The best AI models in the world can't read oracle bone script. That's the takeaway from Chronicles-OCR, a new benchmark released by Tencent's Hunyuan and SSV Digital Culture Lab, along with the Institute of Information Engineering at the Chinese Academy of Sciences. It's the first evaluation to cover the "seven script variations" — from oracle bone to cursive — with 2,800 expert-annotated images that quantify recognition difficulty across these ancient fonts.
The results are brutal. Across 28 major multimodal LLMs tested, performance on early scripts was abysmal. On cross-era character detection, GPT-5 and Gemini 2.5 Pro scored near zero. The best model managed just 16.5. Even when researchers eliminated the localization step by pre-drawing bounding boxes, top accuracy hit only 27.1%, with Gemini 3.1 Pro scraping by at 14.0% on oracle bone script.
This confirms that modern models rely heavily on clean, modern typography priors. Faced with unconstrained, noisy ancient physical media — like turtle shells or bronze with corrosion — their text segmentation mechanisms simply break down. Font classification results further show that models often identify the texture of the carrier (e.g., shell grain or bronze patina) rather than actual character strokes.
The experiments also revealed a counterintuitive finding: enabling reasoning mode actually made recognition worse. Nearly every model that supports such a mode saw performance degrade when it was turned on. When the underlying visual perception is missing, chain-of-thought reasoning doesn't just fail to correct errors — it becomes a hallucination amplifier, confidently outputting wrong answers.