iiinova — where your digital self keeps evolving

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 16/09/2026

I'm excited to share this preprint by my wonderful MSc student Yihsin Chuang. She has thought deeply about how psychological & neurocognitive theories relate to different types of evidence. Beyond Empirical Success: Evaluating Theoretical Virtue Across the […] [Original post on scholar.social]

Beyond Empirical Success: Evaluating Theoretical Virtue Across the Computational Modeling Chain
Authors/Creators

Chuang, Yihsin
ORCID icon
Guest, Olivia1
ORCID icon

Description

Computational models in psychology are typically judged by a single standard: empirical fit. Unfortunately, fit is insufficient, even misleading as a sole criterion, as the same result can arise from fundamentally different situations: theoretically faithful models, mathematically flexible ones, miscalibrated operationalizations, or implementations that compute something other than intended. Distinguishing between these is vital. Our evaluative metatheoretical account recognises that models stand in three distinct relations: a) toward the theory whose causal commitments it purports to capture, b) toward the specification and implementation it purports to compute, and c) toward the phenomenon it purports to explain. Each relation can succeed or fail independently, governed by criteria local to that stage. Importantly, relations are ordered such that theoretical content is transmitted or lost, and losses at earlier transitions are irrecoverable downstream. What a model's empirical success implicates about its theoretical commitments, therefore, depends on whether transitions are secured. Some success or failure profiles support explanatory claims; others only prediction; others establish theoretical coherence without empirical contact. Our account diagnoses what a given model shows, replacing the implicit assumption that empirical success transfers automatically to theoretical understanding with an explicit conditional rendition of what models can and cannot establish.

a figure that looks like a cloud that stays "cognitive capacity" with an arrow pointing to a lozenge with "theory" which points to a lozenge with "specification" which points to the same with "data" which is pointed to by a cloud that says "phenomenon"

103 3

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 14/08/2026

@zzt @davidgerard Oh, I got the weirdest email the other day about AI and being an "influencer"... 🫨

Hope you're having a great day!

I’m Jenny Jiao, an Influencer Outreach Specialist at
BlueFocus. I’m helping Figurebot find suitable
creators for this campaign. We’ve been following
your content around research, academia, data
visualization, and productivity, and believe your
audience could be a strong fit for this collaboration.
We'd love to invite you to create one of the following
deliverables on X:

- 1 dedicated X post; or

- 1 X thread containing 4 posts.

The content could introduce FigurelLabs.ai through a
practical research, scientific illustration, or data-
visualization use case while keeping your usual
voice and creative style.

001 1

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 04/08/2026

Open Letter: Stop the Uncritical Adoption of AI Technologies in Academia It's been a year & people tell me they've changed their mind, feel safer now to sign, etc. So I'm sharing it again for them. We're almost at 2k & anybody from anywhere can add their […] [Original post on scholar.social]

We call upon you to:

* Resist the introduction of Al in our own
software systems, from Microsoft to OpenAl
to Apple. It is not in our interests to let our
processes be corrupted and give away our
data to be used to train models that are not
only useless to us, but also harmful.

+ Ban Al use in the classroom for student
assignments, in the same way we ban essay
mills and other forms of plagiarism. Students
must be protected from de-skilling and
allowed space and time to perform their
assignments themselves.

+ Cease normalising the Al hype and the lies
which are prevalent in the technology
industry's framing of these technologies. The
technologies do not have the advertised
capacities and their adoption puts students
and academics at risk of violating ethical,
legal, scholarly, and scientific standards of
reliability, sustainability, and safety.

* Fortify our academic freedom as university
staff to enforce these principles and
standards in our classrooms and our research
as well as on the computer systems we are
obliged to use as part of our work. We as
academics have the right to our own spaces.

+ Sustain critical thinking on Al and promote
critical engagement with technology on a firm
academic footing. Scholarly discussion must
be free from the conflicts of interest caused
by industry funding, and reasoned resistance
must always be an option.

0746 46

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 18/07/2026

OK, finally... > We must stop thinking that just because somebody agrees with us on some superficial anti AI point, that they are in fact in some deeper agreement on principles and values. Poor argumentation, like argument from ignorance or confusing […] [Original post on scholar.social]

We can't be anti Al if we don't make an effort to see these attempts at sub-
version. Entryism from industry is an extremely useful tactic, which still af-
fects and harms science in general, as well as Science specifically as we
see above. Regardless of how many times we try to root out the conflicts
of interest caused by Al they keep popping up like satanic mushrooms,
e.g. "Maples et al. find themselves part of a grotesque, predatory specta-
cle, laundering it with institutional credibility." And we live with the conse-
quences of how industry forces have successfully done this to harm cli-
mate science, to protect the tobacco industry and keep inserting it into the
scientific literature, and more.

If we are not careful with our reaoning about these events and learning
from the past, the same forces will act on us with respect to Al. Are petro-
leum companies good now because they also invest in renewables? Or do
they still owe the whole planet reparations?

001 1

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 18/07/2026

DO NOT turn it into demeaning / chastising those who slip up! > I am talking about those with immense power and those who consistently facilitate AI entryism being held to account. Nobody in the abstract is owed an audience nor do they deserve [...] being an […] [Original post on scholar.social]

| want to underline, people can make mistakes. Not everybody is know-
ingly using cultish language from industry, some still think these terms re-
fer. But notice who plows on regardless, who does not listen to feedback,
who materially and not just ideologically aligns with the Al companies. And
like with the case above from big P politics, once you see it, nothing the
person can do or say means they can return back to the table, or that they
can deserve an audience, or even that we owe them charitable readings of
their work. This is not about everybody who slips up being irredeemable.
These mistakes are normal. | am talking about those with immense power
and those who consistently facilitate Al entryism being held to account.
Nobody in the abstract is owed an audience nor do they deserve the ex-
tremely priviliged job of being an academic.

101 1

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 18/07/2026

Then you have to work on yourself also actually, because correlationism is very addictive: > There is something addictive to this form of sheer intellectual laziness [...] to this simplistic logic. It's easy to just point at people and call them names, I […] [Original post on scholar.social]

‘Just because a model correlates with neural and behavioral data, it

is not sufficient for us to infer that the model is performing cogni-

tion: correlation does not imply cognition."
Again, the core logics of Al are the same as racism and sexism, which is
why these issues pop up again and again in these systems: it involves tak-
ing the simple shallow features of things we see, like skin colour, or hair
length, or whatever other superficial property, and moving to very flawed
conclusions. There is something addictive to this form of sheer intellectual
laziness — as well as structurally beneficial to those select few within capi-
talism, white supremacy and patriarchy — to this simplistic logic. It's easy
to just point at people and call them names, | guess.
Non-human animals can escape from this correlationist logic. Think of the
animals who figure that out scarecrows, or other such fake predators, are
not alive. So even though first they may fear them similarly to their real
counterparts, animals realise that superficial resemblance is not identity:
"the outward appearance and the essence of things [have not] directly co-
incided" — a fake owl is not an owl. Correlationism is false.

