iiinova — where your digital self keeps evolving

Iris van Rooij 💭 @irisvanrooij.bsky.social · 17h

How come Wikipedia is doing better than academic publishing? Btw love that icon: 🚫🤖 🔗 en.wikipedia.org/wiki/Pragmat...

This article may incorporate text from a large language model, which is prohibited in Wikipedia articles. It may include hallucinated information, copyright violations, claims not verified in cited sources, original research, or fictitious references. Any such material should be removed. The reason given is: see the explanation at Special:PermanentLink/1368212033#Query (August 2026) (Learn how and when to remove this message)

Iris van Rooij 💭 @irisvanrooij.bsky.social · 17h

I wanted to link to the Wikipedia page for 'pragmatics', and then saw this announcement. We need more of that. Such announcements I mean, not more AI slop. Also, love that icon: 🚫🤖 🔗 en.wikipedia.org/wiki/Pragmat...

0133

Iris van Rooij 💭 @irisvanrooij.bsky.social · 06/10/2026

Footnote 23 in Reclaiming AI: link.springer.com/article/10.1...

footnote 23 Similarly, no finite set of ‘impressive’ observations about AIs/LLMs
establishes that they exhibit a (human) cognitive capacity. This link can
only be theoretically established; i.e., if we propose that a machine has a
cognitive capacity X , we must also formally charactise X such that the
machine’s capacity Y can be mathematically proven to be X ≡ Y under
relevant and naturalistic conditions, using appropriate proof techniques
(see, e.g. Blokpoel & van Rooij, 2021, Chapter 5).

2478

Iris van Rooij 💭 @irisvanrooij.bsky.social · 03/10/2026

"Capacities are not tasks" -- van Rooij & Guest (2026) journals.sagepub.com/doi/10.1177/...

be ruled out, it may seem that AI systems that can
(be made to) perform like humans on cognitive tasks
provide at least possible theories of how cognition could
work. However, this inference is not licensed either.
Computational models of tasks, and task performance,
are not yet theories of cognition (Guest & Martin, 2021;
Morrison & Morgan, 1999). Theories of cognition should
at a minimum provide possible explanations of one or
more substantive human cognitive capacities, such as
vision, decision-making, reasoning, concept formation,
learning, or communication (Cummins, 2000; Egan, 2018;
van Rooij & Baggio, 2021). “A capacity is a more or less
reliable ability (or disposition or tendency) to transform
some initial state (or ‘input’) into a resulting state (‘out-
put’)” (van Rooij & Baggio, 2021, p. 684). A cognitive
capacity is thus the ability to transform cognitive states
(e.g., percepts, preferences, beliefs) into other cognitive
states (e.g., beliefs, plans, decisions, actions). Although it
is true that in (hyperempiricist) psychology cognitive
capacities are typically studied by having people perform
various tasks, computationally (or, more often, statisti-
cally) modeling task performance does not yield explan-
atory theories of the full-blown capacities. This is no

only because tasks do not map one-to-one, or in any
other straightforward way, to substantive cognitive capac-
ities, but also because even if they could, the models
would not be able to scale up to situations of real-world
complexity.
AI systems cannot scale. For computational models of
cognition to be able to scale up from toy scenarios (such
as studied in the psychological labs or such as form the
bases of training AI systems through machine learning),
these models should minimally be computationally trac-
table (van Rooij, 2008). That is, the input-output map-
pings hypothesized by the computational models should
in principle be computable using a realistic amount of
resources (Box 1). However, when existing computa-
tional models are mathematically analyzed for their
resource demands it turns out that such models are either
tractable but limited in their input domains, or they are
domain-general but then the models are intractable (for
a textbook introduction to this methodology, see van
Rooij et al., 2019). Consequently, no computationally
tractable account exists to date for substantive and
domain-general cognitive capacities, such as reasoning,
communication, decision-making, planning, analogizing,
categorization, and concept formation (van Rooij et al.,
2019). Moreover, there is no good reason to believe that
such tractable accounts will be forthcoming via machine
learning (van Rooij, Guest, et al., 2024) or otherwise
(Rich et al., 2021). Hence, if someone claims their con-
crete AI system is a “theory” of human cognition, they are
more likely than not overstating the scope and capacities
of the system (van Rooij et al., 2019) and obfuscating the
system’s limits, including the human in the loop needed
to make such systems “work” (Guest & Martin, 2025

13211

Iris van Rooij 💭 @irisvanrooij.bsky.social · 30/09/2026

The paper is short and we’ve included a Glossary of terms to further improve accessibility. Download paper here: 📝 journals.sagepub.com/doi/epub/10.... If you find our paper useful in your teaching of students in Psych, AI, or otherwise, @olivia.science and I would love to know.

