|
|
||
AI-Resilient Science is not about using AI to invent theories, but about using AI to test them. The paper proposes executable theories, cross-model deductive validation, and a clear separation of human authorship from AI computation-turning large language models into logical stress-testers rather than truth oracles. | ||
Abstract
This article proposes a methodology for scientific inquiry in the era of Large Language Models (LLMs) that treats AI not as a source of truth or a generator of knowledge ex nihilo, but as a computational instrument for deductive reconstruction of theoretical systems. The methodtermed AI-Resilient Sciencerests on (i) axiomatic preparation of context (filters), (ii) cross-validation across independent models, and (iii) the construction of executable documents that function as an operational specification of a theory.
The approach formalizes hypothesis generation, improves reproducibility of theoretical inferences, and reframes LLM hallucinations from a defect into a diagnostic element of the scientific method. Its philosophical grounding is discussed through an Augustinian epistemology, according to which the mind does not create truth from nothing but rather actualizes and organizes what is given. As an application case, we briefly consider the Temporal Theory of the Universe (TTU), where AI is used as a strict logical opponent to test the internal coherence of an ontology in which time is treated as a substance.
The method provides a defensible account of AI-assisted authorship by explicitly separating human axiomatic responsibility from AI-based deductive reconstruction.
1. Introduction
Modern Large Language Models (LLMs) such as GPT, Gemini, and Claude are often treated either as universal oracles or as mere statistical simulators of knowledgeand therefore unsuitable for serious science. Both attitudes are methodologically nave. This work adopts a different stance: LLMs can be used as engines of constrained deduction, capable of unfolding complex logical structures provided that the underlying axioms, definitions, and inference constraints are made explicit and supplied in context.
We introduce AI-Resilient Science: a research methodology in which a theory is considered robust when its key consequences can be reconstructed by an independent model from a compact axiomatic specification. Within this approach, AI does not replace the researchers thinking; rather, it amplifies it by acting as a strict, yet ontologically blind, logician. Meaning, axioms, and interpretation remain with the human investigator.
A predictable criticism of AI-assisted texts is the conflation of fluency with truth: this was generated by a chatbot. AI-Resilient Science addresses this directly by treating LLMs as constrained deductive engines and by making the human author explicitly responsible for the axiom set, definitions, and interpretive commitments. In this framing, model errors are not lies but signals of underspecification or inconsistency in the specification.
2. Filters as an Axiomatic Foundation (In-Context Ontology)
The core of the method is context preparation, referred to here as filters. In machine-learning terms, this corresponds to an extended form of in-context learning: what is provided to the model is not merely a set of facts, but a structured domain ontology containing axioms, definitions, and permitted forms of inference.
Crucially, such filters need not be optimized for human readability. They may be optimized for machine interpretability, functioning as an operational specification of reasoning. In this sense, a filter is not simply text: it is an executable document that, once loaded into a models context window, unfolds into a constrained space of conclusions.
Example (TTU)
In work on the Temporal Theory of the Universe (TTU), three foundational documents are used to specify:
Together these documents form an executable theory: when loaded into an LLM, they are not merely read, but operationalized into a coherent set of deductive consequences.
3. Cross-Validation as AI-Based Peer Review
A central element of AI-Resilient Science is cross-validation across multiple models. Each LLM is a probabilistic approximator with its own training distribution, inductive biases, and generation style. Consequently, convergence and divergence between models carry methodologically relevant information.
Formally:
Thus AI is used as an adversarial reasoning systeman analogue of distributed peer reviewwhile the human remains the final arbiter and interpreter. Unlike traditional review, this process is repeatable, can be partially automated, and is transparently reproducible.
A revealing phenomenon is that some systems (e.g., Copilot) may refuse to process documents they themselves generated, treating them as potentially dangerous executable files. This behavior is not merely an implementation error, but a symptom: such texts are no longer perceived as passive narrative content, but as control-like specifications capable of altering the models mode of reasoning.
In other words, these documents function not as descriptions about a theory, but as operational instructions for reconstructing it. They encode axioms, constraints, and permissible inference paths in a form that can be directly enacted within a language models context.
A theory is called AI-Resilient if an independent language model, given its axiomatic specification alone, is able to reconstruct the same system of key conclusions without access to the original expository text.
This definition introduces a novel criterion of scientific reproducibility. What is reproduced is not the surface narrative of a paper, but the structure of thought that generates its conclusions.
Consequently, executable documents shift reproducibility from re-reading to re-running. They enable independent reconstruction, deductive stress-testing, and controlled extension of a theory across different models and contexts. In this sense, the unit of scientific communication is no longer limited to a static PDF narrative, but includes an operational specification that can be executed, compared, and challenged within AI-assisted reasoning systems.