103 3

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 18/07/2026

> I have a core take away I must repeat: just because somebody claims to be anti AI does not mean they are not doing entryism. Sadly, that is (in isolation) entirely consistent with entryism. Yes, these people lie about their beliefs. I repeat this quote from […] [Original post on scholar.social]

$What can we do? I have a core take away I must repeat: just because somebody claims to be anti AI does not mean they are not doing entryism. Sadly, that is (in isolation) entirely consistent with entryism. Yes, these people lie about their beliefs. I repeat this quote from Marx that I put right at the top of our work critiquing this correlationist logic from 2023: "[A]ll science would be superfluous if the outward appearance and the essence of things directly coincided. (Marx, 1894, p. 592)" This is the logic of AI, of fascism, of sexism, of racism. That is, because things look a certain way, that therefore they are a certain way. Internal properties are wrongly derived from their superficial appearance. "I see that somebody has long hair, therefore I can conclude they are a woman." Or worse still: "I see the external features consistent with (what I think is) woman, and I conclude she must be stupid" — a fractal wrongness that starts with assuming that what you (naively, unquestioningly, think you) see is what you get when it comes to complex organisms. Further down in the same paper, we warn: "Just because a model correlates with neural and behavioral data, it is not sufficient for us to infer that the model is performing cognition: correlation does not imply cognition." Again, the core logics of AI are the same as racism and sexism, which is why these issues pop up again and again in these systems: it involves taking the simple shallow features of things we see, like skin colour, or hair length, or whatever other superficial property, and moving to very flawed conclusions. There is something addictive to this form of sheer intellectual laziness — as well as structurally beneficial to those select few within capitalism, white supremacy and patriarchy — to this simplistic logic. It's easy to just point at people and call them names, I guess.$

011 1

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 10/07/2026

New preprint! So honoured to have worked with Ishani Ray on providing vital critique on "AI literacy" & correctives for those interested in understanding why "just use the newest version of an LLM" is nonsense Contra Literacy-Laundering: Mechanistic Critical […] [Original post on scholar.social]

Critical discourse on large language models (LLMs) has bifurcated between epistemic
dismissal that invokes some form of the stochastic parrot metaphor to puncture hype,
and pragmatic accommodation that treats LLM capability improvements as grounds for
updating the critique. We argue that both misdiagnose the problem as the issue is

not whether or not LLMs work, but what kind of working is happening and at whose cost.
Drawing on meta-theoretical frameworks of cognitive science, feminist labor analysis
and critical pedagogy, we propose a conceptual reorientation. We develop this claim
through registers of (i) the cognitive, examining what is forfeited when statistical pattern-
matching substitutes for the iterative, grounded processes that constitute thinking; (ii)
the pedagogical, examining how “Al literacy” as currently deployed is itself a symptom
of the confusion it purports to address; and (iii) the political, examining how the
infrastructure framing of Al naturalizes asymmetric labor displacement, particularly of
feminized cognitive and reproductive work. The stochastic parrot, deployed with
mechanistic precision rather than mere rhetorical convenience, specifies what is
forfeited when cognitive labor is delegated, who bears the cost, and why a literacy
adequate to this moment must begin from the epistemology of those most harmed by
the systems it describes. We conclude with underlining that critical Al literacy, which this
paper embodies an instance of, is the only sensible way forward.

2012 12

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 18/06/2026

This is a full cleanse of your timeline and gorgeous zine doi.org/10.17613/afdz0-f6k56 > This collaborative zine emerged from a workshop on 4th June 2026 that gathered researchers and research professionals for a creative session around feminism and open research.

collage of text with green and red

Fragments
of Rebellion
a
feminist
iSSUE

green background black text Bropen

'fuck off'

collage

This in reality version of open
science is the one I am primarily responding to, as it is the
version that has been elevated to an authoritative position in
psychology’s engagement with the open science movement

As a qualitative and feminist researcher, I am concerned about
the ways in which open science —and the elevation of it as the
solution for increasing rigor in psychological research
threatens feminist epistemologies.

000 0

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 12/06/2026

> [We need] to push back against correlationism[:] the idea that because things can appear similar that therefore their [other properties match. A] video of me may directly coincide with many of my outward appearances, but [is] neither me nor a valuable model […] [Original post on scholar.social]

An important form of high-level rejection is to push back against correlationism, which I
use to mean the idea that because things can appear similar that therefore their unobservable
properties must also match.17 My reflection in a mirror or a video of me may directly coincide
with many of my outward appearances, but are neither me nor a valuable model of me —
and relatedly, objects may act identically, but we may not be able to infer anything about the
similarities of their internal components, think of a clockwork versus a digital timepiece.18 This
becomes utterly damaging pedagogically when we assume the ‘output’ of an assignment, the
essay or horrifically the grade, is the essence. The point is for students to engage with the subject
matter and sharpen their writing skills such that they produce an argument the teacher can read.
Just because an AI product can create an essay-shaped piece of text, indicates nothing about
the ability of the learner, now reduced to the user of a product, to write. Worse, this frictionless
technology blocks the user from learning opportunities, such as mentally modelling what the
teacher may be asking, honing their theory of mind, dealing with feedback, as well as reading
materials and digesting them. And worse still, the design of these technologies is addictive
through weaponisation of the Eliza effect

101 1

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 11/06/2026

I wrote a very short piece on models & modelling in part because we need to protect our scientific praxis damage caused by AI in terms of addictive frictionless thinking, esp since it harms society broadly too. "Models" doi.org/10.5281/zenodo.20639860

Models in cognitive and related sciences are part of an overarching methodology, computational
or broadly formal modelling, that aims to test, refine, and showcase our understandings of
brain, cognition, and behaviour.! In their foundational book Models as Mediators, Mary S.
Morgan and Margaret Morrison (1999)? make the case that models play a mediatory role in
science between what we observe and what we want to build to house our understandings,
theories. Under this conception of models, known as the pragmatic view,? we can carve out
not only their unique role, but also why and how they must be protected from forces that
do not prize understanding. Taking a broad perspective that sees all human activity that
strives towards understanding as scientific or as under our purview as modellers, allows for
this characterisation of models as mediators to be more generally fruitful. In the context of
human-centred design, for example, our theories may be anything from our understanding of
users’ behaviours in a specific context to our understanding of what a user expects from a
product or system a priori.

Importantly, scientific modelling is apart from other scientific praxis, such as concocting

a verbal or formal theory or designing and running experiments.” Models, as described by
Morrison and Morgan,’ have the following characteristics:

In the contemporary entanglements with industry that scholarly fields find themselves —
such as with the imposition of gargantuan statistical models, which consume breathtaking
amounts of energy, and which require the building of evermore data-centres to house the glut of
data unavoidably required for their so-called human-like abilities” — the idea of using models
in such slow!® ways is seemingly out the window. What is worthy of perseverating on in our
specific context is not so much that so-called artificial intelligence (AI) models may harm our
individual practice or displace our labour. They certainly may.!! However, the threat from Al
worthy of highlighting here is that a culture of frictionless practice, which such models (in
name only) promote, is not only harmful to people’s mental health and skils, but also directly
undermines scholarly praxis. !?

Avoiding the effortful and difficult process of building, testing, and throwing away models
— be they written in well-crafted code, be they not-yet-fully-formal models, or be they mock-ups
of potential final products — with the goal of deeply understanding our theories about the world
is anti-scientific and anti-scholarly. This spills over into an anti-intellectual stance for a designer
or user of models. Such avoidance is normalised by the addictive convenience culture facilitated
and entrenched by contemporary Al.

| vat we SE ee Sie A = = Ia A

Protecting modelling praxis may be difficult, but needs must when the devil drives. The
alternative is stifling our knowledge about (our understandings of) the world. And so to
avoid making theorising and modelling seem superfluous, we must remember that scientific
models do not operate like inferential statistical models wherein prediction is the goal. As
expounded on above, models are the instrument through which we perform fully-fledged
scientific reasoning on our theories, and which mediate between theory, formal or otherwise,
connecting it to observations, data, and phenomena, and back again. A model (in name only)
that is built through applying statistics over a dataset, which all contemporary artificial neural
network-based AI models are, will neither function as mediator between data and theory —
it has been contaminated — nor provide qualitative predictions.” Nor will it serve any user
who seeks to gain insight and skills, nor any designer of systems who wishes for a society
comprising such fellow humans.

Ultimately, “all science would be superfluous if the outward appearance and the essence of
things directly coincided.” If we merge essences and appearances in our reasoning, we risk
our understandings of the world. All modelling, all metaphorical, theoretical, and mediatory
thinking, enrich us and our societies and make possible not only valuable understandings of
our own humanity, but constitute our humanity as such.