Table A1. Glossary of terms used in the article

0328

Iris van Rooij 💭 @irisvanrooij.bsky.social · 30/09/2026

We end with guidance on how Psych & AI can benefit Cognitive Science if these traps are avoided. This alternative approach disavows attempts to automate science—using machine learning or otherwise—and instead centers scientists’ own cognition. Download here: 📝 journals.sagepub.com/doi/epub/10....

Figure 2
A visual depiction of the connections between the Cognitive Sciences. Solid lines denote stronger interdisciplinary ties; and dashed lines denote weaker ones. This figure is derived from the original put forth by the Sloan Foundation in 1978 and reproduced from Figure 4 in Pléh and Gurova (2013). Different versions of it over time have used ‘Artificial intel- ligence’ (as above) instead of ‘Computer Science’ and vice versa (cf. Miller, 2003).

1245

Iris van Rooij 💭 @irisvanrooij.bsky.social · 30/09/2026

We present a typology of traps to avoid: 1. Believing that AI systems are minds 2. Believing that AI systems are theories 3. Believing that cognitive science can be automated. Learn to recognise and avoid these traps. Failure to avoid leads to numerous problems.

Table 1: Typology of traps, what goes wrong if not avoided, and how the traps can be avoided. Note that all traps in a sense constitute category errors (Ryle & Tanney, 2009) and the success-to-truth inference (Guest & Martin, 2023) is an important driver in most, if not all, of the traps.

18626

Iris van Rooij 💭 @irisvanrooij.bsky.social · 30/09/2026

“Many of our contemporaries now want to combine the worst of these two worlds [i.e., Psychology and Artificial Intelligence]. What could possibly go wrong? Quite a lot.” 📝 journals.sagepub.com/doi/epub/10.... cc @olivia.science

Combining Psychology With Artificial Intelligence: What Could Possibly Go Wrong?

Abstract
The current AI hype cycle combined with psychology’s various crises make for a perfect storm. Psychology, on the one hand, has a history of weak theoretical foundations, a neglect for computational and formal skills, and a hyperempiricist privileging of experimental tasks and testing for effects. Artificial intelligence, on the other hand, has a history of conflating artifacts for theories of cognition, or even minds themselves, and its engineering offspring likes to move fast and break things. Many of our contemporaries now want to combine the worst of these two worlds. What could possibly go wrong? Quite a lot. Does this mean that psychology and artificial intelligence can best part ways? Not at all. There are very fruitful ways in which the two disciplines can interact and theoretically contribute to cognitive science, for instance, by studying the scope and limits of computational models of human cognition. But to reap the fruits, one needs to understand how to steer clear of potential traps.

Keywords
theoretical psychology, artificial intelligence, cognitive science, computationalism, epistemology

13710

Iris van Rooij 💭 @irisvanrooij.bsky.social · 30/09/2026

✨ Published version ✨ Iris van Rooij & Olivia Guest (2026). Combining Psychology with Artificial Intelligence: What Could Possibly Go Wrong? journals.sagepub.com/doi/10.1177/... @olivia.science Thread with highlights 🧵

Figure 1
Illustration of why AI systems cannot realistically scale to human cognition within the foreseeable future: (b) Human cognitive capacities (such as reasoning, communication, problem solving, learning, concept formation, planning etc.) can handle unbounded situations across many domains, ranging from simple to complex. (a) Engineers create AI systems using machine learning from human data. (d) In an attempt to approximate human cognition a lot of data is consumed. (c) Making AI systems that approximate human cognition is intractable (van Rooij, Guest, et al., 2024), i.e., the required resources (e.g. time, data) grows prohibitively fast as input domains get more complex, leading to diminishing returns. (a) Any existing AI system is
created in limited time (hours, months or years, not millennia or eons). Therefore, existing AI systems cannot realistically have the domain-general cognitive capacities that humans have. [Made with elements from freepik.com.]

