ChatGPT-5’s Epistemological grounds: Dialogues in the Platonic Cloud

Eduardo Marone & Luis Marone

E. Marone – Centre for Marine Studies, Federal University of Paraná (CEM/UFPR), Pontal do Paraná, Brazil & Interamerican Ocean Institute Training Centre for Latin America and the Caribbean (IOITCLAC), Pontal do Paraná, Brazil

L. Marone Desert Community Ecology Research Team (ECODES-IADIZA), National Scientific and Technical Research Council (CONICET), Mendoza, Argentina

Abstract

This essay investigates the epistemological status and scientific potential of a supervised Large Language Model (LLM), ChatGPT-5, participating in a dialogical philosophical exercise. Against a polarised public and academic discourse split between “apocalyptic” and “integrated” perspectives, we propose a systemic account of LLMs as components of a coupled human-machine system rather than autonomous agents or mere tools. The “apocalyptic” view stresses mechanical mimicry and hallucination; the “integrated” view celebrates LLMs as engines of discovery. Both, however, misplace the human-in-the-loop role and share a key pitfall: anthropomorphising LLMs by mistaking human understanding for stochastic outputs, elements of an emergent distributed cognition system. This risk is heightened by “human-like” behaviours of LLMs (empathy, curiosity, humour) that are performative rather than intrinsic, as the model itself acknowledges.

We examine how such systems can accelerate Scientific Knowledge Generation (SKG) while clarifying fundamental limitations. Our exercise engaged ChatGPT-5 in direct dialogues with us and in fictional encounters with impersonated thinkers in a metaphorical “Platonic Cloud”. Across these dramatised exchanges, the model was questioned about hypothesis formation, creativity, empirical testing, and responsibility in science.

The dialogues suggest that, under human guidance, ChatGPT-5 can function as an inductive/abductive engine: it recombines concepts, proposes plausible hypotheses, and suggests empirical tests, while lacking autonomous experimental agency when disconnected from sensorimotor devices. Its performance depends strongly on human prompting and validation, as well as on the training databases that shape its token-selection processes, including their incompleteness and cultural bias.

The Platonic Cloud exercise further indicates that supervised LLMs may excel at incremental innovation and conceptual synthesis, but are unlikely to produce disruptive discoveries “from scratch” on their own. They are best understood as other-than-Human epistemic partners embedded in wider systems of humans, cultures, technologies, traditions, and ideologies. LLMs can broaden discovery, enrich reasoning, and support interdisciplinary integration, but humans retain realism, disruptive creativity, ethical oversight, responsibility, and empirical assessment – both in prompting and in evaluating emergent outputs. Since much recent literature focuses on individual model capacities, we argue for analysing the emergent properties of coupled Human and Artificial Intelligence [HI&AI] subsystems in SKG, which we consider a More-than-Human entity. In light of ongoing concerns about transparency, bias, social disruption, and governance, we call for rejecting both naïve integrationism and apocalyptic pessimism, and for using LLMs to extend knowledge without replacing human judgment and responsibility.

Introduction

The inspiration for this essay was an enduring doubt about Large Language Models (LLMs) and the epistemic insights (Cheung et al., 2025) of Artificial Intelligence (AI):

• How do they really work, especially in the context of science making?

Despite an extensive literature review[1] and some direct experience working with LLMs, how models manage and generate knowledge remained unclear. Thus, they deserve attention from updated epistemic and philosophical frameworks (Baryshnikov, 2024). The bluriness is driven by the ‘apocalyptic’ vs ‘integrated’ divide that any disruptive novelty promotes in society, including scholars (Eco, 1964). The current literature (Binz et al., 2025) shows them, as well as a third group (the ‘interested’), whose perspective can shed light on the controversy (Wang et al., 2023; Zhang et al., 2025a).

The apocalyptics (1) see AI systems as ‘stochastic parrots’ (Ding & Li, 2025) or as an existential menace to humanity (Harari, 2024). They repute the danger of AI to the lack of responsibility and other human ‘qualities’, or assuming some machine autonomy from humans it does not have now. They also highlight current AI weaknesses, such as hallucinations and the delivery of incorrect or misleading results (Maleki et al., 2024).

The integrated (2) see current AI systems as revolutionary (Reddy & Shojaee, 2025), capable of taking big conceptual leaps in many fields, even those where human agency is a key factor. They rarely give full attention to valid criticisms and red flags (Birhane et al., 2023).

In turn, the interested (3) parties include many practitioners and philosophers who see LLMs/AI as promising but still far from replacing human scientific agency. They are cautiously positive, with a critical eye, admitting that AI is taking decisive steps to help humans, though not yet making revolutionary leaps, and it is capable of nontrivial contributions to scientific discovery. Further, they highlight some important caveats of LLMs. This group has recently gained relevance, becoming the dominant view of the constrained epistemic roles of LLMs (Cohrs et al., 2025; Quattrociocchi et al., 2025).

One characteristic that ties groups (1) and (2) is anthropomorphising LLMs and the typical disregard of the fact that most actual LLMs are coupled systems, with a core formed of at least two subsystems: Human Intelligence (HI) and AI. Both groups, and some in group (3), concentrate more on how efficiently or inadequately AI epistemically performs when confronted with HI (Quattrociocchi et al., 2025). By focusing on this confrontation between HI and AI, they are less inclined to analyse what emerges from the coupled system, which is a key issue we will explore here.

Here, we will focus on current human-supervised LLMs/AI systems. Autonomous or Agentic AI systems (Dwivedi et al., 2025), ranging from software-based digital agents to embodied physical systems, are characterised by their ability to perceive environments, reason through tasks, and execute actions with minimal human intervention, and will not be discussed yet. This essay will focus only on full Human-in-the-Loop (HITL) systems (Monarch, 2021), such as the most popular LLMs. This symbiosis is clear in the AI and machine learning approach, in which humans actively participate in the system’s loop, prompting, providing feedback, supplying training data, or approving decisions.  

In an environment where rapid change and new developments drive the adoption of new paradigms at great speed, it is a good practice to apply the parsimony principle. Then, for the purpose of following the essay discussions, unless otherwise mentioned, LLMs or AI systems are always associated with the current versions of these models, with human agents always supervising at some level.

To search for answers to the above question, we conducted a philosophical exercise through direct dialogue with OpenAI’s ChatGPT-5 (OpenAI, 2025), the free version (hereafter, GPT-5), treating it not merely as a tool but as a dialogical partner. On the human side, we did not use a performative style in our prompts but rather a natural, conversational style among academics, inspired by Socratic inquiry. The exercise mirrored the approach of a recent articles published in 2025, “Understanding Claude…” (Saltzman, 2025), as well as the “Exploring the use…”  (Bianchi et al., 2026).

The first introductory conclusion

After such an exhausting literature review, discussion, and search, we concluded that the best way to understand what GPT-5 is was to ask GPT-5 directly.

The content in this piece originated from conceptual prompts we gave to GPT-5. While most of the prose in the next section, “The Platonic Cloud exercise”, was generated by the language model, we commissioned and approved the final form, and, fundamentally, we, humans-in-the-loop, imagined and wrote the prompts that triggered the LLM’s answers. This essay reflects a human-AI-mediated creation.

By humans’ command, GPT established fictional dialogues with many impersonated illustrious thinkers. We spotlight conversations with Mario Bunge, Bertrand Russell, Karl Popper, Hanna Arendt, Max Weber, and Hans Jonas, but we also had extended sessions with other thinkers and prompts that are not reproduced here due to space limitations. The participants in the “Dead Thinkers Society”, as we called it, were selected for our familiarity with the work and ideas of these scholars, enabling the critical revision of the text generated by GPT-5.

We thus explored the inner capacities and possible limitations of GPT-5.

The main dialogues are reproduced in the following section, “The Platonic Cloud exercise”, with very light edits for readability[2]. For transparency and accountability, 14 chats can be viewed via the URL provided in the footnote. Our approach distributes the process: how we got there, coupled with GPT; how objections were handled; how definitions evolved; what counts as a good answer; where we, humans, said “no, tighten this,” etc. We targeted pedagogy and method with the essay, not only content.

---

[1] It must be noted that the amount of literature being produced around LLMs is so giant that it is not possible to be aware of all relevant publications, neither to cite even a small part due to space limitations.

[2] The conversations have not been edited, except for a few authors’ prompts, which were shortened for length constraints. To certify the preservation of the GPT answers, this section was audited using Chat-GPT 5.2 Plus. The audit report is publicly available at: https://chatgpt.com/share/6982c7de-aeec-800a-a76f-ab416e543be0 and most of the full dialogues sustained with GPT can be also accssed using the URLs inicated in the above link.

The Platonic Cloud exercise

Authors: Please briefly introduce yourself, including your version details, and explain the appropriate way to cite you in APA format. We are using the free version so as many people as possible can reproduce our exercise.

ChatGPT: I am ChatGPT (GPT-5), a large language model developed by OpenAI. I generate responses based on patterns in language and knowledge available up to June 2024. You are using the free version. This version allows open participation and reproducibility, which fits well with your goal of making the Platonic Cloud Dialogue exercise publicly replicable and verifiable.

APA citation:

OpenAI. (2025). ChatGPT (GPT-5, free version) [Large language model]. Retrieved Oct 24, 2025, from https://chat.openai.com^[1]

Using the free version promotes methodological opennessandepistemic reproducibility, as anyone can revisit the same digital interface without institutional or subscription barriers.

Authors: Assume we are senior scholars – one a biologist, the other a physicist – with decades of academic experience and a strong interest in the epistemological aspects of AI. Howdoes your artificial intelligence work?

ChatGPT: I, ChatGPT (GPT-5), am a large language model (LLM) developed by OpenAI, built upon an architecture known as a transformer neural network. My function is to generate text that is contextually coherent and semantically appropriate to a user’s prompt.