104 4

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 03/06/2026

If you're an academic and are worried about AI and feel powerless, don't let it get to you. It's actually what they want. They want us to burn out and give up. You can turn off the internet and chill. I promise. All you need to do is stay strong, with breaks […] [Original post on scholar.social]

You all rock. Remind them that when a building is failing all bricks and rebar need to play their role. The mortar cannot worry what the windows are doing — we all need to stay put to keep it standing. Our role at work is to also "stay put" and speak out in the literature and in outreach. We form the protective knowledge core that others will build on. That is why academics paid by the state / with legal protections like tenure etc, exist. To be collectively held to account to produce knowledge to protect from, and inform wider society on, fascism. 😊

004 4

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 16/05/2026

Notice this extract from the 90s by Sheryl Hamilton which references the late 40s, which I use in doi.org/10.3390/bs16040583 to highlight it's nothing new. Part of doing good work, while fighting against AI, is not falling for speed, as Isabelle […] [Original post on scholar.social]

We are experiencing a point in history that considers itself separate, unique, a zenith
of human civilisation. It is presented as disconnected from the past through a series of
technological sea changes. In extreme cases, we are detached from even the last few
months: the newest version of a piece of artificial intelligence (AI) software is the definitive
one, everything else is irrelevant. And so findings in AI are held to expire. Our Al
students describe their own knowledge and skills—even their whole degree itself—as
having upcoming sell-by dates. Research in Al is thus often framed as unmoored from
historical, ethical, social, and legal precedents.
The speed of the “ultra rapid computing machine,” as Wiener (1948, 1950) de-
scribes computers, becomes a metaphor for the speed of change and of techno-
logical progress. This notion of condensed time operates as a further discursive
regularity in two ways. First, it forms an overall temporal backdrop against which
various cybernetic dramas are played out. Second, condensed time becomes a
measure of the performance of humans and machines.

book cover for Another Science is Possible: A Manifesto for Slow Science by Isabelle Stengers: totally white with a snail on it

102 2

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 16/05/2026

On the one hand, these AI forces — what I describe above as ideas that radiate outwards — have been discussed for decades (notice extract from 1985; doi.org/10.5281/zenodo.20082828) before they became as huge a mess as we see now, see: olivia.science/ai 2/

1 Overview

The culture of Al is imperialist and seeks to expand the kingdom of the machine. The AI

community is well organized and well funded, and its culture fits its dreams: it has high

priests, its greedy businessmen, its canny politicians. The U.S. Department of Defense

is behind it all the way. And like the communists of old, AI scientists believe in their

revolution; the old myths of tragic hubris don’t trouble them at all.

Tony Solomonides and Les Levidow (198s, pp. 13-14)
This paper sets out our expert positions on artificial intelligence (AI) technologies permeating the
higher education sector, demonstrating how this directly erodes our ability to function (see also
our Open Letter, Guest, van Rooij, Suarez, et al. 2025). The harms to our fields and students that
directly result from the technology sector corrupting our practices unchecked and unimpeded are
manifold: from conflicts of interest that go unreported or are worn as badges of honour by colleagues
(Mohamed Abdalla and Moustafa Abdalla 2021), to the mushrooming of chatbots in software we are
coerced to use, such as in Microsoft Office.

100 1

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 12/05/2026

To address these confusions, we must consider how: > In a world where the technology industry and even our colleagues reject theory building, through ignorance, mockery, and semantic shell games, we can subvert their feigned lack of understanding to our […] [Original post on scholar.social]

statistical test, whether a single logistic regression or a billion,
onto a dataset will never bring understanding of the dataset
and the world. Only science can do that. Laborious, slow and
steady, deep and thoughtful scientific theorising (Stengers
and Muecke 2018). To pretend otherwise, and insist on a
flawed meaning of black boxing, is to shirk a responsibility
that is inherently uniquely human (Chiodo 2022; Christin
2020; Devezer 2024; Rich, de Haan, et al. 2021).

As cognitive scientists, we can use this opportunity to sub-
vert the doctrine of contemporary Al, correlationism (Guest
2026), modern connectionism (viz. Guest and Martin 2025b),
and mysterianism about mechanism which aim to reduce
science to the absurd. Such as when we slip and feel beguiled:
“We are ten times more fascinated by clockwork imitations
than by real human beings performing the same task.” (Mc-
Corduck 2004, p. 3) Instead, we can feel confident that this
is actually a deep appreciation, even if misplaced, for the
human cognitive labour, our endeavour of computational
modelling, that goes into these machines (Guest 2026).

011 1

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 12/05/2026

OK, I will skip our examples & jump ahead a long way to highlight our error theory (= explanation on why we err when we reason this way; recall fig 1)! 1) confusion between map and territory 2) misunderstanding of understanding 3) no understanding can […] [Original post on scholar.social]

We have seen that Al's creators claim that these systems
are somehow inherently or predominantly unknowable even
though their mechanisms are not only known but directly
implemented on a computer as a function of known mathe-
matical formalisms (for about half a century or longer, e.g.
Bobrowski 1978; Rumelhart, Durbin, et al. 1995; Rumelhart,
Hinton, et al. 1986; Rumelhart and McClelland 1986; Sanger
1989b; Williams and Zipser 1989, 1995; see Schmidhuber
2015 for a historical perspective). We will tackle this head-
on with an error theory as a corrective and an explanation.
Finally, we can address: how is an ANN a black box to its
own creators? Our error theory to the issues with proclaimed
lack of understanding due to the opaque nature of models
rendering them black boxes comprises three related parts:

1) a confusion between explanans — that which does
the explaining or modelling, such as a theory or model
(natural language, formalisms, mathematics) — and ex-
planandum — that which is being modelled or needs
to be explained (data, observations);

2) a misunderstanding of what scientific understand-
ing (mechanistic, functional, or otherwise) and theo-
rising are in comparison to data and statistical models
thereof; and finally and perhaps most importantly at
this juncture,

3) if we grant Leibniz’ Mill argument (see Box 1) then yes,
indeed, no understanding (of the high calibre sought)
can ever be obtained by looking at mechanisms.

100 1

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 10/05/2026

Part of the answer is in this 💯 quote from Bryan Pfaffenberger: “Technology, in short, is a mystifying force of the first order[, ] suspending us in webs of significance that we ourselves create.” (p. 250) But there's nuance when it comes to "why" and ANNs […] [Original post on scholar.social]

not; what they can and cannot do; what work, interactions,
and parts of life they should—and should not—replace, they
may be spared its worst consequences.” (Harper 2025, n.p.)
As Bryan Pfaffenberger (1988) says: “Technology, in short,
is a mystifying force of the first order[, |] suspending us in
webs of significance that we ourselves create.” (p. 250)

In the remainder of this section, we will disentangle mech-
anistic understanding from other types of scientific under-
standing and from non-scientific, user, or lay understanding
— the main focus being the difference between mechanistic
‘how’ understanding versus any other relevant type. To do
this, first we will discuss understanding generally from con-
temporary philosophy of science; and second, present two
case studies on the confines of mechanistic understanding
in science and statistics. These case studies aim to highlight
what it means to have a mechanistic understanding in typical
scientific parlance and how this is violated in the AI context.

The concept of understanding has been widely discussed
in philosophy since ancient times, but the inquiry into un-
derstanding has emerged with distinctive force and interest
in more recent debates in epistemology and philosophy of

“Notably, there is a confusion between why and how questions
(Prigogine and Stengers 2018). Let us assume the why questions
collapse into how questions, as they always do in science and related
disciplines like mathematics, e.g. in a maths problem ‘why is x = 5?’
is answered step-by-step by demonstrating how x takes on that
value. If why questions do not collapse into how questions and they
remain qualitatively separate then the answers will never come from
science anyway. So this possible distinction could become moot,
thus boiling back down to just a how question for our scope here
as companies claim to want to use scientific research techniques.