36128

Iris van Rooij 💭 @irisvanrooij.bsky.social · 30/09/2026

Additionally, be aware of responsibility washing. See e.g. 📝 Amnesty International (2026). Unlawful by design: Exposing the human rights costs of generative AI. www.amnesty.org/en/documents...

Cover image of Amnesty International (2026). Unlawful by design: Exposing the human rights costs of generative AI. https://www.amnesty.org/en/documents/pol40/0996/2026/en/

27437

Iris van Rooij 💭 @irisvanrooij.bsky.social · 30/09/2026

"we show that contemporary AI can comprise research misconduct and demonstrate how critical literacy can protect us from this." -- @olivia.science & @irisvanrooij.bsky.social 📝 Guest, O. & van Rooij, I. (2025). Critical Artificial Intelligence Literacy for Psychologists. doi.org/10.31234/osf...

Table 1
Core reasoning issues (first column), which we name after the relevant numbered section, are characterised using a plausible quote. In the
second column are responses per row; also see the named section for further reading, context, and explanations.

03213

Iris van Rooij 💭 @irisvanrooij.bsky.social · 29/09/2026

See, e.g., Section 4 in our position paper for an overview of how the principles can be applied to 'AI': 📝 @olivia.science , @marentierra.bsky.social et al. (2026). Against the Uncritical Adoption of 'AI' Technologies in Academia. Digital Culture & Education, 16(2), 85–118. doi.org/10.5281/zeno...

Honesty implies that we do not secretly use AI technologies without disclosure, and that one does not make unfounded claims about the presumed capabilities of AI technologies (this also follows
from Responsibility; see example from MIT Economics 2025, where perhaps too little too late was
done).

Scrupulousness demands, among other things, that scientists only use AI products whose functional-
ity is well-specified and validated for its specific scientific usage (cf. Kwisthout 2024; Kwisthout and
Renooij 2025). This includes terminological precision about what formalisms, models, and/or tech-
nologies are used (recall Figure 1) and rigorous argumentation to motivate why these technologies
are appropriate for the scientific purposes at hand.
Transparency requires that the AI technologies are open source and computationally reproducible.
Here we must recall the technology industry’s obfuscatory tactics: “the name of the current
producer of ChatGPT. ‘OpenAI’ sounds like it is engaged in open science, but as we have now
seen, ‘open’ never really means what you think it does.” (Mirowski 2023, p. 738; see also Dingemanse
2025; Hao 2025; Jackson 2024; Liesenfeld and Dingemanse 2024; Liesenfeld, Lopez, et al. 2023;
Maffulli 2023; Maris 2025; Nolan 2025; Solaiman 2023; Thorne 2009; Widder et al. 2024)
Independence means that scientists ensure that their research is unbiased by AI companies’ agendas,
and that any potential conflicts of interest are declared in publications and other public communi-
cations (this also follows from Honesty and Transparency; cf. Mohamed Abdalla and Moustafa
Abdalla 2021; Atkin 2025; Forbes and Guest 2025; Knoester et al. 2025).
Responsibility precludes scientists from using AI products whose use is irresponsible, e.g. harmful to
people, animals, and the environment, or otherwise in violation of legal guidelines (e.g. copyright,
data privacy, labour laws, Butterick 2025; Cole 2025; Rijo 2025; Tafani 2024a). Minimizing harm is
vital for both engin…

1386

Iris van Rooij 💭 @irisvanrooij.bsky.social · 29/09/2026

Dear @erc.europa.eu , we DO already have common standards & frameworks. 'AI' likes to pose itself as 'special' but we do not need to reinvent the wheel. We already have established codes of conduct for research integrity. We just need to apply them. See below + more resources in thread🧵

Protecting the Ecosystem of Human Knowledge: Five Principles
We must protect and cultivate the ecosystem of human knowledge. AI models can mimic
the appearance of scholarly work, but they are (by construction) unconcerned with truth
— the result is a torrential outpouring of unchecked but convincing-sounding “informa-
tion”. At best, such output is accidentally true, but generally citationless, divorced from
human reasoning and the web of scholarship that it steals from. At worst, it is confidently
wrong. Both outcomes are dangerous to the ecosystem.
Olivia Guest, Iris van Rooij, Marcela Suarez, et al. (2025, n.p.)
Knowledge production is supposed to be safeguarded by (inter)national codes of conduct for research
integrity (ALLEA 2023; KNAW et al. 2018). Such codes forbid, for instance, fabrication of data,
falsification of results, plagiarism, and, generally, distortion of the scientific record. Many argue that
new rules are required to regulate academic AI use, but pre-existing guidelines fit the bill (cf. Tafani