The use of large language models in scientific work raises a legitimate and recurring question: who is the author, what exactly does the AI do, and where does epistemic responsibility lie?
AI-Resilient Science provides an explicit and non-ambiguous answer to this question by clearly separating authorship, computation, and validation.
In this methodology, authorship remains entirely human. The human researcher is responsible for the formulation of axioms, definitions, ontological commitments, interpretive claims, and empirical relevance. These elements constitute the semantic and epistemic core of the theory and are not delegated to the AI.
The role of the AI is strictly instrumental and deductive. Language models are used for:
AI does not introduce new ontological assumptions, does not assign meaning, and does not adjudicate truth with respect to reality. It operates as an ontologically blind logical engine constrained by the provided specification.
Validation, finally, is achieved through a dual mechanism:
This division of roles can be summarized as follows:
By making this separation explicit, AI-Resilient Science avoids the conflation of fluency with authorship and computation with understanding. Claims produced through this method are not generated by AI in the epistemic sense, but deductively reconstructed under human responsibility.
5. Philosophical Foundations: Augustinian Epistemology in a Modern Frame
AI-Resilient Science is rooted in a classical epistemological stance, particularly in the thought of St. Augustine. On an Augustinian view, the human mind does not create truth ex nihilothat is the prerogative of God. Rather, the mind connects, orders, and actualizes what is given.
This aligns closely with the mechanics of LLMs:
A vivid metaphor is borscht: the ingredients are not the dish, yet the dish does not arise without them. A new quality emerges through organization and transformation, not from emptiness.
Addendum: Popper and Kuhn
In Popperian terms, cross-model divergence functions as a generator of potential falsifiers, highlighting weak points in a theoretical structure. In Kuhnian terms, paradigms can be represented as executable documents, and their change can be tracked through systematic cross-deduction under competing axiom sets.
6. Application in Physics: Time as Substance (TTU)
The method was tested in the development of the Temporal Theory of the Universe (TTU), where time is treated not as a parameter or coordinate, but as an ontologically real, created substance.
Key propositions include:
In this context AI is not a source of validation of truth; it is an ideal logician used to test internal consistency, definitional clarity, and deductive completeness.
7. Step-by-Step Protocol of AI-Resilient Science
Limitations of the Method
AI-Resilient Science does not replace the experimental method, nor does it remove the dependence of a theory on the quality of its initial axioms. If the axioms are flawed, inconsistent, or misaligned with physical reality, AI may reconstruct only a coherent but ontologically false model. In this sense, AI can support internal coherence, but it cannot guarantee correspondence with nature.
Moreover, the method cannot substitute for empirical verification: it is a tool for deductive analysis, not a source of observational data. Its strength lies in structural stress-testing of theories, exposing hidden premises and logical gaps; final confirmation or refutation remains a matter of experiment and observation.
Accordingly, epistemic responsibility remains with the human author: the method audits coherence, not reality.
A further limitation follows directly from the deductive nature of the method. If the initial axiom set contradicts empirical observations, or encodes physically incorrect assumptions, an AI system will nevertheless reconstruct a logically coherent and internally consistent theoretical structure. In such cases, the method may succeed precisely in what it is designed to dopreserve coherencewhile simultaneously failing with respect to correspondence with physical reality. This highlights the non-negotiable role of empirical constraints: AI-Resilient Science can expose logical weaknesses, but it cannot correct false premises supplied at the axiomatic level.
AI-Resilient Science articulates a new mode of scientific production at the intersection of philosophy, mathematics, and engineering. It does not redefine the aims of science; rather, it restructures the means by which theoretical knowledge is constructed, tested, and reproduced in the presence of large language models.
Within this framework:
Crucially, AI-Resilient Science does not compete with the experimental method, nor does it attempt to automate discovery. Instead, it clarifies and strengthens the deductive phase of theory construction, making explicit what is often implicit in human reasoning.
In this sense, the approach marks the emergence of a new epistemic infrastructure for 21st-century science. Philosophy provides ontological and epistemological foundations, mathematics formalizes axioms and constraints, artificial intelligence performs scalable deductive verification, and empirical observation remains the final arbiter of correspondence with reality.
What is proposed here is therefore not playing with chatbots, but a disciplined and accountable methodology for working with AI as a logical instrument. Beyond amplifying deductive rigor, the method also functions as a protective filter against the growing informational noise of the AI era, explicitly separating coherent theoretical structures from fluent but unconstrained text generation. In this sense, AI-Resilient Science is not a theory of intelligence, but a theory of responsibility under conditions of automated deduction.