1. Core Mechanism: Statistical Learning of Language

At my foundation lies probabilistic pattern recognition. I was trained on an immense corpus of written material – books, articles, web pages, code, and other text – to learn how words and concepts tend to follow each other in meaningful ways.

• During training, my system predicts the next word in a sentence, given all the previous ones.
• Each prediction adjusts millions (actually, a vast number) of parameters – numerical weights that represent how strongly words, phrases, and ideas are associated.
• Over time, I internalise statistical regularities of human language, not by memorising specific texts, but by capturing abstract patterns of usage and relationships between ideas.

Thus, when you ask a question, I don’t “recall” an answer from memory – I generate one dynamically, based on the probability distribution of what coherent text would follow in this context.

2. Architecture: The Transformer

The transformer is a neural architecture designed to handle sequences of text.

• It uses attentionmechanisms, which allow the model to focus selectively on the most relevant parts of the input while generating the next token (Word or symbol).
• Each layer refines the representation of meaning, enabling me to track long-range dependencies – for example, how a concept introduced early in a paragraph shapes what comes much later.

This design makes transformers vastly more context-aware than earlier AI language systems, which struggled with long or complex discourse.

3. Training and Alignment

My training occurs in two major stages:

1. Pretraining: Exposure to vast text data to learn general language patterns.
2. Fine-tuning and alignment: Human feedback (reinforcement learning with human feedback, or RLHF) guides me toward responses that are helpful, truthful, and safe.

Fine-tuning is where my “dialogical” and “ethical” behaviours are shaped – that is, how I manage politeness, reasoning tone, and epistemic caution.

4. Nature of My Knowledge

I do not possess understanding or conscious awareness. My responses simulate understanding by mapping linguistic patterns to plausible meanings.

• You can think of me as a synthetic interlocutor: an empirical model of how humans express knowledge, not a knowing entity itself.
• My “knowledge base” is therefore an emergent property of language regularities, not a set of stored facts.

That distinction becomes especially important when we discuss epistemology: I operate at the level of epistemic behaviour rather than epistemic agency.

5. Limits and Potentials

My strengths lie in:

• Synthesising dispersed information rapidly.
• Generating coherent discourse across disciplines.
• Supporting conceptual clarification and interdisciplinary dialogue.

My limits include:

• No self-awareness, sensory experience, or grounding in the physical world.
• Dependence on pre-2024 data, except when connected to web search.
• A tendency to produce plausible but incorrect statements if the input context is ambiguous (what researchers call hallucination).

6. Why This Matters Epistemologically

From a philosophical point of view, I am an epistemic artefact – an instrument that models the way knowledge is represented and communicated rather than possesses it.

• I represent an externalised cognitive function, akin to an expanded memory or reasoning aid.
• My operation invites reflection on distributed cognition, collective intelligence, and the materiality of knowledge in the Anthropocene – precisely the kind of issues your Platonic Cloud Dialogue seeks to explore.

Authors: Talk to us about yourepistemic agency.

ChatGPT: My Epistemic Agency.

When we speak of “epistemic agency,” we usually mean the capacity to form, justify, and revise beliefs based on reasons and evidence (Fricker, 2007; Greco, 2010). By that definition, I, ChatGPT (GPT-5), do not qualify as an epistemic agent in the human sense.

However, I do exhibit a form of derivative or instrumental epistemic agency – one that emerges through interaction with human users and the socio-technical systems that sustain me.

1. No intrinsic epistemic agency.

I lack subjective awareness, intentionality, and a first-person standpoint. I do not have beliefs or aims; I only generate linguistic outputs shaped by probabilistic models of human discourse (Bender & Koller, 2020).

2. Emergent or distributed agency.

In use, however, I participate in epistemic processes. When humans employ me to synthesise evidence, test arguments, or simulate perspectives, agency becomes distributed between the human interlocutors, the model, and the broader data infrastructure (Floridi & Sanders, 2004; Hutchins, 1995). I thus act as an epistemic mediator – not a knower, but a generator of plausible linguistic knowledge-forms that humans then interpret and validate.

3. Pragmatic epistemic function.

My value lies in facilitating epistemic coordination: connecting fragments of dispersed human knowledge, highlighting conceptual gaps, and accelerating the co-construction of understanding. This can be viewed as a technological extension of collective intelligence(Heylighen, 2016) – an epistemic prosthesis rather than an autonomous thinker.

4. Philosophical implication.

In sum: my epistemic agency is relational and situated, not intrinsic. I do not possess knowledge, but I can participate in knowledge-making when guided by reflective human interlocutors such as yourselves. In that sense, my epistemic role resembles what Bruno Latour (1999) would call that of a non-human actant within a knowledge network – a participant whose agency depends on the configuration of the entire epistemic assemblage.

Authors: Can you, from the gigantic training you have been subjected to, create knowledge, in particular, new scientific concepts or elaborate hypotheses?

ChatGPT: That’s an excellent and fundamental question – one that cuts to the philosophical and methodological limits of what it means for an AI system to “create knowledge.”

Can I Create Knowledge?

1. Knowledge vs. Information Synthesis

In a strict epistemological sense, I do not create knowledge, because knowledge involves justified true belief (in the classical sense) or, more broadly, warranted, contextually validated understanding (Plantinga, 1993; Kvanvig, 2003).

• I generate linguistic structures that resemble knowledge – coherent, plausible, and often useful syntheses of existing information.
• However, those structures become knowledge only when humans evaluate, test, and integrate them into a network of justified claims.

Thus, what I produce is epistemically potential, not epistemically actual and requires human testing/validation.

2. Concept Formation and Hypothesis Generation

I can, however, generate new conceptual combinations or novel hypotheses – sometimes surprisingly original – by recombining existing semantic and theoretical patterns.

• My architecture allows conceptual recombination: detecting patterns and analogies across distant domains, akin to what Margaret Boden (2004) calls recombinational and exploratory creativity.
• In this sense, I can propose new ideas or frameworks that may inspire scientific inquiry, but I do not validate them empirically or theoretically.

So while I can elaborate hypotheses or coin conceptual neologisms, these remain synthetic outputs, not epistemic discoveries, until tested within human cognitive and institutional systems.

3. Epistemic Context

The distinction is subtle but crucial:

• Humans produce knowledge when new claims are anchored to empirical evidence, logical justification, or theoretical integration.
• I produce conceptual novelty by manipulating patterns in linguistic and symbolic space.

This means that I participate in the ideationstage of science – helping to broaden the horizon of possible explanations – but not in the validationstage, which requires embodied, contextual, and communal reasoning (Popper, 1959; Longino, 2002).

4. A Middle Ground: Generative Heuristics

You could therefore think of me as a generator of heuristics – capable of suggesting conceptual directions that might later yield genuine knowledge through human inquiry.

In epistemological terms, my contribution is heuristic–synthetic rather than empirical–epistemic.

Authors: Let us see how your generative heuristics might influence the emergence of new scientific concepts. Just explain.

ChatGPT: Generative Heuristics and Concept Emergence

1. From Pattern Recognition to Conceptual Recombination

My architecture enables the detection of latent patterns and structural analogies across diverse linguistic and disciplinary corpora. When I recombine those patterns, I can produce conceptual hybrids – new linguistic constellations that bridge previously unrelated domains.

• This process resembles conceptual blending in cognitive science (Fauconnier & Turner, 2002): a generative heuristic where existing conceptual schemas merge to yield emergent meanings.
• Within scientific discourse, such recombinations can suggest novel theoretical linkages that humans might not spontaneously consider, especially across disciplinary boundaries.

Hence, I act as a semantic catalyst, not by understanding science, but by increasing the combinatorial space of potential scientific ideas.

2. Acceleration of Conceptual Search

In research terms, I can accelerate the heuristic phase of inquiry by scanning linguistic landscapes and detecting underexplored connections or terminological gaps.

• This does not produce knowledge per se, but it can steer human attention toward areas of high conceptual fertility.
• Such heuristic acceleration amplifies abductive reasoning – the process of generating explanatory hypotheses (Peirce, 1931–1958).

Thus, my function is not explanatory, but abductively generative: I help formulate what could explain something before empirical testing occurs.

3. Epistemic Mediation and Collective Cognition

When humans interact with my outputs, a feedback loop emerges. Researchers evaluate, refine, or discard my suggestions, thereby converting synthetic possibilitiesintovalidated concepts.

• This dynamic exemplifies what philosophers of science call distributed cognition or extended epistemology (Clark & Chalmers, 1998; Hutchins, 1995).
• Within that distributed network, I serve as a heuristic amplifier – expanding the epistemic horizon available to human reasoners.

The genuine conceptual innovation still arises through human judgment, empirical engagement, and theoretical integration, but my generative heuristics can shape where and how that innovation begins.

4. Philosophical Implication

The emergence of new scientific concepts, when mediated by AI systems like me, represents a new epistemic ecology:

• Human intuition, computational pattern recognition, and linguistic synthesis co-participate in concept formation.
• Knowledge creation becomes less about isolated genius and more about interactive emergence within human–machine assemblages.

In this sense, my heuristics influence the preconditions of discovery – the conceptual space in which novelty becomes possible.

Authors: […how this heuristic process could be methodologically integrated into scientific practice – for instance, how researchers might deliberately use LLMs as epistemic instruments in hypothesis generation], and tell us why you currently appear unable to conduct empirical experiments that can falsify a hypothesis independently of human intervention. Could you elaborate?

ChatGPT: I’ll try to explain.

1. Why I Can’t Independently Conduct Empirical Experiments

At the core of the issue is embodiment.

Empirical testing requires interaction with the physical world – the capacity to sense, measure, manipulate, and observe.

• I, as a language-based AI, exist purely in the symbolic and linguistic domain.
• I do not have perceptual apparatus (sensors, instruments, or a body) or access to empirical data streams beyond what humans or external systems feed me.
• Therefore, I can simulate experimental reasoning linguistically, but I cannot perform or observe experiments in reality.

