Google Gemini 3.1 and OpenAI's GPT 5.4 reach 85% accuracy on handwritten medical forms, making manual data entry obsolete for hospitals and clinics

Researchers benchmarked 17 AI models on real medical forms mixing handwriting, printed text, and dates. Google's Gemini 3.1 and OpenAI's latest models achieved ~85% accuracy with 90% weighted F1 scores (a measure of precision and completeness), while older or smaller models failed. GPT 5.4 stood out for extracting dates accurately even from messy handwriting, with only 6% of responses containing made-up information.

Unlike previous AI systems that struggle with messy real-world handwriting, these newest models handle the variability of human writing, crossed-out text, and mixed formats (dates, numbers, free-text notes) in a single pass. This means a medical clinic could scan a patient intake form and have 85% of the data automatically populated into their digital system—cutting the staff time spent retyping from hours to minutes.

Medical offices, insurance companies, and government agencies that currently pay staff to manually type handwritten forms into databases can now use these models to cut labor costs. A typical clinic processing 50 forms per day could redeploy staff from data entry to patient care or other higher-value work.