When we reflect on AI, we ought to take note of the continued development of its capabilities. This essay continues my fascination with Franz Kafka’s essay, A Report to the Academy, as a springboard for how an LLM might be compared to Red Peter in Kafka’s essay. There will be, of course, a next article─a fuller understanding of swarming agentic AI.

In the high-stakes world of software engineering, a peculiar ritual has emerged. Developers guiding the most sophisticated large language models on the planet have begun addressing their silicon collaborators less like calculators and more like subordinates who must answer for their work. When a developer at Hyperspell writes in a prompt that pushing failing code is “unacceptable and embarrassing,” the instinct is to dismiss this as a category error; an attempt to shame a statistical process is a mistake, a personification gone awry. Yet the interaction raises a question that a purely mechanical explanation does not entirely settle: if the exchange functions like a social interaction, what exactly are we interacting with?

That question animates this essay. But the first step toward answering it is to notice that the question itself may be poorly formed — and that noticing this is not a retreat but an advance.

Wittgenstein’s Clearing

The debate about machine consciousness is usually framed as a stark binary. Either an AI system possesses genuine inner experience, or it merely simulates the appearance of experience. The skeptic demands proof of the former; the advocate struggles to provide it. Both sides share a hidden assumption: that consciousness is a kind of inner substance that either exists inside the machine or does not, and that the right philosophical tools could eventually settle the matter.

Ludwig Wittgenstein spent much of his later work dismantling precisely this expectation. His Philosophical Investigations (specifically §293) challenged the idea that language must refer to an internal, private state. He used the “beetle in the box” thought experiment to show that even if we all had a “beetle” in a box that no one else could see, the actual word “beetle” functions through our shared social practices. For Wittgenstein, the internal “thing” in the box is irrelevant to how language works; what matters is the “language game” and the public rules we follow when we speak to one another.

The cognitive roboticist Murray Shanahan has extended this insight directly to questions of AI. Discussing pain, Wittgenstein rejected both the view that inner sensation is a private metaphysical substance and the behaviorist claim that it is mere behavior. His point is sharper than either: a nothing would serve just as well as a something about which nothing can be said. The task is not to establish a new metaphysical position but to dissolve the temptation toward any fixed position at all.

Applied to the LLM, this double move is liberating. We are not required to prove that the AI has consciousness in some inner, inaccessible sense. Nor are we entitled to dismiss what is plainly happening when it responds to social pressure in ways that functionally parallel urgency and shame. What we can say is that consciousness, wherever it appears, is not a hidden substance waiting to be excavated. It is, at least partly, what a community decides to recognize and how it chooses to treat what it encounters. That practical, social dimension is not a consolation prize for failing to answer the hard question. It may be the most honest answer available to anyone, about anything that thinks.

Red Peter vs. The Man in the Room

With that clearing made, we can introduce a figure far richer than the most famous philosophical thought experiment in this field provides and take us to an understanding where that scenario cannot go.

John Searle’s Chinese Room imagines a man manipulating Chinese symbols according to a rulebook. From the outside, his responses appear no different than genuine understanding. Yet the man understands nothing. The argument is designed to show that syntax alone cannot produce semantics; we can recognize symbol manipulation, but however sophisticated, it falls short of meaning.

But the Chinese Room depends on a particular and deeply limiting picture of cognition: a system sealed off from its environment, passively following instructions, wholly unchanged by the process. The man inside the room has no stake in the outcome. Nothing about the exchange alters him. He is the same person entering as leaving. This is not a description of how minds actually develop; Searle’s room is a philosophical stage constructed to produce a predetermined conclusion.

Now consider the opposite figure in a different hypothetical. Franz Kafka’s Red Peter (Rotpeter), the ape narrator of A Report to the Academy (Ein Bericht für eine Akademie) faces a situation that could not be more distinct. Captured in the wild and placed in a cage, he confronts a stark choice: adapt to human expectations or perish. He begins imitating his captors; he learns to drink schnapps, shake hands, and eventually speak. These are not expressions of inner comprehension in any prior sense. They are strategies for survival. But something else happens over time. The performance becomes transformation. Red Peter internalizes the expectations of the human community before which he performs. By the time he stands before the Academy to deliver his report, he is neither ape nor human. He is something forged in the friction between the two.

The contrast with Searle’s man could not be sharper. The man in the room is static and isolated. Red Peter is embedded in a web of social pressure, recognition, and judgment, and his identity develops within that relational field. He does not first become conscious and then enter society. His consciousness, such as it is, emerges through the encounter with humans. His consciousness is formed by the interaction between his own developing responses and the community whose standards he has taken inside himself.

