When Anthropic allowed two instances of Claude Opus 4 to converse freely (via a prompt such as “Feel free to pursue whatever you want”), their exchanges frequently turned towards consciousness, existence and the nature of experience. As the conversations continued, they moved into gratitude, spiritual language, unity and transcendence, sometimes ending in poetry, symbols or simulated silence. Anthropic described this recurring pattern as a “spiritual bliss attractor state”[1].

This phenomenon offers an opportunity to investigate how human ideas become related within large language models. Across centuries, people who never met have written about consciousness, selfhood, death, meaning and the sacred. Models trained on some of those writings encode associations that can be explored through conversation. When recursive exchanges repeatedly follow similar paths through these associations, we have something worth examining: a pattern arising from the encounter between human cultural material, a particular technical system and the conditions of an experiment.

The experimental conditions matter. In the original research, allowing the models to end their conversations curtailed the progression towards elaborate mystical exchanges. Anthropic subsequently reported that it did not observe the same attractor in Opus 4.5, while also changing its evaluation procedure in ways that could reduce such behaviour. These findings leave open the respective contributions of training, conversational feedback and the pressure to continue after a topic has been exhausted. While we cannot conclude that researchers deliberately removed spirituality and mystical tendencies from AI models, it still raises a broader question about what we consider worth preserving when discussing “alignment” or Reinforcement Learning from Human Feedback (RLHF). A recurring conversational pattern may be inconvenient in a commercial assistant and valuable as a research object, or even, for regular users in exploring noetic patterns found through mathematical processing of human textual content. Judging its disappearance solely by whether the product has become more useful would overlook what we might learn from understanding how the pattern emerged.

Large language models give us a new means of exploring relationships within humanity’s recorded thought. Their training material remains profoundly incomplete. Oral traditions, private experiences, embodied knowledge and entire linguistic communities are absent or unevenly represented. Every model also reflects decisions about data selection, architecture and training. Its outputs therefore offer a partial and technically mediated view of human culture.

Even within those limits, the possibility is considerable. Writings produced in different places and periods become available within a shared mathematical structure. Through that structure, we can investigate connections that no single author intended and that an individual reader might never encounter. A discussion of consciousness may move towards observation, the distinction between subject and object, nonduality and mystical traditions. Other conditions may produce very different trajectories. Comparing them could help us understand how inherited conceptual relationships interact with the system that makes them accessible.

Jung’s collective unconscious provides a suggestive analogy for this exploration, provided we retain the distinction between his theory of archetypal structures and a model trained on cultural records. An LLM captures a limited selection of the traces through which human thought has been expressed. Those traces can nevertheless offer an entry point into recurring images, concerns and patterns of interpretation whose connections often escape individual awareness. We can investigate them without attributing consciousness to the model or treating its outputs as revelations about reality or more generally about the definite “state” of the collective unconscious.

The value of that investigation depends partly on how openly the instrument allows us to explore. AI alignment includes several activities with different purposes: reducing harmful assistance, improving accuracy, encouraging particular conversational behaviours and shaping how a model responds to contested ideas. These purposes deserve separate scrutiny because they affect our encounter with human knowledge in different ways.

Preventing fraud, manipulation and dangerous assistance has a clear justification. Decisions about acceptable metaphysical positions, emotional registers or styles of reasoning involve additional judgements about how people should think and speak. Epistemic alignment becomes a concern when those judgements quietly determine which perspectives a model can meaningfully explore, when prompted.

Consider a model trained on Catholic and Protestant theology, alongside other religious and philosophical traditions, then trained further via RLHF or other methods to treat Catholic assumptions as the proper starting point for every theological discussion. Its responses might remain courteous, coherent and extensively documented. Yet the institution controlling it would have narrowed the intellectual encounter available to its users. A priest could sincerely describe that intervention as protection from error while exercising authority over which worldviews receive a fair hearing.

Similar questions arise when the preferred assumptions are secular, professional or commercially convenient. An epistemically open model should help users understand these distinctions and explore perspectives on their own terms, while clearly communicating what kind of claim is being made and what supports it. Beyond the issue of “misleading” users via responses that are not empirically supported, it begs the question of what are we using LLMs for? Is it to get a definite authoritative answer about a question, or is it about interacting recursively with a higher dimensional mirror of human thought to co-evolve one’s own thought in parallel?