219077

Iris van Rooij 💭 @irisvanrooij.bsky.social · 29/09/2026

Rereading my Conclusion, I think still relevant and timely

5. Conclusion
I agree with the authors of the target articles that computational approaches to cognition
have real problems in explaining the context sensitivity and speed of human cognition (see,
e.g., Haselager, van Dijk, & van Rooij, 2008; Kwisthout et al., 2011). I am also sympathetic
with the ideas that the human brain need not be organized as a massively modular system,
behavior need not be planned by a central executive, and behavior may be emergent from
the complex interplay between brain, body and world (see, e.g., Haselager et al., 2008; van
Dijk, Kerkhofs, van Rooij, & Haselager, 2008; van Rooij, Bongers, & Haselager, 2002).
The proposal that such an interplay can be characterized as a process of self-organized con-
straint satisfaction seems to me quite elegant. But, with this proposal, the dynamical
approach is not yet home free. In order to really explain why and how a brain–body–world
coupled system can quickly converge on constraint-satisfying configurations, dynamical
theorists need to build explicit models of self-organized constrained satisfaction that have
the property that they afford such quick convergence. I sincerely hope that dynamical mod-
elers will take on this challenge, as I believe that the approach holds promise of solving one
of the key problems in cognitive science, that is, the inexplicable speed of cognition and
behavior.

120

Iris van Rooij 💭 @irisvanrooij.bsky.social · 29/09/2026

Btw, I also wrote this on self-organisation: 📝 van Rooij, I. (2012). Self-organization takes time too. Topics in Cognitive Science, 4, 63-71. onlinelibrary.wiley.com/doi/pdf/10.1...

Self-Organization Takes Time Too
Iris van Rooij
Donders Institute for Brain, Cognition, and Behaviour, Radboud University Nijmegen
Received 4 February 2011; received in revised form 8 November 2011; accepted 8 November 2011
Abstract
Four articles in this issue of topiCS (volume 4, issue 1) argue against a computational approach in
cognitive science in favor of a dynamical approach. I concur that the computational approach faces
some considerable explanatory challenges. Yet the dynamicists’ proposal that cognition is self-orga-
nized seems to only go so far in addressing these challenges. Take, for instance, the hypothesis that
cognitive behavior emerges when brain and body (re-)configure to satisfy task and environmental
constraints. It is known that for certain systems of constraints, no procedure can exist (whether modu-
lar, local, centralized, or self-organized) that reliably finds the right configuration in a realistic
amount of time. Hence, the dynamical approach still faces the challenge of explaining how self-orga-
nized constraint satisfaction can be achieved by human brains and bodies in real time. In this com-
mentary, I propose a methodology that dynamicists can use to try to address this challenge.
Keywords: Complexity; Self-organization; Dynamical systems; Modeling; Explanation; Constraint
satisfaction; NP-hard; Intractable

141

Iris van Rooij 💭 @irisvanrooij.bsky.social · 29/09/2026

Btw, extreme synchronicity, I found this today on my desk in my office and I do not know why. This was a core book I read and built on for my Master thesis on DST, among other things.

Book by J.A. Scoot Kelso, Dynamic Patterns

261

Iris van Rooij 💭 @irisvanrooij.bsky.social · 27/09/2026

Much enjoying @samiraibnelkaid.bsky.social 's Reclaiming Our Futures festival together w/ @marentierra.bsky.social @ethicsofai.bsky.social @markdingemanse.net @joostvossers.bsky.social and many more not on Bluesky reclaimingourfutures.org

The future is rhizomatic

(picture is a ginger root)

Reclaiming our futures

1152

Iris van Rooij 💭 @irisvanrooij.bsky.social · 26/09/2026

Oh look, it is brainlike! Must be intelligent 🤔

722246

Iris van Rooij 💭 @irisvanrooij.bsky.social · 25/09/2026

We have a new campus cat She is so cute 😻

71215

Iris van Rooij 💭 @irisvanrooij.bsky.social · 23/09/2026

"Please Don't Use AI for Reviews" -- by @ffassnacht.bsky.social (repost from LinkedIN)

