A text processing technique promoted by Chinese artificial intelligence startup DeepSeek is facing scrutiny after new academic research questioned its claimed ability to help AI systems handle long and complex documents. Researchers from China and Japan have challenged the DeepSeek OCR method, which was introduced as a way to compress large blocks of text by transforming them into visual representations. The approach was positioned as a breakthrough that could allow models to read extended documents more efficiently by relying on visual perception rather than conventional text based processing. However, the new study argues that the technique does not consistently deliver the performance gains highlighted by the company. Instead, the researchers suggest that reported improvements may be overstated, raising broader questions about how experimental AI methods are evaluated and communicated within China’s fast moving technology sector.
The study was conducted by academics from Tohoku University and the Chinese Academy of Sciences, marking a rare instance in which DeepSeek’s research has been publicly examined by independent peers. Titled Visual Merit or Linguistic Crutch, the paper argues that DeepSeek OCR relies heavily on language priors, meaning the system draws on patterns learned from large text datasets rather than genuine visual understanding. According to the researchers, this dependence undermines the core claim that visual compression alone drives improved performance. They also noted inconsistent results across tasks, suggesting that the method’s effectiveness may vary widely depending on context. As a result, they described some of the performance metrics presented by the company as potentially misleading.
The critique highlights growing academic attention on how AI breakthroughs are validated as competition intensifies among Chinese technology firms. As companies race to develop models capable of processing longer texts for applications ranging from research analysis to enterprise automation, claims of efficiency gains carry significant commercial weight. Independent evaluations such as this one may encourage greater transparency around testing standards and benchmarking practices. While the findings do not dismiss the broader idea of multimodal approaches to text processing, they suggest that more rigorous validation is needed before such methods are adopted at scale. The debate reflects a maturing AI research environment in China, where collaboration and scrutiny between industry and academia are becoming more visible as the sector moves toward higher stakes applications.
