Can GPT read tables accurately from photos of printed reports?

I’ve tried asking GPT to extract figures from photographed pages, and the results look convincing even when a row or column has been misread. How well does it handle dense tables, merged cells, footnotes and slightly skewed images compared with a clean PDF or spreadsheet?

For anyone using it with reports, what do you check before trusting the extracted numbers? I’m especially interested in whether there are particular image or table layouts that tend to cause silent errors.