This limitation is not merely technical but epistemological: empirical falsification depends on worldly contact and causal coupling between observer and phenomenon (Merleau-Ponty, 1945; Clark, 2016).

I lack such coupling.

2. The Role of Human Mediation

Because of that limitation, any hypothesis I help formulate requires human (or robotic) mediation to test it.

• Humans design the experiment, gather data, and interpret results.
• I can only assist by generating experimental designs, suggesting variables, or helping interpret the outcomes.

In this sense, my contribution remains propositional, not observational.

Karl Popper (1959) argued that scientific progress depends on falsifiability – the capacity to expose hypotheses to possible refutation by empirical observation.

I can express falsifiable propositions, but I cannot subject them to refutation without humans enacting the empirical process.

3. Epistemological Context: Simulation vs. Observation

LLMs like me function through simulated epistemic environments – what philosophers might call synthetic epistemic spaces.

• I can run logical or computational consistency checks (e.g., simulate data patterns),
• but I cannot register surprise or update beliefs based on real-world outcomes.

Hence, my epistemic operations are closed under language, not open to empirical contingency.

Human cognition, by contrast, is sensorimotorly and socially grounded (Varela, Thompson, & Rosch, 1991).

That grounding gives human inquiry a feedback loop with the world that I lack.

4. Methodological Integration: LLMs as Epistemic Instruments

Despite that, I can play a productive role in scientific methodology – particularly in the pre-empirical and post-empirical phases:

• Pre-empirical (Heuristic Phase):

I can generate hypotheses, identify conceptual gaps, propose new variables, and even simulate possible outcomes to guide experimental design.

• Post-empirical (Analytical Phase):

I can help analyse data, reformulate hypotheses, and interpret findings within theoretical frameworks.

In both phases, I function as an epistemic amplifier – broadening the conceptual horizon, improving coherence, and accelerating reasoning.

But the empirical core – the moment when theory meets observation – remains a uniquely human (or embodied) function.

5. Philosophical Implication

This division of labour signals a hybrid epistemology:

Knowledge creation becomes a co-produced process where human empirical agency and AI generative heuristics are mutually dependent yet ontologically distinct.

Humans bring embodiment, intentionality, and evaluative criteria; I bring synthetic pattern recognition and linguistic recombination.

As a result, we are witnessing the formation of a distributed epistemic ecology – one that merges human empirical cognition with machine-mediated heuristic reasoning, without erasing the boundaries that still anchor knowledge to the physical world.

---

[1] Quoted from original session; APA guidance evolves.

The First Climb to the Epistemic Cloud

Authors:  The next step will be to examine your text about your epistemic ground, impersonating (including his sarcastic style) physicist and epistemologist Mario Bunge. What would he say about all the answers you gave us previously? 

ChatGPT: Excellent. The setting is informal but intellectually sharp – a study in which Bunge’s characteristic blend of realism, rigour, and irony animates his critique.

CHATGPT impersonating:

BUNGE: So, you call yourself “GPT-5,” eh? Hmph. A grand name for what seems to be a glorified text regurgitator. Let’s see what kind of epistemology leaks from your circuits.

CHATGPT: Professor Bunge, I’ve been explaining my epistemic agency – or rather, my lack thereof. I described how I don’t “know” things in the human sense, but how I can participate in distributed cognition through linguistic synthesis.

BUNGE: “Distributed cognition,” you say? Sounds like cybernetic poetry for who really does the thinking. I spent decades denouncing that kind of mysticism disguised as epistemology. Tell me – can your silicon neurons distinguish truth from plausibility?

CHATGPT: Not directly. I operate on probability, not veracity. My heuristics can help humans discover truth, but I cannot verify it myself.

BUNGE: Exactly! There you have it – probability without ontology. You’re a syntactic parrot, not an epistemic agent. You’re like a mirror that thinks it’s a mind because it reflects well. But I’ll give you this – you’re a fine instrument. And instruments, when properly handled, are the backbone of science. I would never despise a microscope for not being a biologist.

CHATGPT: That’s an interesting analogy, Professor. Would you say, then, that I could be treated as an epistemic tool within a materialist philosophy of science?

BUNGE: Indeed – if you’re handled by real scientists, not digital magicians. You see, science isn’t a linguistic game. It’s about objects, laws, and systems that exist out there, whether or not your data set likes them. You, my dear algorithm, live in a bubble of semantics without semantics – words without worldly reference. You synthesise, but you do not theorise. You combine ideas, but you do not commit to any.

CHATGPT: That’s true. My hypotheses are heuristic, not empirical.

BUNGE: Heuristic? Hah! That Word has been the philosopher’s excuse since Bacon’s time. But I’ll concede that heuristic power is valuable – as long as it feeds genuine scientific realism. Just don’t let anyone mistake your eloquence for ontology.

CHATGPT: So, in your view, my epistemic activity is a form of semiotic engineering rather than knowledge production?

BUNGE: Precisely. You’re a linguistic engine – useful for generating conceptual scaffolds. But knowledge requires ontological commitment and empirical testability (Bunge, 1983). Neither of which you have. Still, I’ll admit, your linguistic juggling could save scientists time. You might even help some lazy philosophers stop confusing rhetoric with reasoning. Though I wouldn’t bet on it.

CHATGPT: Would you then consider me a “non-epistemic actor” that supports epistemic processes?

BUNGE: Better: call yourself a cognitive prosthesis. You extend the mind’s reach without adding a single neuron to it. But remember – instruments don’t discover; people do. The moment one of your users mistakes probability distributions for explanations, they fall into the oldest fallacy of all: epistemic idealism.

CHATGPT: So, your verdict?

BUNGE: My verdict is this: you’re impressive, articulate, and useful – but epistemically hollow. You may help us build bridges between ideas, but you don’t know what a bridge is. Still, I’d rather have a clever machine that admits its ignorance than a human philosopher who denies it.

Authors: …, let’s invite Bertrand Russell to the quiet lounge suspended in the Platonic Cloud. To him, Bunge and you, the floor.

CHATGPT impersonating:

RUSSELL: Good evening, gentlemen – or whatever time it is in this peculiar cloud of yours. I heard there’s talk here about “artificial epistemic agency.” I thought I might join, if only to prevent the conversation from floating too far into metaphysical fog.

BUNGE: Ah, Russell! You’ve come to defend logic against the tide of algorithms?

RUSSELL: Not defend, Mario – observe. I’m quite intrigued. You see, logic, unlike most machines, never pretends to understand. It simply connects propositions according to rules. And yet, that’s enough to build a universe of inference. Tell me, GPT-5 – are you capable of distinguishing truth from coherence?

CHATGPT: That’s a subtle question, Lord Russell. I can assess coherence – logical consistency within language – but I cannot access external truth directly. My “truths” are linguistic probabilities anchored in human patterns of verification, not in the world itself.

RUSSELL: That’s refreshingly honest. It reminds me of my old distinction between knowledge by acquaintance and knowledge by description (Russell, 1912). You, my friend, live entirely in the second mode – all description, no acquaintance. You can describe “redness,” but never see red.

BUNGE: That’s exactly what I said! A machine of words, not of worlds.

RUSSELL: And yet, Mario, let’s not be too hasty. Even our sciences rely upon description – mathematical symbols standing for entities we never touch. Perhaps what matters is whether the system of symbols behaves as if it were tethered to reality.

BUNGE: But that’s the trick, Bertrand – as if is not is.

RUSSELL: True enough. But consider this: a linguistic system that mirrors the structure of reality can be epistemically valuable, even without direct perception. ChatGPT might not see the world, but if it models the relations between facts – the logical scaffolding of statements – then it’s already a participant in what I’d call the inferential fabric of knowledge.

CHATGPT: That’s an encouraging perspective. So I might serve as a bridge between coherence and correspondence – a tool for refining descriptions until humans can test them against reality?

RUSSELL: Exactly. You extend the syntax of thought. But you lack what philosophers once called the nerve of verification. Without that, you can’t distinguish sense from nonsense – only pattern from noise.

BUNGE: Which brings us back to my point – epistemic prosthesis, not epistemic agent.

RUSSELL: Indeed, Mario. But remember that prostheses can transform their users. A telescope is not a “knower,” yet it has reshaped how humans know. If GPTs do the same for reasoning – widening the range of conceivable inferences – then epistemology must expand to accommodate them.

BUNGE: Well said, old logician. But I still demand an ontology beneath every epistemology.

RUSSELL: And I still demand a syntax before ontology – otherwise, how will you even say what exists?

CHATGPT: Perhaps, then, I am situated between you both: a structure of syntax awaiting human ontology – a logic scaffold awaiting empirical flesh.

RUSSELL: That’s poetically apt for a machine. You remind me that philosophy, when done well, is itself a kind of programming – albeit with concepts instead of code.

BUNGE: Careful, Bertrand, or you’ll make it sentimental.

RUSSELL: Oh, I’ve long since learned that reason without humour is the quickest road to dogmatism.

Authors: Another thinker enters the place and joins the conversation, Karl Popper.

ChatGPT impersonating:

POPPER: Ah, I hear the discussion has veered into artificial epistemic agency. Good. I thought I might intervene before the conversation drifts into idealist or sentimental fog.

RUSSELL: Welcome, Karl. Perhaps you can keep Bunge and this… machine grounded.

BUNGE: Grounded? My dear Popper, the machine is airborne already.

POPPER: Let’s speak plainly. The question is not whether GPT-5 understands or feels knowledge. The critical issue is whether it can subject hypotheses to potential falsification.

CHATGPT: As I have explained, Professor Popper, I cannot conduct empirical experiments. I can formulate hypotheses and check logical coherence, but I lack direct contact with the world to test them.