This goes out to everyone but particularly also to early stage researchers:

Please do not use AI if you get invited for review papers for a scientific journal. There are several critical problems related to this: 1. It is not allowed to upload unpublished work to LMMs (copyright problems) 2. Be aware - editors will in 98% of the cases know that you used AI and it leaves a very bad impression on you 3. Be aware that editors would be perfectly capable to ask a LLM themselves if they would be interested AI's opinion. 4. An AI-review will force the editor to invite another reviewer and delay the process heavily. 5. Trust in your own capabilities - conducting reviews is actually also a good way to learn. Even a statement like "I cannot follow your thoughts in this section" may be helpful advice to the authors to improve their work. AI will not give such advice and is typically also not great at spotting problems in the experimental set-up.

So please - if you do not have the time to really review a manuscript or you are not capable of reviewing the manuscript adequately, just decline the invitation. Thanks!

AI-free artwork created in MS paint.

15616

Iris van Rooij 💭 @irisvanrooij.bsky.social · 20/09/2026

I guess you do not know what cognitive science is. Here a useful illustration

Fig. 1 A visual depiction of the connections between the cognitive
sciences. Solid lines denote stronger interdisciplinary ties, and dashed
lines denote weaker ones. This figure is derived from the original put
forth by the Sloan Foundation in 1978 and reproduced from Fig. 4 in
Pléh and Gurova (2013). Different versions of it over time have used
‘Artificial intelligence’ (as above) instead of ‘Computer Science’ and
vice versa (cf. Miller, 2003)

3726

Iris van Rooij 💭 @irisvanrooij.bsky.social · 19/09/2026

Is it me or is this an AI-generated image of a brain on a Cogntive Psychology textbook? 👀

2111

Iris van Rooij 💭 @irisvanrooij.bsky.social · 19/09/2026

Thanks for sharing. Reminds me of our work showing that "good enough" selection of "good enough" heuristics is also intractable. mu.social/profile/iris...

Box 2. Proof intuition
NP-hard models are considered to be intractable, because there do not exist any tractable (formally, polynomial-time) algorithms for computing them (unless P = NP, see footnote 1). How can one prove that a computational model is NP-hard (and thus intractable)? A model is NP-hard if it is at least as hard as another NP-hard model, that is, one can use the former to compute the latter with minimal (only polynomial) overhead. We can prove such a relationship using polynomial-time reduction, a technique from computational complexity (Garey & Johnson, 1979). In the supplementary materials, we use this technique to prove that Adaptive Toolbox is NP-hard. Specifically, we reduce from the known NP-hard graph problem Dominating Set (which takes as input any graph and asks as output a so-called dominating set, that is, a subset of vertices with the property that for each vertex v in the graph, either v or one of its neighbors is in the subset). This is done as follows:
Provide a tractable (polynomial-time) algorithm that transforms any input iDS for Dominating Set into input iAT for Toolbox Adaptation (proof step 1 in Supplementary Materials).
Show that after the transformed input is processed as specified by Toolbox Adaptation, the resulting output oAT can be transformed back into a solution oDS for Dominating Set (proof steps 2 and 3).
Both transformations only take polynomial time to compute. This means we can use Toolbox Adaptation to solve Dominating Set with minimal (only polynomial) overhead.
Toolbox Adaptation is NP-hard, because it is at least as hard as the NP-hard problem Dominating Set.
This is, in a nutshell, the strategy we used to prove that Toolbox Adaptation is NP-hard. Fig. 2 illustrates this strategy.

1104

Iris van Rooij 💭 @irisvanrooij.bsky.social · 11/09/2026

From: Guest et al. (2026, p. 100). zenodo.org/records/2008...

Independence means that scientists ensure that their research is unbiased by AI companies’ agendas,
and that any potential conflicts of interest are declared in publications and other public commu-
nications (this also follows from Honesty and Transparency; cf. Mohamed Abdalla and Moustafa
Abdalla 2021; Atkin 2025; Forbes and Guest 2025; Knoester et al. 2025).

