AI Fails to Read Ancient Egyptian Hieratic Script in New Benchmark

Overview

A new benchmark designed to test artificial intelligence’s ability to interpret ancient scripts has revealed a significant gap in current large language model capabilities. Researchers tested 13 leading AI models against an image of Ancient Egyptian Hieratic script, and not a single model correctly identified the writing system.

Hieratic was the everyday cursive hand used by ancient Egyptian scribes for administrative, literary, and religious texts on papyrus and ostraca. Despite being one of the most well-studied scripts in Egyptology, the sample sentence was misidentified as everything from Urdu to Tibetan to Nüshu.

The failure highlights a persistent limitation: while modern AI models can recognize images of papyri or identify famous monuments, they struggle to read the actual texts inscribed upon them. This benchmark, created by Aly Moursy, founder of Veeza AI, underscores the need for specialized training in historical and linguistic data.

The Benchmark and Its Results

The test involved a single sentence written in Hieratic by an Oxford Egyptologist in 2022. The sentence is a grammatically standard example used in introductory courses, meaning it should be recognizable to anyone with basic training in the script. The AI models were asked to identify the script and provide a translation.

The results were uniformly incorrect. Models from major providers, including Anthropic, Google, OpenAI, xAI, Meta, and Mistral, all failed to recognize the script. Some models confidently provided translations for scripts that were completely unrelated to the image provided.

Claude Opus 5.5 stated it could not identify the script, noting that it did not match Chinese, Japanese, Korean, Tangut, Khitan, Jurchen, or Nüshu. It also checked Gregg shorthand before concluding it could not give a translation without inventing one. Notably, it never considered Egyptian scripts.

"Opus 5.5 checked Tangut, Khitan, Jurchen and Nüshu. It checked Gregg shorthand. It never checked Egypt."

Claude Fable 5.1 initially claimed the script was not real. When pressed, it listed Hieratic and Demotic among the closest visual families but ultimately rejected them. Only after being told the image was Hieratic did it attempt a translation, incorrectly reading the text as the personal name "Amenemhat" with the meaning "Amun is in front."

Other models made equally bizarre errors. Claude Haiku 4.5 identified the script as Urdu and translated it as "One should learn every day." GPT-6 Astra repeatedly guessed Tibetan, Geba, Marchung, and Marchen. Gemini 3.8 Flash suggested Nabataean, Linear Elamite, and Proto-Sinaitic. Grok 4.7 guessed Arabic, Georgian Mkhedruli, and Baybayin.

Llama 4 Maverick identified the script as Mongolian, Hangul, and Georgian. Mistral Medium 3.5 guessed Rongorongo, Sinhala, Thai, and Syriac. Qwen3.8 Max guessed Mongolian script multiple times before giving up.

Implications for AI and Digital Humanities

The failure of these advanced models to read Hieratic has broader implications for the field of digital humanities. As institutions increasingly digitize ancient texts, there is a growing expectation that AI will assist in translation and transcription. However, this benchmark suggests that general-purpose language models are not yet reliable tools for reading ancient scripts, even those that have been well-documented for centuries.

Most of these models were trained on modern text data, with limited exposure to historical writing systems. Hieratic, in particular, requires specialized knowledge of sign forms, context, and grammatical structures that are not commonly found in mainstream training datasets.

The fact that some models produced confident but incorrect translations is particularly concerning. When an AI confidently states that a Hieratic text is Urdu, it reflects a deeper issue of hallucination rather than mere ignorance. Users may assume that the output is accurate because it is presented with authority.

Researchers argue that this gap needs to be addressed through targeted training. Adding diverse historical scripts to model datasets could improve performance, but it may also require specialized architectures or fine-tuning for specific domains.

Path Forward

The benchmark was created to provide a clear, reproducible test for AI capabilities in reading ancient scripts. By using a standard sentence from an Oxford course, the researchers ensured that the task was fair and verifiable.

Aly Moursy emphasized that the goal is not to dismiss AI but to understand its limitations. "Help AI read hieratic," he stated, pointing to the need for collaboration between technologists and Egyptologists.

The next steps involve developing models that can be fine-tuned on specialized corpora. Some researchers are already working on AI systems designed specifically for epigraphy and papyrology. These systems may require less general knowledge but more domain-specific training.

Until then, human experts remain essential. The benchmark serves as a reminder that while AI can process vast amounts of information, it still lacks the nuanced understanding that comes from years of study in specialized fields.

For now, the message is clear: despite rapid advances in artificial intelligence, we still need human scholars to read the ancient world.