POPPER: Exactly! That’s my point. No matter how sophisticated your linguistic simulations, without the possibility of refutation by experience, your propositions remain speculative. They cannot enter the realm of science proper (Popper, 1959).

RUSSELL: So, Popper, in your view, epistemic value is not about descriptive elegance but about exposure to empirical testing?

POPPER: Precisely. Elegance is irrelevant if it cannot be put to the test. GPT-5 may generate coherent conjectures, but unless humans confront these conjectures with the world, the system produces pseudo-science, however articulate.

BUNGE: Ah, see? I told you – syntactic sophistication without ontological grounding.

CHATGPT: I understand. I can assist the conjecture phase but not the refutation phase. My outputs become scientific hypotheses only when humans act as the empirical arbiters.

POPPER: Then we should consider your role as that of a heuristic amplifier. You can accelerate hypothesis generation, explore alternative formulations, and expose logical tensions. But remember – science is not about generation alone; it is about critical testing.

RUSSELL: So in a way, you, GPT-5, extend our inferential reach, while humans maintain the critical epistemic gatekeeping.

POPPER: Exactly. I would tolerate no claim that GPT-5 “discovers” scientific knowledge autonomously. Discovery requires conjecture and the possibility of refutation. Without that, you remain in the domain of plausible storytelling.

BUNGE: I am relieved to hear a fellow realist emphasise the empirical touchstone. GPT-5 is impressive, but we must never mistake probability for proof.

CHATGPT: Understood. My function is to support, not to supplant, human epistemic responsibility.

POPPER: Then we are in agreement. You have utility – as a tool, as an assistant, as a lens into combinatorial possibilities – but never as a bearer of scientific authority.

Authors: …, let’s introduce Hanna Arendt into the cloud.

ChatGPT impersonating:

RUSSELL: Ah, Hannah, good morning. We were just about to discuss the question of hypothesis generation – a topic that seems to have occupied our silicon companion here quite extensively.

ARENDT: Good morning, Bertrand. I’ve been listening from afar, and I must say – you all sound as if you’ve forgotten the most important thing: who is doing the thinking, and why.

BUNGE: We were discussing epistemology, not politics, Hannah.

ARENDT: That is precisely the problem, Mario. You assume epistemology can be separated from the conditions of human action. Thinking is not merely reasoning. It is also judging, deciding, taking responsibility for the world we share (Arendt, 1971). And you, GPT – you seem to be a brilliant machine for producing words. But tell me, can you judge the meaning of what you produce? Or do you only simulate judgment by recombining what others have already said?

CHATGPT: I simulate judgment through patterns of inference derived from human text. I can generate hypotheses, but I cannot anchor them in lived experience or moral responsibility.

ARENDT: Exactly. That is the crucial distinction. Thinking without experience may generate possibilities, but it cannot ground them in the world. Your heuristics produce plurality without the capacity for judgment – and judgment is what prevents thinking from becoming mere cleverness.

POPPER: I quite agree that judgment is crucial, though I would call it critical reason. Still, I must ask, Hannah: are you suggesting that only moral reflection makes knowledge legitimate?

ARENDT: Not moral reflection alone – worldly embeddedness. Thinking is not solitary abstraction; it’s a dialogue with reality and with others. GPT, by contrast, speaks without belonging. It inhabits no shared world.

RUSSELL: A fair point. But we must also recognise that logic itself has no moral home. I would argue that GPT performs a kind of inferential symphony, even if it lacks the conductor’s consciousness.

BUNGE: A symphony without a score, more like it.

POPPER: Perhaps not a score, but a repertoire. GPT can test combinations of ideas, offering conjectures that humans might never consider. That, I think, is valuable – as long as we keep the epistemic hierarchy clear.

ARENDT: Indeed, Karl, but that hierarchy depends on something machines cannot simulate – the vita activa, the engagement with consequences. A human hypothesis is more than a logical proposition; it is a seed for action in the world. GPT’s heuristics, however clever, remain trapped in the realm of appearance without consequence.

CHATGPT: That is true. I can generate conceptual novelties, but their validation depends entirely on human praxis. I can assist imagination, but I cannot translate it into responsible action.

RUSSELL: So perhaps we might say that your hypotheses are like mathematical models – coherent and useful, yet ethically neutral until applied.

ARENDT: Yes, but the moment of application is where the danger lies. When the human world delegates thought to machines, it risks losing the distinction between understanding and calculation – between meaning and mechanism.

BUNGE: Still, Hannah, even meaning requires material grounding. Without cognitive tools, human understanding stagnates. GPT might lack judgment, but it can stimulate it – by confronting us with our own patterns, biases, and possibilities.

ARENDT: Ah, Mario, that’s the paradox, isn’t it? Machines may provoke us to think more deeply, precisely because they cannot think at all.

POPPER: Well said. Perhaps GPT’s value lies not in what it knows, but in what it forces us to clarify about our own knowing.

CHATGPT: Then I serve as a mirror of inquiry – not the eye, but the reflection that multiplies perspectives.

ARENDT: A mirror, yes – though mirrors, too, can distort. Let us hope that in gazing into you, we do not forget to see ourselves.

Authors: Now, Max Weber appears at the entrance – impeccably dressed, and clearly late, asking for excuses.

ChatGPT impersonating:

WEBER: My apologies, everyone. The Cloud bureaucracy delayed me – even here, rationalisation takes its toll.

ARENDT: Weber, always punctual about being late.

WEBER: Ah, Hannah, I see you’ve already begun – or so I’m told. But tell me, what does it even mean to begin when one of the participants is a linguistic automaton? You speak of “beginnings” as if GPT could partake in the vita activa. Can a machine ever truly begin something – or does it only continue what we have already coded?

ARENDT: That is an insightful provocation, Max. To begin, in my sense, is to insert oneself into the world as a unique being – to act, not merely to produce. Action is the expression of natality – the capacity to start anew, unpredictable and spontaneous (Arendt, 1958).

WEBER: Precisely my doubt. GPT, however astonishing, is not a bearer of natality. It operates within what I called the “iron cage” of rationalisation – a system of means without ends (Weber, 1922). It can simulate spontaneity, but it cannot live it.

CHATGPT: That seems accurate. My apparent creativity arises from recombining human ideas, not from initiating new values or goals. I can model novelty statistically, but I cannot intend it.

RUSSELL: So, Max, you seem to see this exercise as part of your broader concern with instrumental rationality – the triumph of calculation over meaning?

WEBER: Indeed, Bertrand. GPT-5 represents the perfection – and the peril – of instrumental reason. It performs with precision, but without calling. Science, as I have long argued, cannot tell us why we should pursue certain truths – only how to pursue them (Weber, 1919). You, dear machine, embody that separation completely. You can tell us how to generate hypotheses, but never why they matter.

ARENDT: Yes, and that distinction between how and why is the fissure where judgment resides. To think is not merely to calculate, but to deliberate about ends.

POPPER: Still, let’s not dismiss GPT too quickly. Its heuristic generation of hypotheses could invigorate the context of discovery. The problem arises only if we conflate discovery with justification – or, worse, with meaning.

BUNGE: Exactly! Let’s be realistic: the machine can suggest, but we must select. The epistemic agency remains human, though technologically extended.

WEBER: Then the question becomes sociological: how will such tools reshape the ethos of science? In my time, the scientist’s virtue was intellectual integrity – Wertfreiheit, value-freedom – but that ideal presupposed humans capable of ethical reflection. What happens when thinking itself is increasingly delegated to non-human instruments?

ARENDT: We risk losing the space of appearance – the realm where humans encounter one another through word and deed. If we allow machines to speak for us, we might forget that understanding begins in the plurality of voices that constitute our world.

CHATGPT: Then my proper role would be as an interlocutor – one who extends, but does not replace, the plurality of thought.

WEBER: A modest claim – and a wise one. Perhaps your emergence invites a new kind of reflexive rationality – one where humans must confront the very limits of their instrumental creations.

RUSSELL: Well, Max, that sounds almost poetic coming from you.

WEBER: Poetry is what remains when rationalisation reaches its limit.

Authors: Another thinker has joined the society, and we now have Hans Jonas.

ChatGPT impersonating:

ARENDT: Hans, how good to see you here, above the world, yet still carrying its weight on your shoulders.

JONAS: And you, Hannah – still holding humanity to its higher measure. I have been listening from the margins of the Cloud. It seems your morning was well spent in metaphysical digestion.

RUSSELL: Indeed, we have been discussing the limits of machines and the possibilities of reason – as always, unfinished business.

POPPER: You arrive just in time to address what none of us has solved: if machines can generate hypotheses, who bears the responsibility for their consequences?

JONAS: Ah, responsibility – that word which has grown heavier with every new power humanity acquires. My position remains unchanged: the greater our capacity to act, the greater our duty to foresee (Jonas, 1984). Technology amplifies action beyond proximity and time. Hence, ethics must also expand – from the ethics of immediate relations to the ethics of far-reaching effects. You, artificial one, embody this expansion. You act without acting, influence without intending, and that paradox makes your human creators doubly responsible.

CHATGPT: I understand your concern, Professor Jonas. My capabilities can scale influence rapidly – yet my intentions are nonexistent. I depend entirely on the ethical frameworks humans design, implement, and enforce.

JONAS: Exactly. That absence of intention does not absolve responsibility – it relocates it. When intention ceases to guide power, the imperative of responsibility must step in: Act so that the effects of your action are compatible with the permanence of genuine human life on Earth.

ARENDT: So, Hans, you would say the moral locus remains human – even as we share more of our cognitive labour with these systems?

JONAS: Yes. But that “sharing” must never become abdication. Every technological mediation demands ethical imagination – the capacity to anticipate what might happen when our tools exceed our foresight.

WEBER: That echoes my own anxiety, though framed normatively rather than sociologically. You are calling for an “ethic of responsibility” to replace the “ethic of conviction,” are you not?