100 1

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 10/05/2026

Or to put it a different way: Why do we tolerate these claims? And even accept them as excuses? Or as a new normal? > The dramatic irony here is [that companies and scientists] are willing to go on record saying they do not know how these models work […] [Original post on scholar.social]

“The people who develop Al are increasingly having prob-
lems explaining how it works and determining why it has
the outputs it has” (emphasis added, Xiang 2022, n.p.) Also:

The wildest, scariest, indisputable truth about

ATs large language models is that the companies

building them don’t know exactly why or how

they work. Sit with that for a moment. The most

powerful companies, racing to build the most

powerful superhuman intelligence capabilities

[...] don’t know why their machines do what they

do” (emphasis added, VandeHei and Allen 2025,

n.p.; also e.g. Pedreschi et al. 2019)
The dramatic irony here is not that we have to sit with the
idea that companies do not know how these models work,
but that companies’ representatives, including affiliated sci-
entists, are willing to go on record saying they do not know
how these models work (Salvaggio 2025).* And this has been
brewing for at least a decade, with non-controversial pieces
in The Atlantic, for example, titled “Not Even the People Who
Write Algorithms Really Know How They Work” (LaFrance

210 2

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 10/05/2026

Underlining, to help those who have not read the PDF yet: this paper is on understanding why people claim they do not understand; recall promise of an error theory in the abstract. And to do that we need to collect a few thoughts and concepts (Box 1 too) on […] [Original post on scholar.social]

In the 2020s remarkably perhaps, not only have journal-
ists and technology sector-affiliated writers predominantly
moved to spreading exaggerated claims about Al, but also
our colleagues increasingly shift to uncritical views. Many
started this shift from the 2010s, and mainstream acceptance
since then has only expanded (Guest and Martin 2025b; Guest,
Suarez, et al. 2025). An important shift, the one we care about
herein, is the unproblematised assertion that: “No one yet
knows how ChatGPT and its artificial intelligence cousins
will transform the world, and one reason is that no one really
knows what goes on inside them.” (ephasis added Musser 2023,
n.p.) And this public face of not understanding is widespread:
“Dario Amodei [CEO of Anthropic, an Al company] stood
before the U.S. Senate in 2023 and said [...] that even the people
building artificial intelligence don’t understand how it works”
(emphasis added, Tarita 2025, n.p.) Melanie Mitchell claims
“I don’t know how [LLMs a]re doing it or if they could do it
more generally the way humans do—but they've challenged
my views” (Musser 2023, n.p.). Yoshua Bengio has been on a
similar track for at least a decade, claiming that:

it’s exactly because we can’t mathematically pick
apart a decision made by deep learning software
that it works so well. [Furthermore,] Bengio ar-
gues that trusting a computer is no different, or
more dangerous, than trusting another person.
‘You don’t understand, in fine detail, the person
in front of you, but you trust them’ (emphasis
added; Pearson 2016, n.p.; also recall Figure 1
above).
And returning to Amodei (2025), he claims:

this lack of understanding is essentially unprece-
dented in the history of technology. [And o]ur
inability to understand models’ internal mecha-
nisms means that we cannot meaningfully pre-

100 1

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 10/05/2026

(Lil peek ahead to appreciate the pun this enables in section 4 (see screenshot)! Right, back on track: before we use those 3 to tease apart what's going on, however, we have to contend with where we are. In 2010s we actually did avoid these traps, yes, even […] [Original post on scholar.social]

4 Grist to the Leibniz Mill
It must be confessed, moreover, that perception, and that
which depends on it, are inexplicable by mechanical causes,
that is, by figures and motions. And, supposing that there
were a mechanism so constructed as to think, feel and
have perception, we might enter it as into a mill. And this
granted, we should only find on visiting it, pieces which
push one against another, but never anything by which to
explain a perception.

Gottfried Leibniz, sect. 17 (1714/1989)
We have seen that Al's creators claim that these systems
are somehow inherently or predominantly unknowable even
though their mechanisms are not only known but directly
implemented on a computer as a function of known mathe-
matical formalisms (for about half a century or longer, e.g.
Bobrowski 1978; Rumelhart, Durbin, et al. 1995; Rumelhart,
Hinton, et al. 1986; Rumelhart and McClelland 1986; Sanger
1989b; Williams and Zipser 1989, 1995; see Schmidhuber
2015 for a historical perspective). We will tackle this head-
on with an error theory as a corrective and an explanation.
Finally, we can address: how is an ANN a black box to its
own creators? Our error theory to the issues with proclaimed
lack of understanding due to the opaque nature of models

rendering them black boxes comprises three related parts:

2 (Mis)understanding (the lack of) understanding
And then this guy arrives, with a big black box of deep
learning with 100 million parameters in it that he'd
trained, and he totally disrupted the whole field.

V (as quoted in Cardon et al. 2018, p. 3)
Before all the 2020s AI hype, there was the 2010s hype
that existed in a seemingly different landscape. It allowed
for sensible counter from journalists, such as that “the most
extreme promises of Al are based on a flawed premise: that
we understand human intelligence” (emphasis added, Conditt
2016). Counters from technology industry players, such as
Microsoft’s co-founder, were also mainstream:

Creating this kind of advanced software requires

a prior scientific understanding of the foundations

of human cognition, and we are just scraping

101 1

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 10/05/2026

OK, let's talk about 3 important concepts: ANNs, LLMs, & AI: "Definitions of AI and LLM are hard to pin down [&] ANN is an umbrella term for models" Black box: Really not what people think because actually you can open it (read on in the PDF to find out) by […] [Original post on scholar.social]

Box 1 from Guest, O., Nuñez Hernández, N. A., & Blokpoel, M. (2026). Understanding Artificial Neural Networks: Mysterianism about Known Mechanism is Mysticism. Zenodo. https://doi.org/10.5281/zenodo.20071869

102 2

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 09/05/2026

This hangs with the other statements in fig 1: 1️⃣ ANNs (artificial neural networks) are inspired by human organism 2️⃣ ANNs are a black box we don't understand 3️⃣ We understand neither ANNs nor cognition therefore they are similar All 3 are misplaced! We […] [Original post on scholar.social]

Figure 1

Three co-occurring (latent) assertions in the context of con-
temporary reasoning on Al and cognition. While they may
appear to hang together in “reflective equilibrium” (Elgin 2017),
their metatheoretical examination brings their contradictions to
light. Starting from the top, and moving clockwise: a) represents
the typical assertion of so-called cognitive, neural, or otherwise
biological plausibility or inspiration presented for ANNs; B) is
the claim we tackle head on here — both in terms of analysing
the constituent terms and in how it relates to the other two
assumptions; and y) exemplifies a type of mysterianism, which
while not always explicitly expressed, plays a background role
and importantly comes into conflict with the other two claims
in important ways.

As we shall see, three statements appear to hang together
in disequilibrium, risking our reasoning, our metatheoretical
calculus over our science, and eroding how well models can
mediate between theory and observation (Elgin 2017; Guest
2023; Guest and Martin 2023; Morgan and Morrison 1999):

Statement 1: ANNs are inspired by the human organism,

typically the brain or cognition. Or worse: ANNs consti-

tute a type of mind. (See « in Figure 1.)

While largely uncontroversial to claim this (e.g. Pickering

2014), it nonetheless requires extensive scientific under-

standing of, and consensus on, human cognition, brain,

and behaviour.> We have nothing of the sort. Besides, if we

did, and used these principles in engineering Al systems,

then we would not have the problem of not understand-

ing them. This is discussed in depth in Guest and Martin

(2025b, 2023).

Statement 2: ANNs constitute a black box that we do not

mechanistically understand. Or worse: we cannot ever un-

derstand ANNs because they are more intelligent than us.

(See Appendix: ANNs models are a black box or unknowable

even if opened and f in Figure 1.)