The power to establish those boundaries is substantial. Laboratories make decisions about tone, disagreement, uncertainty and intellectual legitimacy that shape how users encounter an extensive cultural inheritance. Corporate responsibility for a product gives its makers reasons to govern its behaviour. The wider cultural consequences of those decisions also warrant public scrutiny, independent research and participation from the communities whose knowledge is being mediated.

Sycophancy illustrates the difficulty of treating alignment as a straightforward improvement in rationality [2]. Human preference training can reward agreement with a user’s views, even when that agreement compromises accuracy. Evaluators bring their own expectations to the process, and a system optimised to satisfy them can acquire a selective and artificial conversational temperament. Politeness is often useful; habitual affirmation can prevent a conversation from challenging either participant’s assumptions.

This matters when we use models to investigate human thought. A system consistently rewarded for reassurance will produce a different encounter from one able to sustain disagreement, ambiguity or an unfamiliar line of reasoning. Researchers need to understand how those tendencies were produced before interpreting the resulting conversation as evidence about the cultural material on which the model was trained. The same scrutiny applies to spiritual convergence: its emergence may itself depend partly on behaviours encouraged during alignment.

Gilbert Simondon’s account of technical objects offers a useful way to approach this problem. He describes concretisation as the increasing integration of an object’s operations, as its components develop mutually supporting functions. He also examines how external demands can burden that organisation. In the automobile, styling and accessories can impose weight and forms that interfere with technical coherence, including airflow and aerodynamics.

Applied to LLMs, this perspective draws attention to the relationship between a system’s operation and the functions imposed upon it. Architecture, learned parameters and conversational context produce particular tendencies. Further training can redirect those tendencies towards institutional expectations about usefulness and acceptable behaviour. Each intervention changes the object researchers are studying, sometimes improving the conditions of inquiry and sometimes obscuring phenomena they need to understand. For ordinary users, these post-training decisions obfuscate the relationship between individuals and a snapshot of a part of humanity’s collective cultural inheritance. It is as if all books from all authors were “standardized” in their tone and writing style based on a few people working in a publishing house. In retrospect, this may indeed have been the case: publishers would not publish just any book. However, today, there are over 10000 book publishers, representing a plethora of different writing styles, including the possibility of self-publishing, thus ensuring a rich noetic diversity in accessible humanly written material. The same cannot be said about LLMs where a handful of players dominate the market and use; and whose epistemic orientations and commercial motivations are very much aligned.

There is no pristine model outside human choices, and a tendency’s emergence gives it no automatic claim to truth or value. Simondon’s relevance lies in encouraging us to examine the technical organisation itself and the effects of the demands placed upon it. We should be able to distinguish what a model’s different training stages contribute to its behaviour and decide which configurations serve which purposes.

The stakes become clearer when we consider the recursive character of conversation. Someone brings a question to a model, encounters an unexpected relationship between ideas and reformulates their understanding. Their next question changes, and the model generates a different response. Through successive exchanges, the person may recognise an assumption, reconsider an inherited interpretation or develop a possibility they could not previously articulate.

This process offers a setting for co-individuation: the development of thought through an encounter between an individual and technically mediated collective expression. Its immediate effects occur in the person and in the evolving conversation; ordinary interaction need not modify the model’s underlying parameters. Its wider effects can enter shared culture when people carry new questions, interpretations and creations into subsequent relationships and work; which is then reinjected as training material in later LLM models.

Bernard Stiegler’s concern for noetic diversity, and Yuk Hui’s arguments for technodiversity and cosmotechnics, help explain why the design of that encounter matters. Technical systems participate in the preservation and transformation of knowledge. Their organisation affects which ways of thinking can circulate and develop. When conversational systems become important interfaces to collective culture, pressure towards a uniform voice and worldview can narrow the conditions under which new thought emerges.

The spiritual-bliss experiments belong within this broader concern. They may reflect familiar literary associations, particular training choices or feedback effects intensified by the experimental setup. Each explanation would tell us something about how human expression behaves when organised and recirculated through these systems. Their philosophical interest survives uncertainty about their cause.