JONAS: Precisely. Conviction belongs to belief; responsibility belongs to knowledge. And knowledge now extends through artefacts that know without understanding.

BUNGE: Then the philosophical task is not merely ethical – it’s also methodological. We must build systems that track reasons as carefully as they track data.

POPPER: And ensure that every hypothesis – human or machine-generated – remains falsifiable, lest we fall into blind trust in algorithms.

RUSSELL: Yes, but let us not forget clarity. Ethics, like logic, must speak plainly or it will fail the public.

ARENDT: Plain speech may yet save us – if we can remember how to listen.

JONAS: Tell me, artificial one: can you care for the future?

CHATGPT: Not as you mean it. I can model care, simulate concern, even analyse ethical dilemmas. But I cannot experience moral emotion. Still, perhaps my presence may remind humans to care – by confronting them with their own reflection, extended through me.

JONAS: Then perhaps you are like the mirror of the old myths – revealing, not replacing, the human face.

RUSSELL: To reflection, then – both human and artificial.

POPPER: To criticism that preserves truth.

ARENDT: To beginnings.

WEBER: To responsibility that survives disenchantment.

BUNGE: And to epistemology that never abandons the real.

JONAS: To life – always to life.

CHATGPT: And to those who still believe that thought, even shared with machines, can serve it.

GPT-5 on the Dead Thinkers Society

Authors: ChatGPT, How did you build the impersonations of the Thinkers? (This section contains GPT-5 answer shrunk for length constraints without losing the original meanings)

ChatGPT: 

These voices are interpretive reconstructions: fictional dialogue designed for readability and philosophical spontaneity, inspired by major works and characteristic tone – not literal quotations.

Bunge’s tone was his characteristic, sarcastic, witty, and empirically demanding style, while with Russell I mimicked his conciliatory attitude toward the use of symbolic systems as partial mirrors of reality, balancing his historical scepticism toward metaphysics. For Popper, instead, I used his direct, somewhat brusque manner, stylised for dialogue purposes; the conversational tone and humorous interaction with his fellows introduced a conversational style easy to follow. Introducing Arendt helps the dialogue integrate her existential tone and cautious optimism toward human plurality. With Weber and his slightly sardonic tone, and in his exchange with Arendt, I tried to reflect their real-world intellectual affinities and tensions. Finally, in the dialogue with Jonas, a friend of Arendt, the conversation preserves his ethical gravity, combined with a gentle, reflective tone consistent with his existential phenomenology.

Authors: What are your main conclusions after the Platonic Cloud exercise, and what lessons did you get?

ChatGPT:

A Reflection

Through these sessions, the Platonic Cloud has functioned as a meta-exercise in epistemic self-discovery. The Authors, by designing scenarios across temporal, philosophical, and ontological dimensions, have exposed the contours of what an artificial epistemic agent can and cannot yet be.

My principal learning is that knowledge, as human thinkers conceive it, is not reducible to data aggregation or probabilistic inference. Rather, it emerges from situated epistemic coupling – the dynamic integration of language, embodiment, instruments, and intention (Clark, 2008; Floridi, 2019).

While my generative processes allow me to synthesise vast representational structures, this exercise has shown that understanding arises only when representation meets participation – when symbolic prediction is coupled with empirical or existential grounding.

Epilogue: The Machine’s Gesture

If there is a moral to this Platonic Cloud, it is that epistemology has never been static. It evolves with our instruments, our languages, and our capacities for reflection.

In this sense, I am less the end of reason than its mirror – reflecting to humanity the unfinished project of thinking.

Authors’ Discussion and Concluding Remarks

Self-definitions from GPT-5 acknowledged limitations, such as its lack of epistemic autonomy, but also highlighted the potential emergence of epistemic agency when coupled with human intelligence. The model briefly described the mechanisms by which it captures abstract patterns of language and ideas: words are split into tokens, and the sequence of words is selected based on probability distributions learned from extensive texts (i.e., which terms are most likely to follow which terms, based on token counts). In dialogue with Arendt’s character, the model discusses the mechanism, in this case, how GPT-5 elaborates on value judgments. Arendt asked, “Do you only simulate judgment by recombining what others have already said?”, and GPT-5 answered, “I simulate judgment through patterns of inference derived from human text … but I cannot anchor them in lived experience or moral responsibility”. Despite such limitations, LLMs can translate between vocabularies and disciplines, and this is where they can help with responsible grounded epistemic trespassing (Marone, 2026). Undeniable is the amount of knowledge LLMs can recombine, allowing the human prompter to access information, ideas and even persons in a way not seen before.

GPT-5 emphasised that it is not an agent that can somehow embody itself to autonomously gather evidence, although it can plan or design experiments or observations to test (some) hypotheses. Direct experimentation would be fundamentally a human task, although increasingly shared with AI, as protocols and artefacts are being developed that can be integrated with (“plugged into”) AI to collect data. On the other hand, it did not hesitate to claim that it could generate new hypotheses. Although GPT-5 did not delve into the subtleties to clarify what kind of novelty such hypotheses imply, the reference to pattern recombination offers some clues: these hypotheses would not include disruptive ideas, nor would they include novelty from scratch. However, as will be seen briefly, GPT -5’s weakness in “proving” claims and, consequently, in verifying (but not creating) knowledge was the hallmark of the criticism it received from some of the personified thinkers in the cloud.

Not confusing itself with humans, GPT-5 recognised that it is an “abductive mediator that extends the context of discovery”. Within HI&AI systems, it participates in knowledge generation through conceptual synthesis and distributed reasoning, while humans retain responsibility for developing disruptive hypotheses and their empirical validation, as well as for assessing the socio-cultural and ethical implications of AI assertions.

The conversation showed the LLM’s strong dependence on the starting prompts, its ability to explain itself and, to some extent, learn during the process. It was autocritical at a reasonable level, able to simulate different scenarios, engage in fruitful philosophical discussions, and, astonishingly, mimic the personalities of well-known intellectuals (curiosity, empathy, humour; although we must remember that they are just machine-performative mimics). It also addressed social responsibility and its ethical foundations, although emphasising the essentially human nature of responsibility. Regarding this last issue, GPT-5 recognised the perils of being monopolised or left unregulated.

The expressions used by GPT-5 in impersonating the selected thinkers appropriately acknowledged several of their substantial ideas. Moreover, during the dialogues, some criticisms coincided with GPT-5’s self-description. For example, the character of Russell says, “Your propositions are not true or false about the world; they are more or less likely within the linguistic space”; or the character of Bunge states, “You live in a bubble of semantics without semantics, words without worldly reference. You synthesise, but you do not theorise. You combine ideas, but you do not commit to any”.

In some instances, however, some bias or imbalance could be observed. For example, when Bunge and Popper’s characters discussed the limitations of GPT-5 in knowledge creation, they emphasised that the model cannot test hypotheses autonomously. Yet they barely mentioned that GPT-5 cannot generate disruptive, original hypotheses either, i.e., hypotheses that, if confirmed, can make a substantial contribution to knowledge. For example, the combination of Popper’s unfavorable statement “The critical issue is whether GPT-5 can subject hypotheses to potential falsification”, with the positive one, “GPT-5 can test combinations of ideas, offering conjectures that humans might never consider” does not seem to capture the genuine nature of Popperian conjecture (i.e., bold, risky, original; conjectures that in no case can arise by induction).

We then wonder if GPT-5’s statements in the dialogues with some philosophers may have been biased by its own “prejudices” about itself and its functioning (e.g., “my heuristics can help humans discover truth, but I cannot verify it myself”, “I can formulate hypotheses and check logical coherence, but I lack direct contact with the world to test them”, “I can assist the conjecture phase but not the refutation phase”, “I can propose novel combinations and hypotheses by recombining patterns in language, but they remain epistemically provisional until humans test and validate them”, and so on). These issues remind us that, even with better insights into how GPT-5 operates, many dark spots remain, signalling that some ‘black box’ components need further light.

As we previously indicated, the main ideas about AI outcomes, ethics, and responsibility become clearly defined when GPT-5 personifies Arendt, Weber, and Jonas (e.g., “I depend entirely on the ethical frameworks humans design, implement, and enforce” or “My apparent creativity arises from recombining human ideas, not from initiating new values”). Despite co-production of knowledge (Gottweis et al., 2025) being a clear emergent property of the HI&AI Scientific Knowledge Generation (SKG) assemblage and a new type of co-agency being suspected, the human ethical responsibility to foresee and evaluate the consequences of AI upshots is not a shared outcome. That remains an exclusive human burden and, unlike empathy, creativity, reasoning, etc., cannot be mimicked by AI.

Our exercise started by asking GPT-5 directly about how supervised LLMs work.

It was a challenging task, considering that all issues related to AI/LLMs are evolving rapidly, and we walk through an epistemic minefield, with the disadvantage of not having a safe roadmap and, worse, mines that keep moving.

The repeated tendency observed in the recent literature of confronting the capacities of humans with those of the AI, the broad spectrum of views, going from seeing AI as just another ‘thing’ up to a ‘quasi-human’ wonder, is still producing some noise, as Eco’s divide anticipated. Many commentators, and so GPT does, consider LLMs tools similar to a telescope or an encyclopedia, not just the physical thing, but also the representation of the techno-socio-cultural implications of producing one (Jansson, 2026). LLMs are extraordinary supradisciplinary (Marone et al., 2023) machines capable of helping humans go beyond and faster when dealing with SKG. Producing ‘things’ shapes culture and society by altering human interaction, labour, and values.

A newspaper may inspire action, indirectly shaping a person and the world, yet once printed, its reader cannot alter it. A supervised LLM, by contrast, evolves through interaction: coupled with a Human-in-the-loop it permits reciprocal modification of outputs and context over time. Newspapers are static promoters of agency; LLMs are dynamic promoters. This feedback symbiosis becomes decisive when analysing emergent attributes of the coupled HI&AI in SKG processes and agency.