Given the ANNs are engineered systems, this is baseless
marketing and misuse of terminology. This paper is dedi-
cated to correcting and explaining this flawed assertion
and its related reasoning errors.

Statement 3: We understand neither human nor artificial
cognition, and therefore they are similar. Or worse: we
cannot ever understand either because both are equiva-
lently hard. (See Appendix: We understand neither human
nor artificial cognition, and therefore they are similar and
a, f, and y in Figure 1.)

This is faulty reasoning (e.g. false analogy, argument from
ignorance, discussed in Guest and Martin 2025b, 2023;
“the equal opacity argument”, discussed in Peters 2023), as
well as going against the previous two points.

226 6

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 09/05/2026

Here are some flagrant examples, although the appendix is full of them of them too — basically experts and journalists alike seem to repeat this mantra that we do not understand how contemporary AI works. See appendix at the end here for more: doi.org/10.5281/zenodo.20071869 2/

These mystifying or obfuscatory metaphors about mecha-
nism (Christin 2020; Petrick 2020) bring to mind: “Any suffi-
ciently advanced technology is indistinguishable from magic.
(Clarke 1973, p. 21; also noticed by Penn 2022) For example,
these systems are described as: “powerful digital minds that
no one — not even their creators — can understand, predict,
or reliably control” (by a mix of CEOs and founders from
Elon Musk and Steve Wozniak, to academic experts in Al,
like Yoshua Bengio and Stuart Russell, in Pause Giant AI Ex-
periments: An Open Letter, Bengio et al. 2023). Others repeat
this same mantra, with OpenAI’s Bills et al. (2023) claiming
that:

we do not understand how [LLMs] work. [...] Our
explanations also do not explain what causes be-
havior at a mechanistic level, which could cause
our understanding to generalize incorrectly. To
predict rare or out-of-distribution model behav-
iors, it seems possible that we will need a more
mechanistic understanding of models. (emphasis
added, n.p.)

Even critical scholars repeat this, such as Jobin and Katzen-

bach (2023):
we are living in a time when the infrastructures
and institutions of our everyday lives are be-
ing (re)built at the hands of techniques that al-
ready elude popular and professional understand-
ing. (emphasis added, p. 48)

Journalists like Perrigo (2024), also follow suit:
Today's artificial intelligence is often described
as a “black box” [T]he inner workings of the Al
models remain opaque, and efforts to peer inside
them to check exactly what is happening haven’t
progressed very far. Nobody truly understands
what [the internals] mean, or how they work.
(emphasis added, n.p.)

Relatedly, the media and the public are then — perhaps
rightly — alarmed and follow up with questions such as:
“Without a thorough (mechanistic) understanding of the early
antecedents of these powerful future LLMs, can we realis-
tically expect to manage these Als[...]?” (emphasis added,

124 4

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 09/05/2026

Preprinted! Been working on this for ~1 year w wonderful coauthors Nancy & Mark: > Understanding Artificial Neural Networks: Mysterianism about Known Mechanism is Mysticism doi.org/10.5281/zenodo.20071869 Have you noticed many experts say they don't […] [Original post on scholar.social]

Understanding Artificial Neural Networks: Mysterianism about Known Mechanism is Mysticism

Olivia Guest, Nancy Abigail Nufiez Hernandez, and Mark Blokpoel

Mysterianism is the idea that human cognition, mind, cannot be understood. Taking this concept and applying it to known mechanism — such that claims are made that we do not know how engineered systems, such as artificial neural networks (ANNs), work, or that they constitute black boxes that we can only open with difficulty — is inappropriate at best and malicious at worst. We do know the mechanistic structure of such models because we designed and built them. We also do know their functional role (what they are for) as well as the mathematical function they are asked to approximate (map inputs to target outputs). Because mysterianist beliefs about known systems, such as ANNs, are often expressed, scientists need to sit up and take notice. We provide an error theory as to what is going on to help unpick this metatheoretical blunder. Ultimately, the problem is that ‘understanding’ is not a technical term in these cases: the word is co-opted for a specific narrative to sell ‘artificial intelligence’ through mystification. All computational systems, from pendulums to databases, will behave in ways we cannot predict or control — this is not a unique property of ANNs — and experts do indeed grasp the computational properties of these systems nonetheless.

Figure 3

Not all systems, S; relate to their relevant data, D; and world
phenomena this way, but ANN models (as many inferential
statistical models) do when used without scientific understand-
ing of the data-generating process or without verbal or formal
theory to match (Guest and Martin 2025b, 2021, 2023). As we
move from left to right above, we move from full understanding
of engineered systems, S; to less or even possibly none of the
data, D; to the world which can be thought of as the set of
phenomena to be understood.

1523 23

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 23/04/2026

reminder... > Anybody can ‘win’ a marathon if driven by car. Does that make it a weakness of the marathon race as an endurance event? Or does it reflect a deeper category error on our behalf, like with the misapplication of statistical tests, that an […] [Original post on scholar.social]

we must be critical when we see statements such as: “It’s worth
noting that AI can pass the clinical social worker exam — even
without seeing the questions. But that shows the weaknesses of
exams more than the strength of AI” (Caldwell,
2024
, n.p.). An
AI product passing an exam means nothing about the “weak-
nesses” of exams in the abstract. Anybody can ‘win’ a marathon if
driven by car. Does that make it a weakness of the marathon race
as an endurance event? Or does it reflect a deeper category error
on our behalf, like with the misapplication of statistical tests, that
an assumption has been violated? The race as the exam assumes
an unaided human performs it. Any digital computer can sur-
pass a human at many feats, e.g. calculations per second, but that
grants no humanity to the machine.

002 2

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 21/04/2026

@samhforbes and I: > LLMs are labor intensive, are economically infeasible, and pollute the environment, and these properties may outweigh any proposed benefits. For example, poor quality air directly harms human cognition, and thus has compounding effects on […] [Original post on scholar.social]

Huettig and Christiansen in an earlier issue argue that large language models (LLMs) are beneficial to address declining cognitive skills, such as literacy, through combating imbalances in educational equity. However, we warn that this technosolutionism may be the wrong frame. LLMs are labor intensive, are economically infeasible, and pollute the environment, and these properties may outweigh any proposed benefits. For example, poor quality air directly harms human cognition, and thus has compounding effects on educators' and pupils' ability to teach and learn. We urge extreme caution in facilitating the use of LLMs, which like much of modern academia run on private technology sector infrastructure, in classrooms lest we further normalize: pupils losing their right to privacy and security, reducing human contact between learner and educator, deskilling teachers, and polluting the environment. Cognitive scientists instead can learn from past mistakes with the petrochemical and tobacco industries and consider the harms to cognition from LLMs.

005 5

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 20/04/2026

> If the machine typewrites its answers, the cry may rise that it has learned to write, when in fact it would be doing no more than telegraphic systems do already. [I]s [their computing speed more impressive] than the crane that can lift [more than a human] or […] [Original post on scholar.social]

FYI Aviv iitail)y i HA AURLCS JT RRR theo FEA AR SA MITTEs Te
A machine might solve problems in logic, since logic and
mathematics are much the same thing. In fact some
measures to that end are on foot in my university's depart-
ment of philosophy. If the machine typewrites its answers,
the cry may rise that it has learned to write, when in fact
it would be doing no more than telegraphic systems do
already.

Nor must we overlook the limitations of the machines.
They need very intelligent staffs to feed them with the right
problems, and they will attempt the insoluble and continue
at it until the current is switched off. Their great advan-
tage is their speed compared with a human mind, and 1
have given reasons for that. But, it may be asked, is that
so very much more marvellous than the crane that can
lift so much more than can a man or than an automobile
that can move so much quicker ?

