum's foundational distinction between judgment 
and calculation (1976).

<Aside>: those familiar with the origins and development of the AoIR 
ethics guidelines for internet research may recognize that phronesis, 
contra rule-based approaches such as utilitarianism and deontology, has 
been central to our development and suggestions for use of these 
guidelines from the start.
Another discussion entirely, but I have the very strong impression, from 
several sources and experiences over the past 2+ decades, that by 
grounding our guidelines on the already solid and fruitful phronesis / 
judgments of both new and more experienced researchers as a primary 
beginning point, coupled with the aim of building and using the 
guidelines in such a way as to foster our own phronetic abilities - 
rather than tick-boxing or, ho ho, asking ChatGPT ... - are primary 
contributors to their extensive and successful use. </aside>

Still more recently, work by, e.g., Virginia Dignum (2019, 2021) and 
Katherina Zweig (2019) (a computer scientist and bioinformaticist, 
respectively, show similar limits to generative AI approaches - 
reiterating the point that AI / ML / LLMs offer only semblances of 
knowledge, and hence represent similar risks of deskilling should we 
turn more and more of our cognitive / writing / ethical, etc. loads over 
to the machines.
All of this captures a central point from the Phaedrus - "know your 
tools," i.e., their strengths and limits, in order to then know how to 
best make appropriate use of them towards these larger ends.

If any of this is of further interest - including those of us who have 
discussed the Phaedrus over the years here - I'm happy to send along the 
paper detailing all of this (forthcoming in the Danish Yearbook of 
Philosophy: Open Source). I conclude there with some examples of such a 
complimentary approach to using LLMs in contemporary research and 
teaching that might be helpful or suggestive in particular.

Of course, critical comments and suggestions always welcome.

Best of luck with it all and thanks again to Peter for raising these issues.

- charles ess
Professor Emeritus, University of Oslo

On 07/03/2026 08:17, Peter Timusk via Air-L wrote:
> Hello I haven't really posted to this list for a decade or more I think. I
> work in statistical computing these days and rarely spend time on my
> employer's ( Statistics Canada) Internet Use surveys though I did write one
> short study in 2009.
> 
> At work we do a lot of statistical computing and even analysts are
> expected to be able to calculate statistics for their own papers. We are
> moving off the statistical computing language SAS and going open source
> with R and Python. We are also expected to start exploring AI for use in
> making things more efficient and saving money.but also to do so with high
> ethical standards and of course working within our culture of citizen and
> business data privacy.
> 
> I just wanted to write this email to see if there is any discussion
> possible here via email.
> 
> I have used AI for the past few months to do some volunteer work. Here I am
> making an online collection of PDFs into a searchable set of blog posts on
> a Word Press site. I have learned that one can ask long search questions to
> ChatGPT and it will produce code examples and workflows, say to show me how
> to use R and the PDFTools an R package to extract text from these PDFs. I
> generally have to know what I am looking for to ask the questions but
> am happy so far with the results of this AI use. I also have to read the
> code and understand the loops and logic going on which allows my
> programming experience to validate the results.
> 
> I am about to use Python suggested by ChatGPT to make an offline archive of
> all my Facebook posts from 2007 to now and also publish these in some neat
> fashion to a WordPress blog.
> 
> anyways hoping to join the discussions on AI bias and ethics that have been
> going on for awhile.
> _______________________________________________
> The Air-L at listserv.aoir.org mailing list
> is provided by the Association of Internet Researchers http://aoir.org
> Subscribe, change options or unsubscribe at: http://listserv.aoir.org/listinfo.cgi/air-l-aoir.org
> 
> Join the Association of Internet Researchers:
> http://www.aoir.org/