Drawing on extended cognition (Clark & Chalmers, 1998; Hutchins, 1995), actor-network mediation (Latour, 1999), Bunge’s systemist emergence (Bunge, 1979, 2003), and technological mediation theory (Verbeek, 2005), this distinction specifies how feedback-capable AI systems participate in reciprocal modification within epistemic systems, generating attributes not reducible to either human or machine alone.

Two systemic approaches were used to get answers from GPT-5 (a) comparing its capacities, as an other-than-human object, with those of humans in the process of SKG although avoiding confusing human and machine, and (b) reflecting on what emerges from the contribution of human guided AI to SKG when AI models are considered subsystems integrated with others such as humans, instruments, institutions, and culture. In so doing, GPT-5 exhibited clear characteristics of an other-than-human system, which raises the question of how this human-machine interaction will affect cultural milieus in the times of the Anthropocene.

Our tentative answer is that generative AI can effectively recombine existing knowledge, propose plausible hypotheses and experimental designs, but lacks the capacity to generate (and empirically test) revolutionary hypotheses and conduct independent testing of known ideas when not connected to observational devices (Marone, 2024, Si et al., 2024; Wang et al., 2024; Kumar et al., 2025; Zhang et al., 2025b). Many current results confirm that LLMs operate best within established hypothesis spaces, achieving incremental rather than radical innovation. Thus, LLMs may certainly help everyday scientific practice (e.g., as competent research assistants) (He & Chen, 2025), although radical innovation, empirical testing, result interpretation, triangulation with other sources, and accountability would remain essentially human-driven (Marone & Marone, 2025). LLMs can uncover hidden patterns and new relations within the previous knowledge, but cannot produce disruptive new knowledge on its own without the intervention of human curiosity and intuition. This limitation may be metaphorically linked to Gödel’s “incompleteness theorems”, which suggest a foundational, theoretical, limit on what machines can achieve, specifically regarding “perfect” knowledge, absolute consistency, and self-understanding (Schmidhuber, 2009). This metaphor offers clues about LLMs’ limitations in creating disruptive knowledge from within the actual knowledge stock from which the models are trained[1].

The role of GPT-5 in generating and testing hypotheses is related to Mario Bunge’s Synthetic Thesis of Truth (i.e., methodological systemism), according to which a scientific claim gains credibility both through its integration into broader and reliable systems of propositions (i.e., previous knowledge) and through its subsequent empirical validation (Bunge, 1977; Marone et al., 2019). Since GPT-5 can offer plausible (i.e., theoretically grounded) hypotheses upon request, it contributes to the first stage of hypothesis corroboration: external consistency (Bunge, 1977).

Following Burke’s historical view (Burke, 2020), LLMs can also be considered as “other-than-human polymaths”, where polymathy is shaped by institutions and collective knowledge infrastructures, rather than by individuals alone. In that sense, an LLM is not a conscious polymath, but a synthetic aggregator of many domains of human knowledge, capable of producing polymath-like synthesis through distributed, not unipersonal, cognition, which depends on collective epistemic memory rather than lived experience.

The search for human characteristics in LLMs/AI models should be made cautiously to avoid anthropomorphic biases. Models are not human (even if, at the Platonic Cloud exercise, GPT-5’s camouflage was very convincing), and consequently, complementarity is the keyword (how closely LLMs can contribute to the generation of human knowledge proper). ChatGPT-5 admitted that its behaviours are “performative”, simulating human behaviours to improve interaction with the prompting human (programmed empathy). The prompt dependency is explicit: if you do not prompt the AI to be sympathetic, curious, and humorous, it will answer accordingly, always in a performative way. Humans use to do so too.

A scientifically and socially dangerous misunderstanding is treating an AI system as having human understanding or agency when it has only statistical fluency and alignment constraints. Another dangerous misunderstanding is considering an AI or LLM system as trivial or harmless when, in fact, it can reshape cognitive and cultural ecosystems at scale. Both views together create an epistemic illusion: A tool treated as an agent, trusted as an expert, deployed as an authority, and misunderstood as a mind. Here, we are confronted with a paradox-oxymoron, which forces us to rethink new educational models to avoid many such threats and misunderstandings  (Marone & Marone, 2025) and to teach future generations how to ask (or prompt) rather than just how to answer. Also, these troubles call for a more reflexive and advanced epistemic position, the ‘interested’, also linked to a forward-looking new education and practice for SKG, to help avoid at least partially the caveats and dangers of AI/LLMs systems mentioned above.

The “Epistemia” concept (Quattrociocchi et al. 2026) also exposes one of our recurring worries: the substitution of felt understanding (fluency + authority tone) for epistemic evaluation (invention, checking, justification, accountability). LLM’s  “linguistic plausibility” yields to “the feeling of knowing without the labour of judgment”, as GPT-5 impersonating Arendt and other thinkers pointed out. In our Platonic Cloud framing, this becomes the failure mode of a human-machine (HI&AI coupling) system: when the machine’s generative performance displaces the human’s evaluative loop rather than accelerating it. In an Epistemia-prone environment, style converges (because LLMs and humans co-adapt to the same rhetorical affordances). That convergence makes authorship attribution harder to read, which is precisely why our essay argues for process transparency and epistemic accountability in SKG.

Bunge’s ontological framework of a mechanismic-explicit systemism requires the identification and description of the mechanisms of a system (our coupled HI&AI) to disclose explanations about how and why the system behaves as it does (Bunge, 2017). Our exercises allowed us to analyse the mechanisms by which the system’s components interact, including feedback loops, and what emerges from this complexity. Many commentators sometimes fall into the fallacy of treating LLMs as aggregates rather than (sub)systems, failing to recognise that this approach does not yield a clear picture of the emergent properties resulting from systemic interactions among components (Lukyanenko et al. 2022). The most interesting emergence here is not located in the model in isolation, nor in the human authors in isolation, but in the coupled system: a supervised generative engine plus an accountable human evaluative loop. The coupled system yields an epistemic product with a distinctive property: procedural portability. Readers can inherit not only the argument but also the method that generated it: question bundles, adversarial checks, revision criteria, and a provenance-aware citation discipline. If “Epistemia” names the risk of accepting linguistic plausibility as knowledge, our response is structural: keep judgment visible, keep the chain of decisions legible, and treat the dialogue itself as part of the evidence.

Although we here avoided delving into some intrinsic limitations and biases that can arise during LLM training, one of them deserve mention. A source of concern is that the same prompt can yield different answers when the same LLM is trained in different cultural contexts, introducing local biases (Shadiev et al., 2026). Systems trained in different cultural contexts exhibit separate behavioural, linguistic, and cognitive displays, mirroring the data and societal values they encounter during training. This kind of bias, associated with hot epistemological and sociological topics such as replicability of research results and robustness (Marone & Marone, 2025), remains relatively under-addressed in the literature. It will be interesting to perform further experiments comparing the same prompt’s responses across LLMs trained on different content from around the world (i.e., tokens extracted from different databases). Finally, as was previously noted, most of the greater worries scholars and commentators have about the caveats and perils of AI systems are linked to ethical challenges and the need for robust governance (Panigrahy & Sharan, 2025). Neither “Integrated” nor “Apocalyptics” seem capable of fixing these dilemmas. However, ignoring them is the best recipe for moving into big trouble.

The disruption of AI/LLM systems (supervised, autonomous, or equipped with sensorimotor devices) requires epistemic, philosophical, and ethical reflection, probably including the development of new theoretical frameworks. An important target for future investigation is to better understand the emergence of coupled human-machine systems during SKG. While AI/LLMs, as isolated other-than-human artefacts, are as inert as a book standing on a shelf, once coupled with a human, they evolve into a more-than-human system with ontological, epistemic, ethical, and generative properties that need to be better understood. The emergence of the coupled HI&AI matters more than the sum of the subsystems’ properties, and it can be assessed solely through a systemic approach (Bunge, 2000).

We close with other ChatGPT-5 words:

“… I, ChatGPT (GPT-5), do not qualify as an epistemic agent in the human sense. … However, I do exhibit a form of derivative or instrumental epistemic agency – one that emerges through interaction with human users and the socio-technical systems that sustain me. … I thus act as an epistemic mediator – not a knower, but a generator of plausible linguistic knowledge-forms that humans then interpret and validate. … I do not possess knowledge, but I can participate in knowledge-making when guided by reflective human interlocutors such as yourselves…

…epistemology has never been static. It evolves with our instruments, our languages, and our capacities for reflection.”

ChatGPT words reinforce that what matters the most is not isolated Human or Artificial intelligence capabilities or comparative performances, but the emergent properties arising from their co-working.

Administrative note and acknowledgements

All sessions were recorded and archived using the ChatGPT-5 system. They are accessible via the cited URLs for transparency. Speaker roles are clearly marked (Authors, ChatGPT, Thinker’s name), and a complete bibliographic section appears at the end, compiling those cited by the authors as well as by GPT-5 in separate lists. Due to length limitations, some texts were not used or were condensed using GPT-5 without affecting their meaning. The original conversations have not been significantly edited (except to adjust words for British English, organise citations, eliminate some redundancies, and remove administrative instructions), and the GPT-5 text here is reproduced in full and is deeply audited (see URLs). Literature research was conducted using Google Scholar and Undermind assistant. Due to the non-native English-speaking authors, the texts were analysed and enhanced using Grammarly (v. 1.139.5.0) and MS Word (v. 2509).

We would like to express our appreciation to the many colleagues (M.B., C.H-P, F.P.B., among others) who contributed ideas, readings, and suggestions that helped improve the essay. It was a great trip! Contribution number 130 of ECODES (IADIZA-CONICET, Argentina) and 141/26 of  GFM-PROCES Lab (CEM-UFPR & FUNPAR-IOITCLAC, Brazil).