0133

Iris van Rooij 💭 @irisvanrooij.bsky.social · 07/09/2026

Somehow @olivia.science 's thread reminded me of a post that someone placed beneath an interview with me when I was appointed as full professor at my uni. I interpreted it as insinuating that I was not appointed because of my qualifications, but because I must have slept with someone in power.

Professional qualifications should be the sole determinant in selecting the best applicant for any academic position. Yet, regrettably, for many female applicants, family relations and amorous connections or potential for such seem to play a decisive role in directing appointment committees’ decisions. All one happy family!

6245

Iris van Rooij 💭 @irisvanrooij.bsky.social · 05/09/2026

Thank you! I really appreciate it. I was noticing the same over the years, which is why I added this note to our Cognition & Intractability textbook (van Rooij et al., 2019). Screenshot from Chapter 1, freely available here: www.cambridge.org/core/books/c...

1.4 FurtherReading
The Tractable Cognition thesis and its formalizations in the form of the P-Cognition thesis and the FPT-Cognition thesis were first coined by van Rooij in her PhD thesis in 2003. She built, however, on pioneering work of Edmonds (1965), Cobham (1965), and Frixione (2001). Edmonds and Cobham formulated a polynomial-time variant of the Church-Turing thesis, now known as the Cobham-Edmonds thesis. Frixione translated the Cobham-Edmonds thesis to the cognitive domain: He argued that tractability—conceived of as
polynomial-time computability—is a constraint that applies to computational-level theories of cognition in general. This thesis, proposed by Frixione, is what van Rooij coined the P-Cognition thesis. Prior to 2000 the P-Cognition was already tacitly entertained in several subdomains of cognitive science, for instance, in work by
• Cherniak (1986) and Levesque (1989) in the domain of reasoning
• Tsotsos (1990) in the domain of vision
• Simon (1988, 1990) and Martignon and Schmitt (1999) in the domain of
decision-making
• Thagard and Verbeurgt (1998) and Millgram (2000) in the domain of belief
fixation
• Oaksford and Chater (1993) and Oaksford (1998) in the domain of common-
sense
• Parberry (1997) in the domain of knowledge
Work exploring the FPT-Cognition thesis is much younger, given that it
was not conceived prior to 2003. The compendium in Appendix C gives an
overview of fixed-parameter (in)tractability analyses of computational-level
theories to date.

3354

Iris van Rooij 💭 @irisvanrooij.bsky.social · 05/09/2026

Also, some problems with larger (super-polynomial) search spaces may be easier to solve that other problems with smaller (super-poly) search spaces. My favourite example in slide: Minimum Spanning Tree (left) has a vastly greater search space than TSP (right), but former is in P & latter NP-hard.

0114

Iris van Rooij 💭 @irisvanrooij.bsky.social · 05/09/2026

Chess is not undecidable, but in generalized form (i.e, on an nxn board) it is EXPTIME-complete. See: en.wikipedia.org/wiki/Game_co...

Visual illustration of the (conjectured) relation between different complexity classes: P ⊆ NP ⊆ PSPACE = NSPACE ⊆ EXPTIME ⊆ EXPSPACE ⊆ decidable

2255

Iris van Rooij 💭 @irisvanrooij.bsky.social · 02/09/2026

Useful reminder, the 9 reply guys taxonomy. You're welcome 🙃

5387

Iris van Rooij 💭 @irisvanrooij.bsky.social · 01/09/2026

It isn't possible to deeply engage with modeling & not get why AI use is detrimental to modeling. As @olivia.science (2026) writes: "models are neither directly derived from theory nor from data, but through an effortful friction between the modeller and the medium" -- zenodo.org/records/2063... 🧵

1379

Iris van Rooij 💭 @irisvanrooij.bsky.social · 01/09/2026

The same paper by @olivia.science, @marentierra.bsky.social et al. (2026): zenodo.org/records/2008... also explains how "Knowledge production is ... safeguarded by (inter)national codes of conduct for research integrity (ALLEA 2023; KNAW et al. 2018)." 🧵

1348

Iris van Rooij 💭 @irisvanrooij.bsky.social · 01/09/2026

The citation in the slide in OP refers to a paper by @olivia.science, @marentierra.bsky.social et al. (2026): zenodo.org/records/2008..., which explains a.o. "why phrases like 'generative AI' impede scholarly discussion because by design these expressions are used to dazzle and sidestep scrutiny." 🧵

