Epistemology: Truth-Grounding and the LLM Hallucination Critique

Epistemology is the philosophical study of knowledge—its nature, sources, and limits. While traditionally focused on human cognition, it has become the primary framework for evaluating the reliability of Large Language Models (LLMs). The central crisis of AI—**Hallucination**—is fundamentally an epistemological failure where a system produces "Belief" without "Truth" or "Justification."

1. The Tripartite Theory: Knowledge as JTB

The classical definition of knowledge, derived from Plato, is **Justified True Belief (JTB)**. For an LLM to "know" a fact $P$, three conditions must be met:

1. **Belief Condition:** The model's weights must assign high probability to $P$. (The model "asserts" $P$).

2. **Truth Condition:** $P$ must correspond to the external world (correspondence theory) or be logically consistent (coherence theory).

3. **Justification Condition:** There must be a reliable, traceable reason why $P$ is asserted (e.g., training data, RAG source, or logical deduction).

Concrete Example: The Hallucination Failure

An LLM asserts: *"The 2024 Olympic marathon was won by a runner named John Doe."*

- **Belief:** The model is 99% confident.

- **Truth:** In reality, no such person won. (Truth condition fails).

- **Justification:** The model "hallucinated" by merging unrelated training tokens. (Justification condition fails).

- **Verdict:** This is **not knowledge**; it is a "Gettier-like" failure where the internal coherence of the model's weights masks a lack of correspondence with reality.

2. Coherence vs. Correspondence in LLMs

LLMs primarily operate on a **Coherence Theory of Truth**. They generate text that is statistically consistent with their training data.

- **Problem:** A model can be perfectly coherent (producing fluent, logical-sounding text) while being entirely false.

- **Solution:** To achieve knowledge, we must shift toward **Correspondence Theory** via **RAG (Retrieval-Augmented Generation)**. RAG forces the model to ground its "Belief" in an external, verifiable "Justification" (a retrieved document).

3. The Critique of "Hallucination"

What the industry calls "hallucination," philosophers call **"Epistemic Irresponsibility."**

- **Global Skepticism:** The "Brain in a Vat" thought experiment is the ultimate analogy for an LLM. The model has no direct access to the physical world; it only sees a "vat" of text.

- **Truth-Grounding:** To bridge the gap, we use **Knowledge Graphs** (like the Wikantik KG) to provide structural "Foundationalism"—a bedrock of atomic facts that the model cannot override with probabilistic drift.

4. Epistemological Frameworks for AI Practitioners

Foundationalism (The Knowledge Graph)

Knowledge rests on basic, indubitable truths. In our system, the **Structural Spine** and **Canonical IDs** serve as the foundation. If the KG says `ID_123` is "Paris," the model's probabilistic output must be anchored to this constant.

Reliabilism (The Evaluation Loop)

Justification is defined by the **reliability** of the process. If a model uses a "Google Search" tool and cites three high-authority sources, its belief is *justified* because the process is reliable, even if the model doesn't "understand" the concept of truth.

The Problem of Induction

LLMs are inductive engines. They predict the future based on the past. Hume's critique reminds us that just because an LLM was correct 1,000 times does not mean it is logically "justified" in its 1,001st prediction.

Summary: From Probability to Knowledge

| Feature | LLM Probabilistic Output | True Knowledge (JTB) |

| :--- | :--- | :--- |

| **Source** | Statistical distribution | Grounded justification |

| **Goal** | High coherence (fluency) | Correspondence with reality |

| **Failure Mode** | Hallucination | Falsehood / Lack of Justification |

| **Anchor** | Entropy minimization | Verifiable evidence (RAG/KG) |