Key Terms
AI-Resilient Science A methodology in which a theory is considered robust and reproducible if its key conclusions can be deductively reconstructed by an independent language model from a provided axiom set.
In-Context Learning A models ability to use information provided within the current prompt/session without updating its internal weights.
Deductive Reconstruction The restoration of a system of consequences from axioms and inference constraints, executed by an LLM within a bounded context.
Executable Documents (Operational Specifications) Texts structured not only for human reading but also for machine interpretation as an axiomatic substrate defining ontology and inference rules.
Cross-Validation (Between Models) A consistency check in which identical axiomatics are fed to multiple LLMs; convergences and divergences serve as signals of coherence and weak points.
Ontological Blindness of AI The lack of direct empirical or intuitive access to reality; LLMs operate over probabilistic token relations within context.
Adversarial Deduction The use of multiple models (or prompting regimes) to generate and mutually challenge deductive chains, analogous to structured scientific debate.
Hallucinations as a Diagnostic Tool Treating uncontrolled model outputs as signals of underspecified axioms, missing constraints, or ambiguous definitions.
Augustinian Epistemology (in an AI context) The view that neither human nor artificial intellect creates truth from nothing, but organizes and actualizes what is given (experience, revelation, or prompt).
Popperian Falsifiability (in AI-Resilient Science) Cross-model divergences act as potential falsifiers or stress tests of theoretical coherence.
Kuhnian Paradigm as an Executable Document A paradigm interpreted as a formalizable system of axioms and methods that can be executed for comparison and critique using AI.
Coherence vs. Correspondence AI can support internal coherence; correspondence to empirical reality requires observation and experiment.
Context Window The token-limited working memory of an LLM, crucial for implementing in-context ontology for complex theories.
Emergent Deduction The ability of an LLM to derive consequences not explicitly present in training data but implicitly contained in the supplied axiom system.
Core References
Epistemology & Philosophy of Knowledge
Philosophy of Science
Large Language Models & In-Context Learning
Critique & Limitations of AI
Philosophy of Information & AI
Ontology of Time
Authors Works
Appendix A. Extended Annotated Review: Conceptual Sources of AI-Resilient Science
A.1 Augustine: Why Knowledge Is Never Created Ex Nihilo
Augustine articulates a fundamental constraint on any finite intellect: it does not create truth, but recognizes and actualizes it within an already-given order of being. This premise becomes methodological in AI-Resilient Science: the model is not a truth-oracle, but a mechanism for unfolding consequences inside a specified ontology. The boundary between generation and truth is therefore drawn at the level of axiomatics and interpretation, not fluency.
A.2 Polanyi: Tacit Knowledge and the Limits of Formalization
Polanyi shows that much of human knowledge is tacit, embodied, and context-ladenhence not fully formalizable. AI-Resilient Science does not deny this limit; it incorporates it. The model operates within the formalizable layer, while the human remains responsible for ontology, meaning, and empirical judgement.
A.3 PopperKuhnLakatos: From Falsification to Program Stability
Popper frames divergence as a resource: discrepancies between models can function as generators of potential falsifiers that reveal weak points in a theory. Kuhn allows paradigms to be treated as structured packages of assumptions; here they become executable documents. Lakatos provides a lens for separating a research programmes hard core (axioms) from its protective belt (auxiliary assumptions). AI-Resilient Science operationalizes these ideas as a repeatable computational workflow.
A.4 Transformers and In-Context Learning: Why Filters Work
Transformers enable context-sensitive generation; in-context learning demonstrates that models can temporarily adopt provided rules and ontologies without parameter updates. Filters work because they impose a constrained space of reasoningeffectively a set of a priori forms of inference instantiated in the prompt.
A.5 Hallucinations: From Defect to Diagnostic Signal
Where hallucinations differ across models under the same axiomatics, the theory is likely underspecified: definitions are ambiguous, constraints are missing, or the axiom set is inconsistent. Hallucinations are thus reinterpreted as a diagnostic indicator of weak points in the specification.
A.6 Floridi & Shanahan: AI Without Truth Access
Floridi and Shanahan provide philosophical legitimacy for treating LLMs as agents without intentionality and without direct access to truthuseful as logical engines but not as epistemic subjects. This supports the methods central separation of coherence (AI-checkable) from correspondence (empirically grounded).
A.7 Time Without Time: Why TTU Functions as a Counter-Paradigm
Rovelli and Barbour represent influential attempts to downplay or eliminate time as a fundamental entity. TTU, by contrast, treats time as causally primary. This makes TTU an especially stringent test-bed for AI-Resilient Science: the theory either holds under deductive stress or it breaks quickly when the axiom set is weak.
|