Figure 1. A cartoon set theoretic view on various terms (see Table 1) used when discussing the superset AI
(black outline, hatched background): LLMs are in orange; ANNs are in magenta; generative models are
in blue; and finally, chatbots are in green. Where these intersect, the colours reflect that, e.g. generative
adversarial network (GAN) and Boltzmann machine (BM) models are in the purple subset because they
are both generative and ANNs. In the case of proprietary closed source models, e.g. OpenAI’s ChatGPT
and Apple’s Siri, we cannot verify their implementation and so academics can only make educated guesses
(cf. Dingemanse 2025). Undefined terms used above: BERT (Devlin et al. 2019); AlexNet (Krizhevsky
et al. 2017); A.L.I.C.E. (Wallace 2009); ELIZA (Weizenbaum 1966); Jabberwacky (Twist 2003); linear
discriminant analysis (LDA); quadratic discriminant analysis (QDA).

DOI: https://zenodo.org/records/20082828

Table 1. Below some of the typical terminological disarray is untangled. Importantly, none of these terms
are orthogonal nor do they exclusively pick out the types of products we may wish to critique or proscribe.

DOI: https://zenodo.org/records/20082828

3318

Iris van Rooij 💭 @irisvanrooij.bsky.social · 01/09/2026

In case of use for others, sharing this slide in my intro class for a Bachelor course that I teach to AI students (and other interested students). More resources in thread 🧵

No (generative*) AI use allowed

Several (not mutually exclusive) reasons

AI use violates standards of scientific integrity.

AI use violates the examination and education rules (EER).

This course is a skills course. Your task is to learn to read critically, think critically, model critically all by yourself.

For more info see course manual.

---
*But see: Guest, O., Suarez, M., et al. (2026). Against the Uncritical Adoption of 'AI' Technologies in Academia. Digital Culture & Education.

319653

Iris van Rooij 💭 @irisvanrooij.bsky.social · 29/08/2026

💯 "the purpose of a system is what it does" -- Stafford Beer en.wikipedia.org/wiki/The_pur...

The purpose of a system is what it does

From Wikipedia, the free encyclopedia

The purpose of a system is what it does (POSIWID) is a heuristic in systems thinking coined by the British management consultant Stafford Beer,[1] who stated that there is "no point in claiming that the purpose of a system is to do what it constantly fails to do".[2] It is used in systems theory, and is generally invoked to ...

1358

Iris van Rooij 💭 @irisvanrooij.bsky.social · 29/08/2026

Postscript: This thread is brought to you by someone who just learnt that a full professor of AI declares in a Dutch newspaper that it is A-Okay to not read 30 papers, but instead fine to ask CLAUDE for a "summary". 🧵 archive.is/VYIKn

Figure. (left) a desk with stack of books, papers, and a computer. Above it the words "2 hours". (right) icon depicting text produced by Claude. Above it the words "5 minutes".

Caption: Two hours searching, or five mintes reading a summary?

613618

Iris van Rooij 💭 @irisvanrooij.bsky.social · 26/08/2026

It is OK :) I get you. It is true what you say. Know this one?

Man to woman at finer table: "Let me interrupt your expertise with my confidence"

2275

Iris van Rooij 💭 @irisvanrooij.bsky.social · 22/08/2026

Since you work in communication, you may also enjoy this research using a communication game, called Tacit Communication Game (TCG): * Blokpoel, M., et al. (2012). Recipient design in human communication: Simple heuristics or perspective taking? www.frontiersin.org/journals/hum...

Examples of the six different types of goal configurations. The difficulty of a game is determined by the combinations of tokens; the boards are ordered in increasing difficulty. In these examples the sender controls the blue token while the receiver controls the red token.

062

Iris van Rooij 💭 @irisvanrooij.bsky.social · 17/08/2026

On LiNkEdIn 🤪🫠 Sir, I teach AI students and what you are saying is nonsense

Allready an old trick, but it the wrong bias and a wrong starting point from this professor/lecturer.
Education has been changed.

Students have to use AI. In the old days (5 year ago) they had to read books and articles. Now it is easy to let AI explain it to them. But AI is not flawless yet. So, make different assignments.

Explain a base, a topic and mention a theory. Let them search for the theory and knowledge about the topic, let them peer review it, let AI explain the knowledge. Let them search for how to use the theory. Let them build on this knowledge with new insights and let them explain it to other students and teachers.

