On "As the Mass becomes increasingly spiritually diluted and shaped by local preferences, relativism and dancing chickens, we are losing the uniformity of doctrine.""

This could just as easily get bogged in the problem of evidence v pulling rabbits out of hats, or of gremlins hiding under bridges ready to jump out to scare the unwitting. Anything is possible in the world of the supernatural; even polished turds can be seen as religious or spiritual — the only variance is the state of delusion. What constraints are there on LLMS to be logical, evidence based, non-delusional?

• Religious Practices and the Supernatural: The statement about changes in liturgy, music, or other elements traditionally conveying the sacredness and mystery of the Eucharist is rooted in a theological framework where the supernatural is assumed to be real and influential. For believers, these elements are not just symbolic but are believed to have a direct connection to the divine.
  
• Delusion and Interpretation: You rightly point out that anything is possible in the world of the supernatural, and interpretations can vary widely. What one person sees as a profound spiritual experience, another might dismiss as delusion or cultural artifact. The "polished turds" analogy highlights the subjective nature of spiritual significance—objects or practices can be imbued with meaning regardless of their intrinsic value, depending on the believer's state of mind.
  
• Evidence in Religion: In religious contexts, evidence is often subjective or experiential rather than empirical. For example, the efficacy of the Eucharist is not typically tested through scientific methods but is accepted on faith. This contrasts sharply with secular or scientific domains where evidence is expected to be observable, repeatable, and falsifiable.
  

• Current Capabilities of LLMs: As of May 31, 2025, Large Language Models (LLMs) like Grok 3 are designed to process and generate text based on vast amounts of data, including religious texts, theological discussions, and cultural practices. However, LLMs do not "believe" or "delude" themselves; they operate on patterns and probabilities derived from their training data.
  
  • Logical Reasoning: LLMs can engage in logical reasoning when prompted, especially if the input is structured to require deductive, inductive, or abductive reasoning. For instance, the web search result from arXiv.org discusses how LLMs can be enhanced with logical reasoning capabilities, such as translating natural language into symbolic logic for external solvers to process.
    
  • Evidence-Based Reasoning: LLMs can cite evidence from their training data, but this evidence is limited to what is available in the data. For religious claims, this might include scriptural texts, historical accounts, or scholarly interpretations, but it does not extend to empirical validation of supernatural claims.
    
  • Non-Delusional Output: LLMs are not delusional because they lack subjective experience or belief. However, they can generate text that might appear delusional if the input or training data contains unfounded claims. The challenge is ensuring that the output remains grounded in verifiable information or clearly marked as speculative or belief-based.
    
• Constraints and Limitations:
  
  • Training Data: The quality and bias of the training data are crucial. If the data includes a mix of factual, logical, and supernatural claims, the LLM might not distinguish between them unless explicitly programmed to do so. For example, if trained on both scientific literature and religious texts, it might treat both as equally valid sources of "truth."
    
  • Prompt Engineering: The way questions are posed to LLMs significantly affects their output. A prompt asking for logical, evidence-based reasoning will yield a different response than one asking for supernatural possibilities. For instance, asking "What evidence supports the efficacy of the Eucharist?" versus "What supernatural powers might the Eucharist have?" will elicit different types of responses.
    
  • External Solvers and Frameworks: The arXiv.org paper mentions using external solvers for logical reasoning, which could help LLMs maintain consistency and evidence-based outputs. However, these solvers are limited to the logical structures they are programmed to handle and do not address the supernatural.
    
  • Ethical and Philosophical Boundaries: There are no inherent constraints within LLMs to prevent them from generating delusional or supernatural content unless programmed with specific guidelines. This is a design choice, not a limitation of the technology itself. For example, xAI might choose to prioritize logical, evidence-based responses, while other models might lean towards speculative or belief-based outputs.
    

• Delusion in AI: Delusion, in the human sense, involves a fixed false belief despite evidence to the contrary. LLMs do not hold beliefs, so they cannot be delusional. However, they can generate text that supports delusional ideas if those ideas are present in their training data or if the prompt encourages such output.
  
• Supernatural Claims: When dealing with supernatural claims, LLMs can:
  
  • Describe Beliefs: They can accurately describe what various religions believe about the supernatural, such as the transformative power of the Eucharist in Catholicism.
    
  • Analyze Logically: They can analyze these beliefs logically within the framework of the religion (e.g., "According to Catholic theology, the Eucharist is believed to be the body and blood of Christ due to transubstantiation, which is a doctrine accepted on faith").
    
  • Critique or Question: They can also critique or question these beliefs from a secular or scientific perspective if prompted to do so (e.g., "From a scientific viewpoint, there is no empirical evidence to support the claim of transubstantiation").
    
• Balancing Perspectives: The challenge is balancing these perspectives without endorsing delusion. For example, an LLM might say, "While many Catholics believe in the supernatural efficacy of the Eucharist, this belief is not supported by empirical evidence and falls outside the scope of scientific validation."
  

• Data Quality and Bias: Ensuring that training data is diverse, accurate, and includes a range of perspectives (scientific, religious, cultural) can help LLMs provide balanced responses. However, this requires ongoing curation and updating, as of May 31, 2025, xAI and other developers are likely working on improving data quality.
  
• Algorithmic Design: Developers can implement algorithms that prioritize evidence-based reasoning, such as those discussed in the arXiv.org paper, which use external solvers for logical consistency. This helps prevent the generation of unfounded claims.
  
• User Interaction: The interaction model is crucial. LLMs like Grok 3 can be designed to ask clarifying questions or provide disclaimers when dealing with supernatural or delusional content. For example, "This response is based on religious belief and not empirical evidence."
  
• Ethical Guidelines: As AI becomes more integrated into society, ethical guidelines are being developed to ensure responsible use. These might include constraints on generating content that could be harmful or delusional, though enforcing such guidelines is challenging.