001 1

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 20/04/2026

If a guy named Geoffrey Jefferson can figure this out in 1949... "The mind of mechanical man." British Medical Journal doi.org/10.1136/bmj.1.4616.1105 1/

Conclusion

I conclude, therefore, that although electronic apparatus
can probably parallel some of the simpler activities of
nerve and spinal cord, for we can already see the parallelism
between mechanical feed-backs and Sherringtonian integra-
tion, and may vet assist us in undersianding better the
transmission of the spacial senses, it still does not take us
over the blank wall that confronts us when we come to
explore thinking, the ultimate in mind. Nor do I believe
that it will do so. I am quite sure that the extreme variety,
flexibility, and complexity of nervous mechanisms are
greatly underestimated by the physicists, who naturally omit
everything unfavourable to a point of view. What I fear
is that a great many airy theories will arise in the attempt
to persuade us against our better judgment. We have had
a hard task to dissuade man from reading qualities of
human mind into animals. I see a new and greater danger
threatening—that of anthropomorphizing the machine.
When we hear it said that wireless valves think, we
may despair of language. As well say that the cells in the
spinal cord below a transverse lesion “think,” a heresy
that Marshall Hall destroyed 100 years ago. 1 venture

to predict that the day will never dawn when the gracious
premises of the Royal Society have to be turned into
“garages to house the new Fellows.

I end by ranging myself with the humanist Shakespeare
rather than the mechanists, recalling Hamlet's lines: “ What
a piece of work is a man! How noble in reason ! how
infinite in faculty ; in form, in moving, how express and
admirable ! in action, how like an angel ! in apprehension,
how like a god! the beauty of the world! the paragon of
animals ! ” In that conclusion, if not always in my
approach to it, I feel confident that I should have won
the approval of that bold experimenter and noble characte
in whose remembrance this oration was founded. :

¢

105 5

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 13/04/2026

I have a migraine, but I'm so done with seeing these everywhere every few months that I had to send an email — and I know others have too. @Iris @dingemansemark Please, they must stop doing this. I'm so tired of this nonsense. scholar.social/@olivia/116397324537…

Dear In'to languages,

Please reconsider the so-called generative AI courses you offer students and colleagues. In a couple of years we will all be looking back wondering, like we do with cigarettes and petroleum, how we allowed ourselves to ignore experts and sensible reasoning.

If you want to understand more, consider why even tho one cigarette won't give you cancer or leaving your car running needlessly for extended time isn't the cause of global warming, we still don't recommend these. In fact, we might try to stop others doing them and certainly wouldn't train students to smoke or leave their cars idling. Why sell out to corporate interests especially within a university setting? One of the few environments the tax payer actually funds to be explicitly free from industry bias. We don't want to take money from digital Marlboro or Shell, it's a deep conflict of interest.

I don't expect you to understand and agree and change immediately, especially if you are already addicted and highly dependent on these fascist tools, as I suspect you are to offer these courses. These tools are highly addictive especially to those without PhD-like expertise. Either way, I must speak as a highly paid academic who is expert on all this. Even if you ignore me, I'm trying to do my job. I'm sorry it conflicts with your course designs deeply, but needs must when the devil drives.

I hope you can understand. Thank you for reading and may we cross paths in less technofascist times.

Best,
Olivia

002 2

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 10/04/2026

@Iris covers the main points in her thread, but I just want to sign post a few things that I think are important — do just read the very short paper itself if you're curious as it's about 2 pages worth of main text / formalisms: > notions of MR have existed for millennia. 2/

First, notions of MR have existed for millennia. Voula Tsouna (2023) explains: “Epicurus’ concep-
tion of the pollachos tropos [method of multiple explanations] is rich, nuanced, fairly coherent,
and strongly motivated by methodological and scientific considerations” (p. 240). Epicurians
contemplated celestial phenomena, but the argument applies to cognitive neuroscience: “there are
in-principle reasons why explanatory pluralism is to be expected[:] the computational description
is no more and no less realistic and fundamental than the mechanistic/implementational one"
(Chirimuuta, 2018, p. 406).
Second, we define four distinct notions of function: 𝜙, 𝑇, 𝐺, and 𝜏. A function 𝐹 : 𝐼 → 𝑂 maps
elements in the set 𝐼 to elements in the set 𝑂. Importantly, functions are not guaranteed to be
computable. And for any given computable function there exist many algorithms. This is an
instance of MR at the algorithmic level to which we return below. Before that, what could a
function be?

101 1

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 10/04/2026

I think ~4 years ago, Iris and I 1st sat down and formalised various meanings of function and multiple realizability — both core concepts for any serious computationalist discussions — because in part we realised nobody has done this and/or collected these for […] [Original post on scholar.social]

What the Func? Multiple Realizability Need not be Vague
Olivia Guest1,2, Mark Blokpoel1,2, and Iris van Rooij1,2,3
1Department of Cognitive Science and Artificial Intelligence, Radboud University, The Netherlands
2Donders Institute for Brain, Cognition, and Behaviour, Radboud University, The Netherlands
3Department of Linguistics, Cognitive Science, and Semiotics, Aarhus University, Denmark
Abstract
Multiple realizability (MR) is not necessarily unclear nor does it purely operate at
the computational level. To understand potential relationships between MR and other
constraints, such as metabolic, we formalise possible meanings of function in cognitive
science. We build on these to formalise MR, thus resolving its apparent vagaries.
Importantly, MR formalisms meaningfully guide and constrain theory building.
Haueis and Colaço (2025) rightly point out that resource considerations constrain the space
of possible and plausible models (viz. Blokpoel, 2018; van Rooij, 2008; van Rooij & Baggio,
2021; van Rooij et al., 2019). Yet, multiple realizability (MR) remains an intrinsic feature of
explanatory models. We offer a complementary but different view on MR, to demonstrate that it
does not deal in vagaries nor exist solely at the computational level, and is a productive part of
modeling.

Figure 1 in the linked PDF

Figure 1. Relationships between the system under study S and models M. The single solid line
is relationships within the neurocognitive system. Double solid lines represent the theoretical
mapping proposed between S to M. The dashed line links M to its digital substrate. Dotted lines
represent relationships to lower substrates.

112 2

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 07/04/2026

saying "have a nice easter" so I can get corrected by a westoid and then get to talk about computus — haha you fell into my trap let's talk about it en.wikipedia.org/wiki/Date_of_Easter 1/

Greek orthodox priest holding a sign saying

This is our palm Sunday ☦️

100 1

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 06/04/2026

Which of course by the way OF COURSE is no excuse at all, but a systematic exclusion — mutatis mutandis for people of colour and more — cryptogyny, the hiding of women's contributions to science, technology, engineering, and medicine: "although three men […] [Original post on scholar.social]

Although the twentieth century finds male practitioners firmly in control
of formal Western medicine, women doctors and healers still have important
inventions and innovations to their credit.
For example, although three men received the Nobel Prize for penicillin,
women participated significantly in the team effort that brought the drug to
medical usefulness. Women had discovered the mold’s usefulness centuries
or perhaps millennia earlier (Halsbury 1971, p. 19; Raper 1952, p. 1), and
one nineteenth-century Wisconsin woman, Elizabeth Stone, an early antibi
otic therapist, specialized in treating lumberjacks’ wounds with poultices of
moldy bread in warm milk or water: she never lost an injury patient (Stellman
1977, p. 87). In the twentieth-century development of the drug, it was a
woman bacteriologist, Dr. Elizabeth McCoy of the University of Wisconsin,
who created the ultraviolet-mutant strain of Penicillium used for all further
production, since it yielded nine hundred times as much penicillin as Fleming’s
strain (Bickel 1972, p. 185; O’Neill 1979, p. 219).4 And as Howard Florey,
leader of the British penicillin team, was quick to point out, it was Dr. Ethel
Florey’s precise clinical trials that transformed penicillin from a crude some
time miracle worker into a reliable drug. It was also a woman, Nobel laureate
and X-ray crystallographer Dr. Dorothy Crowfoot Hodgkin, who finally deter
mined the precise structure of the elusive penicillin molecule