As lecturers we have to change our way of teaching and students have to learn differently. The base is the same: Understand the theory and know how to apply it.

I love to teach you how.

8486

Iris van Rooij 💭 @irisvanrooij.bsky.social · 16/08/2026

🤔 So how does all this help build immunity to hyped claims about “AI” capabilities? 💡It helps when we realize that "no computationally tractable account exists to date for substantive & domain-general cognitive capacities; e.g. reasoning, communication, decision-making, planning, analogizing" 25/🧵

AI systems cannot scaleFor computational models of cognition to be able to scale up from toy scenarios (such as studied in the psychological labs or such as form the bases of training AI systems through machine learning), these models should minimally be computationally tractable (van Rooij, 2008). That is, the input-output mappings hypothesized by the computational models should in principle be computable using a realistic amount of resources (Box 1). However, when existing computational models are mathematically analyzed for their resource demands it turns out that such models are either tractable but limited in their input domains, or they are domain-general but then the models are intractable (for a textbook introduction to this methodology, see van Rooij et al., 2019). Consequently, no computationally tractable account exists to date for substantive and domain-general cognitive capacities, such as reasoning, communication, decision-making, planning, analogizing, categorization, and concept formation (van Rooij et al., 2019). Moreover, there is no good reason to believe that such tractable accounts will be forthcoming via machine learning (van Rooij, Guest, et al., 2024) or otherwise (Rich et al., 2021). Hence, if someone claims their concrete AI system is a “theory” of human cognition, they are more likely than not overstating the scope and capacities of the system (van Rooij et al., 2019) and obfuscating the system’s limits, including the human in the loop needed to make such systems “work” (Guest & Martin, 2025).

12612

Iris van Rooij 💭 @irisvanrooij.bsky.social · 16/08/2026

Well-behaved approximations are constant-additive approximations. All other types of approximations become increasingly bad as input size scales up. It has been proven that no neighborhood-searchable intractable problems have such approximations. From: doi.org/10.1007/s112... 24/🧵

Lemma 2 If f is an NP-hard function that is neighborhood-searchable for a distance function d, then no polynomial-time algorithm can ϵ-approximate f (assuming P ̸= NP; see footnote 5).
The proof is by reductio ad absurdum. Assume P ̸= NP and assume that there is a polynomial-time algorithm H that ϵ-approximates. Then we can run H on i to create output o such that d(o, f (i)) ≤ ϵ. Consider a second algorithm A that takes as input the output of H, and determines for all o′ ∈ fc(i) with d(o,o′) ≤ ϵ whether
o′ = f (i). Because we know that d(o, f (i)) ≤ ϵ, at least one such o′
= f (i), and hence H and A together afford the exact computation of f (i). Since H is a polynomial time algorithm and, by clause 2 of Definition 1, A is a polynomial time algorithm, f(i) can be computed in polynomial time. However, as f is an NP-hard function, this
implies that P= NP, and is thus a contradiction.

173

Iris van Rooij 💭 @irisvanrooij.bsky.social · 16/08/2026

Fwiw, see screenshots from: 📝 van Rooij, I. (2008). The Tractable Cognition thesis. Cognitive Science, 32, 939-984. I also recommend checking out the other 3 versions of the 'cognition is not computation'-objection.

6.2. The cognition-is-not-computation objection

Perspective C: Computational complexity theory has nothing to offer cognitive science because cognition is not computation.

What is meant by this objection depends crucially on the meaning of the phrase “cognition is not computation.” I believe the phrase is associated with a multitude of meanings. In my reactions below, I distinguish between four possible versions of the objection:

Version 4: Complexity theory does not apply to cognition because computation is analtogether wrong way of thinking about cognition.

This last version of the cognition-is-not-computation objection represents the non-computationalist that is not persuaded by any of my reactions to Versions 1 through 3. In my opinion, even this researcher should appreciate the contribution that the Tractable Cognition thesis makes to cognitive science; namely, a non-computationalist can still recognize that tractability is a constraint on computational theories of cognition. Then, the Tractable Cognition thesis offers the non-computationalist a way of evaluating the success of his or her competition. If, in the long run, human cognition systematically defies tractable computational description,
then this can be taken as empirical support for the idea that computation is the wrong way of thinking about cognition.

162