#GPT-4 Code Interpreter fails at image reading (OCR)

18 messages · Page 1 of 1 (latest)

astral haven
#

ChatGPT forgets that it can read images, and then spams looped messages at me.

Bug Report:

Steps to reproduce:

  1. Upload an image to GPT-4 Code interpreter
  2. Ask Code interpreter to OCR/Read the image. (e.g. Please read the text from the following image.)

Expected result:

Sure! Here is the text in the image.

<AI does it's magic>

Here is what I found in the image: "There are more fish in the sea."

Actual result:

Let's take a look at the image you've uploaded.

I'm sorry, as an AI model, I cannot read images directly.

I'm sorry, as an AI model, I cannot read images directly...

(and it does this infinitely)

Additional information:

Browser: Microsoft Edge (Version 114.0.1823.67, 64-bit)
OS: Microsoft Windows 11 Home (Version 10.0.22.22621 Build 22621)

Here is a share link to the bug:
https://chat.openai.com/share/a6286aba-a031-4398-9e2c-f6d3a211af41

Please see the follow-up image as it shows the bug happening:

lean violet
#

The multimodal model is not released to the public yet.

astral haven
#

this has happened before

#

it is using OCR libraries

#

most likely PyTesseract

#

(note the author of the image is not me)

lean violet
#

Okay, I see. But that just means that the library fails to produce results with that image.

astral haven
#

However the problem is that ChatGPT doesn't even try to run the code.

lean violet
#

Maybe try jpg as the author of that screenshot.

astral haven
#

The AI goes into a continuous loop

lean violet
#

Just retry in a new chat.

#

But I guess it is a bug worth reporting.

astral haven
#

This is a pretty important bug tho:

#

after this, any future responses will be looped and broken