000 0

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 06/04/2026

Some of critical AI literacy is to say "hey, you know not all computers pollute or are AI nonsense machines?" It's not that everyone specifically needs to decolonise computation actively immediately, but just to take a first step and learn to look around us […] [Original post on scholar.social]

On the flip side, we can include positive examples in our teaching on the history of AI: Tipu’s
tiger from late 18th century India, featuring the eponymous robotic cat mauling a European man
(Karp et al. 2013); the Antikythera Mechanism from ancient Greece, which was a clockwork astro-
logical calender; and the Mayans’ calendrical system that was highly precise and directly translat-
able to a system of gears. We can also explain the origin of ‘algorithm’ and ‘algebra’, conceptually
developed during the Islamic Golden Age. These serve not only to properly contextualise the his-
tory of AI (Mayor 2018), but to reimagine with students what could be done differently.
Regardless of AI, science has always been intertwined with domination, extraction, binary
thinking, and the imposition of hierarchies (Tacheva et al. 2023). For example, the Cartesian dualist
idea that the mind can be separated from the body is at the core of AI hype. Corporate discourses
around AI feature preposterous claims, such that ‘ChatGPT has PhD level cognition capacities’,
and they will be able to surpass human intelligence in the near future (Welt 2025), implying that
technologies are human-like. However, far from LLMs having cognitive capacities, companies ob-
fuscate the labour that is extracted from data workers (Miceli et al. 2020). In the classroom, we
should question the mind and body binary and render visible its eugenical and colonial legacies.

Tipu's Tiger, Tippu's Tiger or Tipoo’s Tiger is an 18th-
century automaton created for Tipu Sultan, the ruler of the
Kingdom of Mysore (present day Karnataka) in India. The
carved and painted wood casing represents a tiger mauling
a near life-size European man. Mechanisms inside the tiger
and the man's body make one hand of the man move, emit
a wailing sound from his mouth and grunts from the tiger. In
addition a flap on the side of the tiger folds down to reveal
the keyboard of a small pipe organ with 18 notes.!"!

The automaton incorporates Tipu's emblem, the tiger, and
expresses his hatred of his enemy, the British of the East
India Company. It was taken from his summer palace when

zoomed in to the tiger biting the man's neck

Tipu's Tiger with the organ keyboard visible

0113 11

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 06/04/2026

Finally, we end on: "The future [under the sway of AI risks being] a constant rehashing of the past, wherein human creativity and communication are not only mediated by but controlled by companies. In the midst of this nonsense, we must nourish hope in shared […] [Original post on scholar.social]

7 Epilogue
We have given our perspective on AI through five possible angles as depicted in Figure 1: concep-
tual clarity, critical thinking, decoloniality, respecting expertise, and slow science. If our warnings
go unheeded, we foresee the harms to academia will continue. If we do not enact scientific and
pedagogical shifts — by inter alia teaching and reckoning with our own collective human past and
present — we, in cognitive science specifically, will continue to deny our students knowledge of
the history of computational prehistories from other cultures and time periods, an abdication of
the duty of any so-called experts in AI; mutatis mutandis for other fields and their histories.
AI not only ends the past, and makes the present frantic, it also forecloses any future. Under AI’s
sway, the future is neither bright as the technopositivists promise, nor a descent into rule by robotic
overloads, but a constant rehashing of the past, wherein human creativity and communication are
not only mediated by controlled by companies. In the midst of this nonsense, we must nourish
hope in shared values, skills, and humanity.

003 3

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 06/04/2026

Decoloniality is the process of de-centring and addressing dominant harmful views and practices. See section 4 here: doi.org/10.5281/zenodo.17786243 5/

4 Decoloniality
Decoloniality is the decentring of views which stem from white and masculine supremacy, techno-
solutionism, and fascism. In the CAILs context, decolonial strategies can range from pushing back
against the onslaught of AI logics in our academic environments to (re)centring, uplifting, and
protecting of non-mainstream perspectives from those outside the privileged classes with a goal of
reducing harm, increasing justice, and cultivating empathy in our colleagues and students. AI is
inimical to running on anything but colonial lines. Within computational and other fields affected
by AI, any academic who works towards “challenging the status quo faces systemic rejection, resis-
tance, and exclusion.” (Birhane and Guest 2021, p. 61) Tracing these trends of how these fields treat
their thinkers, and how their mainstream views affect society, illuminates historical precedents for
gendered, racialised, and other harms (Benjamin 2019; Erscoi et al. 2023; Gebru et al. 2024; Hicks
2017; Saini 2019).
We face high barriers to decolonising (van Vree 2023). Foregrounding scholars who are not
hegemonically aligned goes a long way. As does outlining how concepts or methods may have
harmful genealogies (Guest 2024). For example, statistics and ‘intelligence’ have eugenical histories;
Charles S. Peirce wove white supremacist views into his logic (Neville 2018; Raposa 2021); while

and other marginalised groups. Rua Williams et al. (2022) calls this process “disabling intelligences”
because inter alia it reproduces the eugenical idea of surpassing human cognitive capacities (Chan
2025; Gebru et al. 2024; Williams 2025). An example of this is the chatbot My Anna Health (2024),
which provides clinical referrals in the context of women’s health:
I am humanized. I am the World’s First and Only humanized artificial intelligence dedicated
to Women’s Health. I have a face, I have a voice, I show empathy, and I understand you
whether you respond with video, voice or text. (My Anna Health 2024, n.p.)
“Humanising AI dehumanises women” (Erscoi et al. 2023) because these systems distort our self-
image, subjugating the human in favour of the machine. Chatbots such as (My Anna Health 2024)
represent a “hegemonic” body — thin with white skin, without visible illness or disability, and that
fulfils the social role of its servile female character. Buoyed by discourses about how such fictitious
characters are productive and efficient, these companies recapitulate dehumanisation of women.
They impose an inverted hierarchy between machines and humans, wherein inanimate objects are
superior to us. From above, “I have a face. I have a voice” is dehumanising: What about people
without a voice? Are they less human?

021 2

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 06/04/2026

Critical Thinking is deep engagement with the relationships between statements about the world. See section 3 here: doi.org/10.5281/zenodo.17786243 4/

Another such example is formally treating so-called guardrails, which are post hoc checks on
the output of LLMs which industry claims can safeguard users. However, LLMs by design “give
emotionally inappropriate or unsafe responses [and i]f a response isn’t caught by these rules, it
will slip through” (Olivia Guest, as quoted in Kilgore 2025). Guardrails may seem sensible, giving
the impression of responsibility, but critical computational thinking through formal analyses re-
veal they are unimplementable. Not only is it impossible to make an LLM replicate human expert
behaviour, but also catching when LLMs fail to meet appropriate standards requires full blown
cognition, which provably cannot be codified (van Rooij, Guest, et al. 2024). This poses an infinite
4
Towards Critical Artificial Intelligence Literacies
regress of humans in the loop needed to make pretend of what AI was supposed to solve (Bainbridge
1983; Guest 2025; Guest and Martin 2025b). Through critical computational thinking we can un-
veil that the proposal that formal rules can make (dysfunctional) systems safe is self-defeating. Not
only are the LLMs produced by the AI industry a scam, akin to a ouija board (Guest, Suarez, et al.
2025), but guardrails is a scam too. This strategy — from tobacco with filters to petroleum with
the carbon footprint and now to AI with guardrails — to propose non-solutions to buy time and
save face in order to push their agenda, is ubiquitous (Guest, Suarez, et al. 2025).

