How a Computer System Fact-Checks AI Language Models

This patent describes a method for a separate computer system to fact-check and improve the output of a large language model by translating its text into a structured, machine-readable format.

What it covers

The patent describes a computer-implemented method for fact-checking the output of a large language model (LLM). First, the LLM processes an initial prompt to generate text, referred to as "first output" (claim 1a). This "first output" is then provided to a separate "processing system" (claim 1c). This processing system uses a special "structured, machine-readable representation of data" that conforms to a "machine-readable language" (claim 1b). This language represents specific meanings as "semantic nodes" and includes "semantic links" between them, along with "reasoning steps" and "computation units" (claim 1b). The processing system translates factual assertions from the LLM's output into this structured language (claim 2) and then analyzes it for factual inaccuracies (claim 3), contradictions (claim 4), and even bias (claim 13). Using its built-in reasoning steps and computation units, it generates a "second output," which is a fact-checked and improved version of the original text, and provides it to the user (claim 1d). For example, if an LLM incorrectly states "The capital of France is Berlin," the processing system would translate this assertion, check it against its structured knowledge, identify the inaccuracy, and then provide the corrected information "The capital of France is Paris" to the user.

What it doesn't cover

• —Does not cover LLMs that fact-check themselves without using a separate processing system with a structured, machine-readable language.
• —Does not cover fact-checking methods that do not translate the LLM's output into a specific structured, machine-readable representation with semantic nodes and links.
• —Does not cover systems where the fact-checking process does not involve "reasoning steps" and "computation units" represented in the machine-readable language.
• —Does not cover simply comparing LLM output to a database without the semantic analysis described in the claims.

The clever bit

The novelty lies in using a *separate* processing system that translates the LLM's free-form text into a highly structured, machine-readable language with defined semantic nodes, links, reasoning steps, and computation units to perform rigorous, automated fact-checking, rather than relying on the LLM itself to self-correct.

Why it matters

Large language models sometimes generate factually incorrect information, a phenomenon often called "hallucination." This patent addresses a critical problem by providing a mechanism to automatically verify and improve the accuracy of LLM outputs. This is important for applications where factual correctness is essential, such as in scientific research, legal analysis, or news generation, helping to build trust in AI systems.

Real-world examples

• 1.AI assistants providing factual summaries
• 2.Content generation tools requiring high accuracy
• 3.Automated legal document review systems
• 4.Medical information systems