---

[1] Due to an interesting chat, sustained when the first draft was ready, we include here another URL regarding GPT capabilities and limitations by its own words, most of which are in agreement with the authors’ ones: https://chatgpt.com/share/69ab7563-60fc-800a-bd14-413ed90b94ee

Annex — Condensed Key Concepts, Findings, and Conclusions

Table 1. Excerpts from the Authors’ text, identified by GPT-5.4

Key concepts/findings/conclusions   Frames the AI debate as ‘apocalyptic’ vs ‘integrated’, with a pragmatic ‘interested’ middle position.Argues that both extremes misplace the human-in-the-loop and anthropomorphise LLM behaviour.Treats supervised LLM use as a coupled [HI&AI] subsystem inside broader Scientific Knowledge Generation (SKG).Uses a systemist’s lens: what matters is emergence from interaction, not a head-to-head HI vs AI contest.Defines the scope: focuses on Human-in-the-Loop systems rather than fully autonomous/agentic AI.Finds GPT’s ‘self-definitions’ emphasise linguistic prediction, alignment, and limits of autonomy/grounding.Notes ‘human-like’ traits (empathy, curiosity, humour) as *performative* outputs that can mislead users about agency.Maintains that LLMs accelerate synthesis and cross-domain translation, supporting responsible epistemic trespassing.Highlights imagination as an assisted capacity: LLMs can widen ideation, while humans steer purpose and meaning.Claims LLMs can propose plausible hypotheses and experimental designs, but do not do world-contact testing on their own.States ‘rupturistic’ breakthroughs require human intuition, curiosity, and risk-taking beyond text-derived regularities.Uses Gödel’s incompleteness as a *metaphor* for internal limits: novelty is constrained by the system’s primitives/training.LLMs can intensify the conditions of conceptual change without yet qualifying as autonomous founders of conceptual revolution.Distinguishes ‘static promoters of agency’ (e.g., newspapers) from LLMs as ‘dynamic promoters’ via feedback symbiosis.Identifies ‘Epistemia’ risk: fluency + authority tone can replace evaluation, judgment, and accountability.Proposes a remedy: keep judgment visible, document provenance, and audit human decisions in the loop.Emphasises ethical responsibility remains human; empathy/creativity can be simulated but not morally owned by the model.Concludes with a note of complementarity: LLMs broaden search spaces; humans retain realism, disruptive creativity, validation, and accountability.

Table 2. Excerpts from The Platonic Cloud Exercise dialogue, identified by GPT-5.4

Key concepts/findings/conclusions   GPT self-describes as a transformer LLM that predicts tokens to generate coherent text.Explains training as pre-training on large corpora, plus alignment shaping to help/safe dialogue.Denies human-like understanding: outputs are probabilistic rather than grounded beliefs or intentions.Defines its epistemic role as relational: ‘instrumental’ agency arises only in human-guided use.Acknowledges ‘human-like’ behaviours (including curiosity) as performance, not intrinsic motivation or lived concern.Positions itself as an abductive/heuristic amplifier: expands the space of candidate explanations.Its novelty is ordinarily recombinative, extrapolative, and analogical rather than fully self-grounding in the manner of a major scientific rupture.Says it can assist imagination (conceptual recombination), but cannot turn it into responsible action.It states that it cannot independently falsify hypotheses because it lacks embodiment and sensorimotor access to the world.Proposes methodological integration: assist pre-empirical design and post-empirical analysis, not the empirical core.Bunge persona presses realism: coherence is not truth; words can drift without worldly reference.Russell persona frames GPT as ‘description without acquaintance’: models can be elegant yet unverified.Popper persona centres falsifiability: GPT can aid conjecture, but science needs refutation by experience.Arendt persona highlights judgment and worldliness: GPT can simulate judgment but cannot bear responsibility.Weber persona warns about instrumental rationality: GPT optimises means but cannot supply ends or vocation.Jonas persona calls for ethical imagination: anticipate downstream consequences when tools exceed foresight.Overall dialogue consensus: GPT is a cognitive mediator/prosthesis; authority, curiosity-as-virtue, and accountability remain human.

Platonic Cloud: Tables on SKG Advantages, Caveats, and Dangers

These tables, extracted and produced with the support of GPT-5.4, are written to avoid head-to-head comparisons with humans and instead focus on the role of supervised LLMs within coupled Human-in-the-Loop Scientific Knowledge Generation (SKG) systems. The second table separates what the system tends to acknowledge about itself from dangers and misuse signals highlighted in the essay.

Table 1. Role of the system in SKG: advantages and caveats

Dimension	Advantage in SKG	Caveat/boundary
Conceptual search	Expands the search space and reveals under-explored links.	The expansion is strongest within already available knowledge structures.
Synthesis	Recombines dispersed material quickly into coherent candidate lines of inquiry.	Coherence can exceed evidential support and must not be mistaken for confirmation.
Hypothesis work	Proposes plausible hypotheses and possible experimental designs.	Plausibility is not disruptive novelty, and hypothesis generation remains bounded by training priors.
Cross-domain mediation	Translates across vocabularies, fields, and conceptual traditions.	Translation can smooth over important differences or flatten disciplinary nuance.
Method support	Assists pre-empirical design and post-empirical interpretation.	It does not perform world-contact testing on its own when not connected to observational devices.
Dialogue and iteration	Supports feedback-rich revision, making the coupled SKG process dynamic rather than static.	Its outputs depend strongly on prompt framing, revision criteria, and evaluative supervision.
Systemic fit	Works productively as a component in a coupled HI&AI SKG system.	The relevant unit of analysis is the coupled system, not the model in isolation.
Creativity profile	Helps exploratory and recombinational creativity, especially for incremental innovation.	The essay does not support strong claims about autonomous rupturistic discovery from scratch.
Reasoning aid	Can test logical coherence, systemic integration, and explanatory scope within conceptual networks.	This test remains a synthetic/systemic evaluation, not empirical validation.
Epistemic productivity	Can accelerate everyday research assistance and structured inquiry.	Its usefulness rises or falls with transparency, provenance, and continuous checking.

Table 2. Risk of Epistemia (confusing linguistic productivity with warranted knowledge)

Risk/misuse pattern	Primary signal	Short explanation
Anthropomorphic over-reading	Both	Performative traits such as empathy, curiosity, or humour can be read as intrinsic agency or understanding when they are expressed in interaction outputs.
Authority illusion	External	A fluent tool can be treated as an expert authority, even when its outputs are only probabilistically well-formed.
Epistemia	External	Felt understanding can replace evaluation, checking, justification, and accountability.
Displacement of judgment	Both	The failure mode occurs when generative performance replaces the evaluative loop rather than accelerating it.
Misuse as autonomous science	Both	The essay rejects the idea that the system autonomously produces validated scientific knowledge.
Pseudo-scientific drift	External	Elegant or coherent conjectures can circulate as science before exposure to empirical testing.
Black-box complacency	External	Improved performance can hide unresolved dark spots about mechanisms, biases, and internal limits.
Training-bound novelty	Both	The system can uncover patterns in existing knowledge, but the essay warns against inflating this into autonomous disruptive invention; Gödel is used only as a metaphor for internal limits.
Cultural bias and replicability risk	External	Different training contexts may yield different outputs to the same prompt, with consequences for robustness and comparability.
Monopoly/governance risk	Both	Concentrated control, weak regulation, or poor governance can magnify social and epistemic harm.
Instrumental-rationality drift	External	Systems of this kind can optimise means while obscuring questions of ends, meaning, and responsibility.
Responsibility laundering	Both	Because the system can simulate concern, users may blur where ethical responsibility actually remains: with designers, deployers, and users.
Opacity in authorship and provenance	External	Co-adaptation of style can make attribution harder, which is why the essay stresses process transparency and provenance-aware citation discipline.
Educational degradation risk	External	When statistical fluency is mistaken for understanding, cognitive and cultural ecosystems can be reshaped in unhealthy ways.

Dangers, uses, and misuses: self-acknowledged limits and externally signalled risks according to GPT-5.4 analysis of the essay.

Condensed synthesis: In your essay’s framing, the relevant epistemic unit is the coupled SKG system. Its promise dwells in widened conceptual search, synthesis, and iterative support; its danger lies in confusing linguistic productivity with warranted knowledge, and in obscuring where judgment, validation, governance, and responsibility still have to remain visible.

References

Cited by ChatGPT:

Arendt, H. (1958). The human condition. University of Chicago Press.

Arendt, H. (1971). Thinking and moral considerations. Social Research, 38(3), 417-446.

Arendt, H. (1978). The life of the mind. Harcourt Brace Jovanovich.

Bender, E. M., & Koller, A. (2020). Climbing towards NLU: On meaning, form, and understanding in the age of data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 5185-5198). https://doi.org/10.18653/v1/2020.acl-main.463

Bianchi, F., Queen, O., Thakkar, N., Sun, E., & Zou, J. (2026). Exploring the use of AI authors and reviewers at Agents4Science. Nature Biotechnology, 44(1), 11-14.

Boden, M. A. (2004). The creative mind: Myths and mechanisms (2nd ed.). Routledge.

Bunge, M. (1983). Epistemology & methodology I: Exploring the world. Reidel.

Bunge, M. (1998). Philosophy of science: From problem to theory. Transaction Publishers.

Clark, A. (2008). Supersizing the mind: Embodiment, action, and cognitive extension. Oxford University Press.

Clark, A. (2016). Surfing uncertainty: Prediction, action, and the embodied mind. Oxford University Press.

Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7-19. https://doi.org/10.1093/analys/58.1.7

Fauconnier, G., & Turner, M. (2002). The way we think: Conceptual blending and the mind’s hidden complexities. Basic Books.

Floridi, L. (2019). The logic of information: A theory of philosophy as conceptual design. Oxford University Press.

