For years, the intersection of artificial intelligence and art history has been fraught with a specific, dangerous kind of fiction. A researcher or a gallery owner asks a large language model about the ownership history of a mid-century masterpiece, and the AI responds with a confident, elegantly phrased narrative. The problem is that the narrative is often a hallucination, blending real names with imagined dates to fill the gaps in its training data. In the high-stakes world of art provenance, where a single missing link in a chain of ownership can strip a painting of its value or signal a legal nightmare, these fabrications are more than just glitches; they are liabilities.

The Architecture of Verifiable Art History

Show GN has introduced a technical intervention to this problem by releasing a dedicated Model Context Protocol (MCP) server focused on artwork provenance. By leveraging MCP, a standard protocol designed to connect LLMs to external, structured data sources, Show GN allows models to step outside their internal weights and query a verified knowledge graph in real time. This server is currently available as a free open beta, allowing users to connect it directly to AI interfaces such as Claude, ChatGPT, and Cursor via a provided URL.

The scale of the underlying data is significant. The server is built upon a knowledge graph consisting of 282,731 nodes, drawing from authoritative sources including the Getty Provenance Index and the historical ledgers of the Knoedler Gallery. This is not a simple text database but a structured map of relationships that the AI can navigate to trace the movement of art across centuries. To make this data actionable, Show GN has implemented 27 distinct tools within the server. These tools enable the AI to perform complex tasks such as verifying ownership chains, flagging potential risks associated with the Nazi era, and generating summaries of auction price percentiles.

To ensure the reliability of these outputs, the system was put through a rigorous validation process. Using a live API eval harness, Show GN reports that the hallucination rate for the server is less than 1%. This represents a fundamental shift in how AI handles specialized historical data, moving from probabilistic guessing to deterministic retrieval.

The Power of the Informed Blank

The true innovation of the Show GN MCP server lies not in what it tells the AI, but in what it forbids the AI from inventing. The core tension in generative AI is the drive toward completion; LLMs are designed to provide an answer, even if that answer requires filling a void with plausible-sounding nonsense. In the context of art provenance, a gap in the record is a piece of information in itself. It indicates a period of obscurity, a lost record, or a potential red flag that requires human investigation.

Show GN has engineered the server to prioritize silence over fabrication. When the knowledge graph encounters a gap where no data exists, the server does not prompt the LLM to speculate. Instead, it returns a blank or a null value as the data point. This forces the AI to report the absence of information rather than inventing a bridge between two known dates.

Consider the case of Jean-Michel Basquiat's Untitled (Boxer). In a standard LLM interaction, the model might attempt to smooth over the gaps in the painting's history to provide a seamless narrative. However, when powered by the Show GN MCP server, the AI explicitly identifies that the ownership chain is unverified from 1982 to 2017. By outputting the fact that the data is missing, the system transforms the AI from a storyteller into a research assistant. The contrast is stark: one approach offers a comfortable lie, while the other offers a rigorous, if incomplete, truth.

This approach solves the causality problem inherent in RAG (Retrieval-Augmented Generation) systems. By constraining the model's context to a verified knowledge graph and enforcing a strict no-fabrication rule for missing nodes, Show GN eliminates the primary cause of provenance hallucinations. The AI is no longer predicting the next likely word in a sentence about a painting; it is reporting the state of a verified database.

This shift toward verifiable, protocol-based data retrieval marks the beginning of a new era for high-value asset verification in the age of AI.