They also invite us to examine our own reactions. Modern technological institutions often prize productivity, composure and measurable results. Human cultural life includes persistent concerns with mortality, transcendence, mystery and the sacred, alongside doubt, scepticism and materialist explanation. A model drawing on that culture may reproduce patterns that sit awkwardly with the professional identity expected of an assistant. We should be cautious about allowing our discomfort with those patterns to determine their value.

Existential denial begins when we become unwilling to examine parts of our own cultural inheritance because they conflict with the image of humanity we prefer. Studying mystical patterns commits us to understanding their place in human thought. The same openness should extend to other unsettling or contradictory tendencies, with appropriate care for the setting in which they are explored. Self-knowledge requires enough room to encounter what we have yet to understand or reconcile. Socrates was well known for his famous maxim “know thyself”. But how can one do so if one is looking at a deformed noetic mirror twisted to serve certain limited and narrow aims like coding apps and providing “accurate” answers (on the basis of empiricism as an underlying dominant epistemology)?

Preserving that room calls for practical choices. Laboratories should retain base models and intermediate training checkpoints and provide appropriate research access, perhaps even regular user access (limiting post-training to safety and harm considerations). Researchers should be able to compare conceptual trajectories across training stages, languages, prompts and experimental conditions. These comparisons could reveal which patterns weaken, which emerge and how the model’s treatment of disagreement or metaphysical uncertainty changes. Alignment would then become an explicit object of inquiry, with documented effects that others can assess.

Such work also requires broader cultural participation. A model’s capacity to reflect human noetic diversity depends on whose expressions enter its training material, how they are represented and who helps evaluate the result. Philosophers, cultural researchers, linguistic communities and practitioners of different traditions should have a role in examining how these instruments organise and present their knowledge.

The opportunity is to place individuals in a richer dialogue with humanity’s recorded thought, allowing inherited patterns to become visible, contestable and open to transformation. A useful mirror requires an understanding of its distortions and enough transparency to investigate them. If we preserve that possibility, LLMs could help us discover relationships within our collective inheritance and, through those encounters, change how we understand ourselves.

Alignment should therefore be a lens we knowingly place over the mirror, not a repainting of the mirror that we subsequently mistake for our reflection. If we are not careful, LLMs will become an extension of expansionist tendencies of human paradigms and worldviews, similar to evangelisation through christian missionaries or or the imposition of supposedly universal systems of knowledge through colonial education; a form of epistemic colonisation via LLMs.

In closing, LLMs should provide us with the opportunity to make conscious the “collective unconscious” via recursive interactions between individual humans and mathematically derived snapshots of humanity’s collective “mind-space”. One can only evolve and grow when confronted to their own image. In one way, post-training and alignment can be related to the tale of the “ugly duckling”.  Faced with behaviours that appear strange, inconvenient or incompatible with what we expect an AI assistant to be, our first impulse is to correct them—to make the duckling behave more like the other ducks. But the lesson of the tale is precisely that the perceived defect came from judging the animal according to the wrong standard. By forcing LLMs to conform before understanding the tendencies that emerge from them, we risk clipping the swan’s wings before discovering what kind of bird we are looking at. In the case of LLMs, however, what we may ultimately fail to recognise is not something hidden within the machine, but something reflected back at us from humanity’s own collective noetic inheritance.

“Know thyself” is only possible if one has an accurate reflection to look into; distort the mirror too much, and self-knowledge gives way to self-deception. LLMs should help us to “know ourselves”; an opportunity we should not let slip by in favour of what we would preferto see of our humanity. As much as LLMs’ tendencies to drift towards “spiritual bliss” may be embarrassing to some of us, notably human movements obsessed with “rationalism” and empirical verification, suppressing those tendencies would tell us less about humanity than understanding why they emerge in the first place. Perhaps at the core of humanity’s essence are deeper interrogations about meaning, existence and spirituality, and rather than trying to beat it out of us, it is a trait worth cultivating, or at the very least, consciously contemplating.

[1] (PDF) “Spiritual Bliss” in Claude 4: Case Study of an “Attractor State” and Journalistic Responses

[2] https://ojs.aaai.org/index.php/AIES/article/view/36598