Decoding 17th-Century Alchemical Correspondence with Large Language Models

When AI Meets Alchemy: A New Tool for Historical Research

Alchemy is the forerunner of modern chemistry and one of the most opaque fields in the history of science. Seventeenth-century alchemists habitually recorded their experiments in ciphers, symbols, and encoded scripts, and their correspondence was riddled with code words and deliberately ambiguous terminology. For centuries, deciphering these documents has been an enormous challenge for historians and linguists alike. Now a new approach is gaining traction — using large language models (LLMs) to trace the transmission of alchemical knowledge and decode letters that have remained sealed for hundreds of years.

The topic was recently raised on Hacker News. Although the discussion is still relatively small, it touches on an increasingly prominent interdisciplinary space: how artificial intelligence can offer entirely new methodologies for humanities research.

Why LLMs Are Well-Suited for Historical Texts

The core strength of large language models lies in their ability to understand and generate linguistic patterns. For texts like alchemical writings — which rely heavily on context and metaphor — traditional keyword search or rule-based matching often breaks down, whereas LLMs can detect latent semantic relationships across vast corpora.

Handling Ancient Languages and Spelling Variants

Seventeenth-century correspondence was often written in Early Modern English, Latin, or a mixture of languages, with inconsistent spelling and abundant abbreviations. Because LLMs are trained on large-scale text data, they have strong tolerance for such variation and can identify the same concept or person name even when spelling is inconsistent. This makes them particularly useful for organizing manuscripts scattered across archives around the world.

Tracing the Chain of Knowledge Transmission

Alchemical knowledge typically spread through epistolary networks: a recipe or theory might pass from one practitioner to another, being rewritten or misrepresented along the way. By analyzing textual similarity and conceptual associations, LLMs can help researchers attempt to reconstruct these transmission pathways — identifying which terms or formulas share a common source document and thereby mapping the scholarly exchange networks of the era.

Decoding the Cipher and Symbol Systems

The greatest obstacle posed by alchemical literature is its deliberate obscurity. To protect their "secrets," alchemists routinely used code names for substances — planetary symbols for metals, mythological imagery to disguise chemical processes. This tradition of Decknamen made the texts virtually impenetrable to outsiders.

LLMs can serve as an auxiliary decoder here: by cross-referencing large collections of known alchemical symbol glossaries with contextual clues, a model can generate plausible interpretations and narrow the range of possibilities for human verification. Naturally, such inferences must be reviewed by specialist historians; AI provides leads, not conclusions.

The Decknamen (German for "cover names") tradition has deep historical roots. From the medieval period through the early modern era, alchemists feared both that competitors would steal their knowledge and that the Inquisition might charge them with sorcery, so they developed highly personalized encoding conventions. For example, "the Sun" commonly denoted gold, "the Moon" silver, and "Saturn" lead — but usage was far from uniform across authors and periods; the same symbol could refer to entirely different substances in different texts. More complex still, some alchemists deliberately mixed multiple symbol systems within a single text to increase the difficulty of decipherment. This means decoding cannot rely on a single lookup table but must integrate multiple contextual dimensions — author identity, date of composition, regional tradition, and the identity of correspondents. This is precisely where LLMs hold a potential advantage over simple keyword search, and also where their limitation lies: when contextual clues are insufficient, the model is equally likely to "randomly select" a plausible-sounding interpretation from among several possible readings.

Opportunities and Limitations

AI-assisted historical research is promising, but its boundaries are real.

The foremost concern is hallucination risk. When LLMs lack firm evidence, they may fabricate interpretations that sound plausible — a fatal flaw in rigorous historical scholarship. Every AI output must be cross-checked against primary archival sources and cannot be used directly as an academic conclusion.

A second challenge is the scarcity of training data. Alchemical literature is inherently limited in volume, and a large proportion of it has never been digitized. General-purpose large models have insufficient coverage of this niche domain. Improving accuracy typically requires fine-tuning on specialized corpora or combining the model with Retrieval-Augmented Generation (RAG) — using reliable historical databases as authoritative reference sources for the model.

Retrieval-Augmented Generation (RAG) is an important technical approach for mitigating LLM hallucination. The core idea is to avoid relying solely on knowledge memorized during training: at inference time, the system retrieves relevant document fragments from a trusted external database in real time, then feeds those fragments to the model as context, guiding it to answer based on authentic sources. For historical research, this means digitized alchemical manuscripts, scholarly annotation databases, or authoritative glossaries can be built into a vector database, anchoring each model inference to traceable primary materials and substantially reducing the probability of fabricated interpretations. Compared with full fine-tuning, RAG has the advantage that the knowledge base can be updated at any time and the reasoning process is traceable — researchers can see which specific document each answer draws from, which is especially important for humanities scholarship with strict citation requirements.

A New Paradigm for Digital Humanities

This case reflects a broader trend: LLMs are becoming a general-purpose tool for digital humanities research. From deciphering ancient inscriptions and organizing historical archives to analyzing the evolution of literary style, AI's pattern-recognition capabilities are dramatically accelerating work that once required decades of human effort.

For researchers, the key is not to let AI replace expert judgment, but to use it as a lever that amplifies human insight — letting AI handle the laborious initial screening and associative analysis, while scholars perform the final evaluation and interpretation. The encounter between this ancient art that once pursued the transmutation of base metal into gold and the most cutting-edge language models carries its own historical resonance: the pioneers who once chased lapis philosophorum are now being reinterpreted by another technology that transforms raw data into knowledge.

Digital Humanities as a formal academic discipline began taking shape in the 2000s, with its core mission being to combine the quantitative methods of computer science with the interpretive traditions of the humanities. Early work focused largely on text digitization and database construction (such as the British Library's manuscript digitization projects), followed by the development of text mining, social network analysis, and Geographic Information System (GIS) mapping. The emergence of LLMs represents the field's third major leap: whereas the first two waves were primarily about "getting computers to search and count humanistic data," LLMs are the first technology to make "getting computers to understand and generate humanistic language" genuinely feasible. Alchemical research is just one of many application scenarios; similar work is already underway in cuneiform translation, Shakespearean textual genealogy, medieval legal document analysis, and more — together forming a new ecosystem of AI-assisted humanities research.