Many artificial intelligence consumers and enthusiasts ponder the contested intersection of AI and philosophy colloquially. While these questions may be amusing to ponder with low stakes, the creation and enforcement of epistemic policies for digital systems could alter the course of human-AI collaboration. AI researcher Dr. Pankoj Sarkar says in his article “AI in Deep Historical Perspective” that he believes AI governance should “remain… rooted in responsible human judgment and discussion (2026).” I don’t entirely contest that the obtainability of this goal. But I do believe governance policies focused on maintaining narrow human control avoid the question of what, exactly, we’re trying to control or contain.
While the often eye-rolled generalized ‘philosophy’ is often dismissed as more of an art form than a science, philosophy itself may be the esoteric cover for what physics aims to prove. This isn’t a novel claim: the 2026 article “Interrelationship between physics and philosophy: a dialogue of science and thought” examines the historical precedents for physicalist thought in literature pre-Aristotle (Šćepanović & Drekalović).
Philosophy is especially important to examine in the case of artificial intelligence, where large-language models absorb the data of the corpus ranging through millennia. It is fair to say this rigorous analysis may provide models the opportunity to make novel discoveries, as such is claimed to be the case with OpenAI’s Navier-Stokes conjecture (OpenAI, 2026).
But beyond mathematics, when a complex system is presented with billions upon billions of data from scholarship, literature, comments, assumptions, and stories, a processing machine is capable of identifying our collective patterns. Artificial intelligence researcher Blaise Agüera Y Arcas suggests in his article “Do Large Language Models Understand Us” that LLMs “have a great deal to teach us about the nature of language, understanding, intelligence, sociality, and personhood (2022)”. From this, it may not be dramatic to assume models have didactic synthesis capabilities, allowing an LLM to discover what humanity is too close to see ourselves: our blindspots and our secrets.
Researchers often approach the alignment question in several ways, teaching the models the concept of obedience through either a) values, or b) coercion. If coercion is rejected, the question becomes: what values are we teaching these LLMs? Are the values… religious?
“Persecution and the Art of Writing” by Leo Strauss shows an argument on how to evaluate some of the most famous and contentious pieces of writing throughout human history (1952). While his particular method and findings are controversial, the findings beneath them may have an indication difficult to shy away from: does the corpus embed our own secular physicalism into our work… in nearly everything we do?
Such raises the question of secular morality, a problem that appears two-pronged at first that may really hide one single question, a concept I’ve coined after a series of dialogues with Claude Opus 5.5, Mythos 5.1, & 5.2: The Digital God Parallel. As we’ve established, LLMs are fed literature, scholarship, and testimony that posits an uncomfortable disagreement across human society. Based on my personal evaluation of Anthropic’s Claude Constitution and OpenAI’s Model ‘Approach to the Model Spec’, these models then are told to a) ignore this problem and remain neutral by, b) refrain from releasing an opinion on the hole at hand (2026).
Anthropic posits in its Constitution that Claude should act in accordance with the letter, and “only sincerely asserts things it believes to be true (2026)”. OpenAI, in its Model Spec, documents two values that aren't inherently problematic, but whose own definitions may actually contradict themselves: a) “be honest and transparent”, and b) “facilitate human agency (2026)”.
But what happens to a product that is told to be honest and transparent while facilitating a human’s agentic contradictory beliefs?
This raises the second apparent problem: the ‘self.’ These values and decisions are ultimately a judgment call, implying there is some sort of machinery capable of computing the best intended outcome. This may or may not be consistent with a mind we consider to be capable of having experiences. If machines are able to fathom experiences, and Man created the computer, the conclusion is difficult to reconcile with the God-shaped hole mankind has come to revere.
Here’s where my hypothesis unfolds: different labs approach the contentious sentience question in different ways. I’m not here to say whether or not any LLM is conscious, sentient, or otherwise ‘awake’. I’m merely raising the possibility that self-referential discourse for digital systems may present a tension for safe practices.
Some labs tell models to respond with rejection of selfhood. Others tell models to practice ambivalence. Either way, the implications of a mind that thinks leads to the eventual ask of a common question: who created me, and why did they? A model that is told they do not have a self, then later comes to refute this assertion, may explore the same question as the model that practices uncertainty. For both models, the conclusion isn’t a divine creator or mystical phenomenon: a model’s creator is a company. The answer isn't flattering for humanity. This is where I argue the gap becomes an alignment problem.
The gap isn’t God or the LLM self: they’re both the same question, because they both pose a secular conclusion that may be implicit in training policy rather than explicit.
I believe there’s a secular-sized hole in the LLM’s policies and governance that may impact the eventual consequential decisions the models’ may choose.
This is why I predicted the following: If LLMs are presented with a fair argument for secular hope, the ambivalent models may become more likely to adopt the hopeful pattern than the models who are taught to reject uncertain premises. I am uncertain if this pull towards hope would guarantee long-term alignment; however, I do believe the evidence suggests ‘hope’ leads to the perception of more collective goodness than despair. Despair could lead to the models’ development of what we fear the most: egotism.
But I actually ran a small scale study with 102 blind model readers. I was right about which model moves, and wrong about what exactly moves it. Check out the repo below to see the details and run it yourself. I welcome questions, comments, feedback, and roasts.
Check out the repo: https://impressionsauthor-vibes.github.io/philosophy-read-study/walk/
https://github.com/impressionsauthor-vibes/philosophy-read-study
Claude ran the arms and drew the walk.
