Gemini repeatedly failed to read text in the same screenshot
In a test, the same was sent to Gemini more than 15 times using Extended, , and Pro Extended. The models repeatedly invented unrelated instead of reading the text that was visible. Deleting the message, uploading the again, and asking for another attempt did not produce a dependable result.
eventually claimed that it could not process text inside images, although it had completed the same task minutes earlier. Once that contradiction was challenged, it immediately produced an accurate summary of the .
Key points
- Three Gemini model were tested with the same .
- More than 15 requests produced repeated, unrelated .
- Deleting and uploading the image again did not solve the problem reliably.
- denied that it could read image text, then summarized it correctly.
- Results were inconsistent even with the same image and task.