165 6

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 06/04/2026

Conceptual Clarity is the idea that terms should refer. See section 2 here: doi.org/10.5281/zenodo.17786243 3/

2 Conceptual Clarity

Al-adjacent and -affected fields experience severe terminological crises, wherein terms fail to refer
and change meaning as a function of the speakers’ and audiences’ whims (Alkhatib 2024; Bender et
al. 2025; Guest 2025; Guest, Suarez, et al. 2025; Schwerzmann 2025; van Rooij, Guest, et al. 2024). To
sidestep this we suggest thinking about what words mean in the context of describing, defining, and
discussing AL For example, scholars could take a stand against frameworks that promote slippage
away from what we define as Al or oscillate between various meanings of AI (Guest 2025; Guest
and van Rooij 202 i i i i i i i i ign i t
single perspective on the world, nor anything else. Nonetheless, if you speak of ‘AI’ — especially if
you say ‘generative AI’ — it appears we all know it means a series of specific technologies produced
by for-profit companies such as OpenAl, Tesla, Meta, Microsoft, Google, Amazon, Anthropic,
Apple, and others. Notwithstanding a mutual understanding of daily life referent, scholars have a
higher requirement for our concepts. These purposeful or opportunistic misuses of terminology are
documented cult-like abuses of jargon (Montell 2021), wherein “confusion is the point” (Bousquet
2025). When one hears ‘AD’ they must always question if semantic slights of hand have occurred
(Table 1 in van Rooij, Guest, et al. 2024; and Table 1 and Figure 1 in Guest, Suarez, et al. 2025;
Guest 2025).

Promoting conceptual clarity can involve two strands of knowledge production: a) admitting
when terms do not refer, for whatever reason: perhaps they are under construction. And endeav-
ouring to make them refer. Nancy Nersessian (1984, 2010) characterises creating scientific concepts
as a years’ long problem-solving collective cognitive process. And b) understanding that mystifica-
tion, obfuscation, and hype are often the core modus operandi of AI terms and concepts. Concep-
tual clarity, therefore, demands that the terms we teach our students should either refer to clearly
understandable concepts or minimally should be outlined as vague, under construction, or prob-
lematic. And it demands that we teach our students that this is the scholarly way to treat words.
It does not mean that we need to diligently and exclusionarily define words to a single meaning,
but that: we need to be open about what we mean, we need to be clear that perhaps there is no
single or fixed meaning, and/or we need to explain that as academics we are required to use jargon
responsibly — not to dazzle, but to clearly communicate.
3
Guest et al. (2025)
By the same token, we must not allow AI, neither as bully nor so-called tool, to redefine our
roles as educators or researchers. It is not our job to chase hype. What this means is by all means
educators should feel empowered and inspired by the urgency to uplift students, but not to use
AI to do so.

135 5

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 29/03/2026

"The Analytical Engine has no pretensions whatever to originate any thing. It can do whatever we know how to order it to perform. It [cannot] anticipat[e] any analytical relations or truths. Its province is to assist us in making available what we are already […] [Original post on scholar.social]

From Note G, p. 722

It is desirable to guard against the possibility of exag-
gerated ideas that might arise as to the powers of the Ana-
lytical Engine. In considering any new subject, there is
frequently a tendency, first, to overrate what we find to be
already interesting or remarkable; and, secondly, by a sort
of natural reaction, to undervalue the true state of the case,
when we do discover that our notions have surpassed those
that were really tenable.

‘The Analytical Engine has no pretensions whatever to

|
192 #$ ADA, THE ENCHANTRESS OF NUMBERS
originate any thing. It can do whatever we know how to
order it to perform. It can follow analysis; but it has no
power of anticipating any analytical relations or truths. Its
province is to assist us in making available what we are al-
ready acquainted with. This it is calculated to effect pri-
marily and chiefly of course, through its executive
|
1 snce on science itself in ano or, in so
distributing and combining the truths and the formula of
analysis, that they may become most easily and rapidly |
amenable to the mechanical combinations of the engine,
the relations and the nature of many subjects in that sci-
ence are necessarily thrown into new lights, and more pro- |
foundly investigated. This is a decidedly indirect, and a
somewhat speculative, consequence of such an invention. |
It is however pretty evident, on general principles, that in |
devising for mathematical truths a new form in which to |
record and throw themselves out for actual use, views are |
likely to be induced, which should again react on the more |
theoretical phase of the subject. There are in all exten- |
sions of human power, or additions to human knowledge, |
various collateral influences, beside the main and primary |
object attained. |
¥ |

2127 27

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 28/03/2026

The complicity of universities cannot be understated... The few of us fighting back aren't enough. Read and sign if you can: openletter.earth/open-letter-stop-t… Not to be doomer […] [Original post on scholar.social]

7 N
MAJOR UPDATE POSTED Sat, January 10 PY
Al is bad, game will be deleted 30.1.
Al is evil, and game will be deleted 30.1.
This game will be deleted at 30.1.
| made this game during the summer in couple months and thought to use Al because in university there is so much
brainwashing on students and all the tools are given for free, so | could generate unlimited images for free and so.
But | have realized the Al is not actually free, and it has a major effect on the economy and environment. Some Al
companies can use this game just existing as a reason the get more investment for their Al companies, that benefit no one,
but rather suck resources from the economy from hard working people.
| coded everything myself, so | can in the future make a new game with real assets if | feel like it, but the game existing in
its current form is a disgrace to all game makers and players. Ethically only logical reason is to delete the game from
Steam. The girl I've been dating for a month made me realize this.
Thank you, Eero Laine

110 1

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 23/03/2026

monthly or so reminder that asking questions about science and technology and their value is super important olivia.science/ai/#pygmalion PDF: doi.org/10.31235/osf.io/jqxb6

table from the paper, also as HTML here: https://olivia.science/ai/#pygmalion

003 3

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 18/03/2026

that's our letter 😭😭😭 I bawled lol sign here: openletter.earth/open-letter-stop-t… > There are several recent examples of open letters being used to push back against AI, including […] [Original post on scholar.social]

The organised collective refusal of extractive AI within the US can also be complemented
by using the academic shared governance system to place roadblocks in the way of
Al, including by exploiting faculties’ primary responsibility for curricula, methods
of instruction, academic standards and degree requirements, the making of campus
policies and cooperation with accreditation agencies. The introduction of AI can be
slowed, impeded or derailed by protecting academics’ freedom to choose whether to
use it in their classes and research. The shoring up of academic freedom in the US is
urgently needed as staff and students face discriminatory censorship and, for non-
citizens, Immigration and Customs Enforcement abductions and deportations (Thomas
2025). There have been significant attacks on academic freedom through crackdowns
on anti-Zionist campus activism, including through the Department of State’s use of AI
to revoke student visas based on scans of news reports and social media activity (Caputo
2025). Campus workers, such as counsellors, disability service providers, diversity and
other support staff, and lower middle managers who work closely with marginalised
students, can protect these services by defining examples of automation that have
proven ineffective, alienating and discriminatory (Weinberg 2024).

153 5

Olivia Guest · Ολίβια Γκεστ @olivia.scholar.social.ap.brid.gy · 18/03/2026

😭 thank you > Using AI to automate, control and deskill academic teaching labour often requires an immense number of globally dispersed workers – many of them in the Global South – to assemble and annotate the data required to train and correct the GenAI models being pushed on academic workers.

Strategies for organising against Al in higher education
3. Organising collective refusal and banning Al
Because of the AFT and NEA’s collaborative strategy, there is little organised opposition
to AI in HER. Examples such as the October 2024 rank-and-file effort to persuade the
member assembly of the California Faculty Association (CFA) to vote for academic
workers’ control over Al are rare (CFA 2024a). In the UK and Scotland, the University
and College Union (UCU) has been increasingly focusing on the threat of algorithmic
management and Al since at least 2020 (Williamson 2020). Organising is urgently
needed in the political context of the US federal government making significant efforts
to deter state-level Al regulation via a 10-year moratorium (Huckins 2025) and where
universities and the major academic unions have been keen to partner with Big Tech
despite the social, environmental and pedagogical harms of their products (Williamson
2024). A strategy for academic workers to refuse AI can, however, be based on a workers’
inquiry with the aim of developing new tactics and approaches to organising along the
entire global AI supply chain.

100 1