New tool makes historical books readable for AI

Old books are full of scanning errors that confuse AI models. Researchers tested fourteen different tools on two thousand pages to see which ones performed best. They wanted to find a way to fix these mistakes cheaply and quickly.

The best tool they found reaches over 97 percent accuracy. It costs less than two dollars to process a thousand pages. This level of quality is finally good enough to use for training large language models.

While this is a big step forward, the team says it is not perfect yet. It works well for AI, but it is not quite accurate enough for scholarly research. They plan to keep improving the technology for better results.