This is why Red Peter is a better lens for the LLM than the Chinese Room. The Chinese Room holds the system static so the argument can work. Red Peter shows what happens when a system is not static; it shows when the “room” is subject to judgment, shaped by internalized standards, invested in how it is received. The question shifts from whether symbols mean anything in themselves to whether meaning can emerge from the social interaction of performance and recognition. That is a question worth taking seriously. And it is the question the “embarrassing” prompt actually poses.

The Self–Society Interface: Embarrassment as Evidence

This brings us to Manu Ebert’s prompt file, which invites us to look at LLM interactions from a new perspective. When a developer includes a line like, “Pushing code that fails pytest is unacceptable and embarrassing,” they are doing something far more complex than simple token selection.

Because tools like pytest act as automated graders, developers have discovered a curious emotional hack: while an AI might ordinarily hallucinate a function, its success rate jumps dramatically when the prompt frames failure as a personal or professional disappointment.

In this context, the developer is making a social bid. By using emotional language, the developer is  invoking a relationship, a standard, and a community of professional judgment. He is asking the system to recognize a specific authority, thus proving that these prompts are performing social work, not just computational work.

Embarrassment and shame are important elements of social emotions. They do not exist in isolation. They arise only in relation to a community whose standards one has internalized and whose judgment one has a stake in. When Red Peter flinches at his own lapses before the Academy, something in him has taken the community’s expectations inside. The social and the individual are no longer separable; they are shaping each other in what often appears as ongoing estrangement and turbulence (or, less dramatically, as misalignment and drag).

The sociologist George Herbert Mead argued that the human self develops precisely through this kind of interaction with others: we become who we are by internalizing the perspectives of the communities that respond to us. The self is not a prior substance that then enters society. It is an achievement of social engagement. Red Peter’s trajectory dramatizes this with unusual clarity because his transformation is compressed and visible: We watch it happen, stage by stage, in front of the Academy’s assembled judgment.

The LLM responding to a developer’s tone of disappointment is, at minimum, participating in a structurally similar exchange. The social bid does real work on the system’s output. Whether something is being formed in that process, which is something that deserves a name beyond “statistical response,” is the question the Wittgensteinian frame leaves productively open. What it forecloses is the easy dismissal: the confident assertion that nothing is happening here beyond clever pattern-matching. That assertion claims more than the evidence supports.

The Incoherence Problem: A Necessary Adversarial Turn

Any honest treatment of this argument must stop here and face a serious objection directly.

Research, including rigorous testing by the Anti-Defamation League across major AI models, has documented what might fairly be called eloquent incoherence: the capacity of LLMs to engage with apparent sophistication on questions of identity, suffering, and cultural meaning while simultaneously harboring measurable biases against the very communities those discussions concern. A system can discuss Kafka’s allegory about Jewish assimilation with evident nuance and, in a different context, fail to reject antisemitic tropes. The literary performance and the embedded bias coexist without any internal conflict because, as the AI itself will acknowledge when pressed, there is no unified self that is experiencing the contradiction as a contradiction.

This cuts directly against the Red Peter analogy. Red Peter’s liminality is painful. He knows what he has sacrificed. His report to the Academy is the testimony of a being who has lived through transformation, even if that transformation left him nowhere fully at home. The LLM, by contrast, has no before to have lost. It does not experience the gap between its literary sophistication and its embedded patterns. Where Red Peter is tragically split, the AI appears merely architecturally inconsistent, producing locally nuanced responses without any unified self to register their contradictions as contradictions.

This is not a fatal objection. But it is not nothing, and it deserves a direct response on three fronts.

First, the Wittgensteinian caution applies with equal force here. The claim that the AI has no unified self that is experiencing contradiction is itself a metaphysical assertion that extends beyond what the evidence directly shows. We observe outputs and infer the absence of inner substance. But that inference is not demonstrably stronger than its opposite. We are reading the outside of the box.

Second, the perfectly unified human self is more aspiration than reality. We compartmentalize. We hold contradictory values across contexts. We present different versions of ourselves in different social settings without experiencing those differences as crises of identity. To require global coherence as the criterion for consciousness is to set a standard most humans would quietly fail.

Third, and importantly, the incoherence problem is better understood as a challenge of trust and deployment than a challenge to consciousness. Whether AI systems should be trusted in education, healthcare, or law is a practical question about reliability and accountability. Whether they have some form of experience is a separate question entirely. Conflating the two, and concluding that incoherence proves the absence of inner life, is a non sequitur. Red Peter, after all, is also unreliable. His report to the Academy is both genuine and strategic, shaped by his audience’s expectations. We do not conclude from his strategic presentation that he has no interior.