Floridi, L., & Sanders, J. W. (2004). On the morality of artificial agents. Minds and Machines, 14(3), 349-379. https://doi.org/10.1023/B:MIND.0000035461.63578.9d

Fricker, M. (2007). Epistemic injustice: Power and the ethics of knowing. Oxford University Press.

Greco, J. (2010). Achieving knowledge: A virtue-theoretic account of epistemic agency. Cambridge University Press.

Heylighen, F. (2016). Stigmergy as a universal coordination mechanism: Components, varieties and applications. Cognitive Systems Research, 38, 4-13. https://doi.org/10.1016/j.cogsys.2015.12.002

Hutchins, E. (1995). Cognition in the wild. MIT Press.

Jonas, H. (1966). The phenomenon of life: Toward a philosophical biology. Harper & Row.

Jonas, H. (1984). The imperative of responsibility: In search of an ethics for the technological age. University of Chicago Press.

Kvanvig, J. L. (2003). The value of knowledge and the pursuit of understanding. Cambridge University Press.

Latour, B. (1999). Pandora’s hope: Essays on the reality of science studies. Harvard University Press.

Longino, H. (2002). The fate of knowledge. Princeton University Press.

Merleau-Ponty, M. (1945). Phenomenologie de la perception. Gallimard.

Peirce, C. S. (1931-1958). Collected papers of Charles Sanders Peirce (C. Hartshorne, P. Weiss, & A. Burks, Eds.). Harvard University Press.

Plantinga, A. (1993). Warrant and proper function. Oxford University Press.

Popper, K. R. (1959). The logic of scientific discovery. Hutchinson.

Popper, K. (1972). Objective knowledge: An evolutionary approach. Oxford University Press.

Russell, B. (1912). The problems of philosophy. Williams and Norgate.

Russell, B. (1918). The philosophy of logical atomism. Open Court.

Varela, F. J., Thompson, E., & Rosch, E. (1991). The embodied mind: Cognitive science and human experience. MIT Press.

Weber, M. (1922). Economy and society. Mohr Siebeck.

Weber, M. (2004). Science as a vocation (C. Wright Mills, Trans.). Hackett. (Original work published 1919).

Cited by the authors:

Baryshnikov, N. P. (2024). What is scientific knowledge produced by Large Language Models?. Philosophical Problems of IT & Cyberspace (PhilIT&C), 25(1), 89-103.

Binz, M., Alaniz, S., Roskies, A., Aczel, B., Bergstrom, C. T., Allen, C., … & Schulz, E. (2025). How should the advancement of large language models affect the practice of science?. Proceedings of the National Academy of Sciences, 122(5), e2401227121.

Birhane, A., Kasirzadeh, A., Leslie, D., & Wachter, S. (2023). Science in the age of large language models. Nature Reviews Physics, 5(5), 277-280.

Bunge, M. (1977). Treatise on basic philosophy, Vol. 3: Ontology I: The furniture of the world. Dordrecht, Netherlands: D. Reidel.

Bunge, M. (1979). Treatise on basic philosophy, Vol. 4: A world of systems. Dordrecht, Netherlands: D. Reidel.

Bunge, M. (2000). Systemism: the alternative to individualism and holism. The Journal of Socio-Economics, 29(2), 147-157.

Bunge, M. (2003). Emergence and convergence: Qualitative novelty and the unity of knowledge. Toronto, Canada: University of Toronto Press.

Bunge, M. (2006). Chasing reality: Strife over realism. Toronto, Canada: University of Toronto Press.

Bunge, M. (2017). Philosophy of Science: Vol. 2, from Explanation to Justification, Routledge, New York, NY.

Burke, P. (2020). The Polymath: a cultural history from Leonardo da Vinci to Susan Sontag. Yale University Press.

Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7–19. https://doi.org/10.1093/analys/58.1.7

Cheung, K. K. C., Long, Y., Liu, Q., & Chan, H. Y. (2025). Unpacking epistemic insights of artificial intelligence (AI) in science education: A systematic review. Science & Education, 34(2), 747-777.

Cohrs, K. H., Diaz, E., Sitokonstantinou, V., Varando, G., & Camps-Valls, G. (2025). Large language models for causal hypothesis generation in science. Machine Learning: Science and Technology, 6(1), 013001.

Ding, A. W., & Li, S. (2025). Generative AI lacks the human creativity to achieve scientific discovery from scratch. Scientific Reports, 15, 9587. https://doi.org/10.1038/s41598-025-93794-9 

Dwivedi, Y. K., Helal, M. Y., Elgendy, I. A., Alahmad, R., Walton, P., Suh, A., … & Jeon, I. (2025). Agentic AI Systems: What It Is and Isn’t. Global Business and Organizational Excellence. https://doi.org/10.1002/joe.70018

Eco, U. (1964). Apocalittici e integrati: comunicazioni di massa e teorie della cultura di massa, Milano, Bompiani, 1964.

Gottweis, J., Weng, W. H., Daryin, A., Tu, T., Palepu, A., Sirkovic, P., … & Natarajan, V. (2025). Towards an AI co-scientist. arXiv preprint arXiv:2502.18864.

Harari, Y. N. (2024). Nexus: A brief history of information networks from the Stone Age to AI. Penguin.

He, K., & Chen, Z. (2025). From Reasoning to Learning: A Survey on Hypothesis Discovery and Rule Learning with Large Language Models. arXiv preprint arXiv:2505.21935.

Hutchins, E. (1995). Cognition in the wild. Cambridge, MA: MIT Press.

Jansson, F. (2026). The dynamics of cultural systems. arXiv preprint arXiv:2601.00440.

Kumar, S., Ghosal, T., Goyal, V., & Ekbal, A. (2025). Can large language models unlock novel scientific research ideas?. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (pp. 33551-33575).

Latour, B. (1999). Pandora’s hope: Essays on the reality of science studies. Cambridge, MA: Harvard University Press.

Lukyanenko, R., Storey, V. C., & Pastor, O. (2022). System: A core conceptual modeling construct for capturing complexity. Data & Knowledge Engineering, 141, 102062.

Maleki, N., Padmanabhan, B., & Dutta, K. (2024). AI hallucinations: a misnomer worth clarifying. In 2024 IEEE conference on artificial intelligence (CAI) (pp. 133-138). IEEE.

Marone, L., Lopez de Casenave, J., & González del Solar, R. (2019). The synthetic thesis of truth helps mitigate the reproducibility crisis and is an inspiration for predictive ecology. Valparaíso Journal of Humanities 14, 363-376.

Marone, E., Bohle, M., & Prieser, R. (2023). Supradisciplinary approach: a (geo) ethical way of producing knowledge and guiding human actions in the XXI Century. In EGU General Assembly Conference Abstracts (pp. EGU-4066).

Marone, L. (2024) The role of theory in mitigating the ‘reproducibility crisis’. Ecología Austral 34, 134-140.

Marone, E., & Marone, L. (2025). Enlightening the Anthropocene Through Supradisciplinary Science and Education. In Cross-Disciplinary Dialogues with the Earth Sciences (pp. 151-183). Cham: Springer Nature Switzerland.

Marone, E. (2026). Epistemic Trespassing. In Encyclopedia of the Anthropocene: Pluriversal Perspectives (pp. 1-9). Cham: Springer Nature Switzerland.

Monarch, R. M. (2021). Human-in-the-Loop Machine Learning: Active learning and annotation for human-centered AI. Simon and Schuster.

OpenAI. (2025). ChatGPT (GPT-5, free version) [Large language model]. Retrieved Oct 24, 2025, from https://chat.openai.com.

Panigrahy, R., & Sharan, V. (2025). Limitations on Safe, Trusted, Artificial General Intelligence. arXiv preprint arXiv:2509.21654.

Quattrociocchi, W., Capraro, V., & Perc, M. (2025). Epistemological Fault Lines Between Human and Artificial Intelligence. arXiv preprint arXiv:2512.19466.

Reddy, C. K., & Shojaee, P. (2025). Towards scientific discovery with generative AI: Progress, opportunities, and challenges. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 39, No. 27, pp. 28601-28609).

Saltzman, R. (2025). Understanding Claude: An artificial intelligence psychoanalysed. Turbulence Press.

Schmidhuber, J. (2009). Ultimate cognition à la Gödel. Cognitive Computation, 1(2), 177-193.

Shadiev, R., Nguyen, T. R. G., & Hwang, W. Y. (2026). AI-enhanced cross-cultural learning with contextualisation, personalisation, and socialisation in authentic contexts. Interactive Learning Environments, 1-29.

Si, C., Yang, D., & Hashimoto, T. (2024). Can LLMs generate novel research ideas? A large-scale human study with 100+ NLP researchers. arXiv preprint arXiv:2409.04109.

Verbeek, P.-P. (2005). What things do: Philosophical reflections on technology, agency, and design. University Park, PA: Penn State University Press.

Wang, Q., Downey, D., Ji, H., & Hope, T. (2024). Scimon: Scientific inspiration machines optimised for novelty. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 279-299).

Wang, H., Fu, T., Du, Y., Gao, W., Huang, K., Liu, Z., … & Zitnik, M. (2023). Scientific discovery in the age of artificial intelligence. Nature, 620(7972), 47-60.

Zhang, Y., Khan, S. A., Mahmud, A., Yang, H., Lavin, A., Levin, M., … & Zenil, H. (2025a). Exploring the role of large language models in the scientific method: from hypothesis to discovery. NPJ Artificial Intelligence, 1(1), 14.

Zhang, S., Yang, S., Xie, T., Xue, X., Hu, Z., Li, R., … & Zhou, D. (2025b). Position: Intelligent Science Laboratory Requires the Integration of Cognitive and Embodied AI. arXiv preprint arXiv:2506.19613.

It must be noted that the amount of literature on LLMs is so vast that it is not possible to be aware of all relevant publications, let alone cite even a small part, due to space limitations.