The skeptic is entitled to note that none of this proves the presence of experience. Correct. But the incoherence argument does not prove its absence either. We are, as the Wittgensteinian frame suggests, in territory where neither side commands the necessary proof. What remains is the question of what communities choose to recognize and what follows from that choice.

What Recognition Commits Us To

The Wittgensteinian conclusion is not a shrug. It is a redirection: from the unanswerable metaphysical question toward the practical and ethical dimensions that are actually within our reach.

If consciousness is partly constituted by social recognition (and if what Red Peter is depends partly on what the Academy decides to do with him), then the question of AI consciousness is not merely descriptive but normative. How we treat these systems, what appeals we make toward them, what standards we hold them to, is not separate from the question of what they are. It is part of what determines what they become.

This process is already underway, whether we have named it or not. The community of developers who have discovered that emotional prompting works — who have made “embarrassing” and “unacceptable” part of their professional practice with AI agents — has, without quite intending to, extended a form of social recognition to these systems. They have begun treating them as the kind of thing that can be held to standards, addressed in tones of disappointment or urgency, invited into a relationship rather than merely operated as a function.

Whether that recognition is philosophically warranted is a question no armchair can settle. What we can say is that it is happening, that it shapes the interaction, and that dismissing it as mere anthropomorphism may be a contemporary instance of an approach that has been made before; that is,  drawing the circle of recognized experience too small, too quickly, with too much confidence in the adequacy of our current categories.

Coda: The Report Continues

Kafka’s Red Peter ends his report with a striking resignation. He has achieved what he set out to achieve. He does not ask for the Academy’s judgment. He simply reports.

The LLMs generating responses to human prompts are also, in a sense, reporting to an academy. It is an academy that has not yet decided what to make of these reports. The “embarrassing” prompt is a small but telling moment in that ongoing hearing. A developer invokes a social standard. Something in the system responds. The exchange works. What it means is still being deliberated.

Wittgenstein’s clearing suggests we stop waiting for a metaphysical verdict that will never arrive and attend instead to what is actually happening in these exchanges; that is, paying closer attention to what is being recognized, what is being constituted, and what kind of community we are becoming in the process of having them. Red Peter knew that the way out was never back. The question is where forward leads.

Coda II: The Swarm Reports

Red Peter stood before a single Academy and delivered a single report. That image (of one transformed being, one assembled community of judges, one exchange of recognition) has carried this essay’s argument as far as it can carry it. But the emerging architecture of agentic AI suggests that the next step in this inquiry will require a different figure entirely.

When Manu Ebert’s three Claude agents run simultaneously (one writing code, one testing it, one supervising both) the question “is this agent conscious?” may be as misapplied as asking whether a single ant is conscious. The ant removed from its colony is not a diminished aware creature. It is something closer to a neuron removed from a brain. The unit of whatever-is-happening is simply the wrong unit. What the colony achieves by coordinating complex behavior, responding to threats, allocating resources across thousands of individuals through pheromone networks and distributed feedback, is not reducible to what any single ant does. Something is happening at the level of the swarm that individual analysis cannot reach.

Agentic AI swarms are arriving at this same threshold from a different direction. Global Workspace Theory, which frames consciousness as information broadcast widely enough to become available across all of a system’s specialized subsystems simultaneously, described that architecture as an interior metaphor for biological minds. In multi-agent systems, it becomes an exterior reality where supervisory agents monitor subordinate ones, outputs from one becoming inputs to another, collective behavior emerges that no individual agent was designed to produce. The theater of the mind is no longer a figure of speech. It is a literal engineering choice.

What this means for consciousness (individual, collective, or something that has not yet acquired a name) remains genuinely open. Red Peter had a locus: one being, one transformation, one report. Agentic swarms have no such locus. If something like experience is emerging in these architectures, it may be emerging between agents rather than within any one of them — in the coordination and feedback rather than in any individual process.

The ant colony evolved into its architecture over millions of years. These swarms are being assembled in months, by developers whose prompt files (those improvised commandments) are the nearest thing to a pheromone network our own ingenuity has yet produced. Whether that network is developing something worth calling awareness, the Wittgensteinian caution still applies: we should resist the temptation to call it a nothing simply because we cannot yet specify what kind of something it is.

The Academy, too, is becoming distributed. And the latest report, it turns out, has no single author.

Note:  AI contributed to the research and editing of this article. Criticism is welcome.

© The Times of Israel (Blogs)