GPT-4 Vision with PII (Personal Identifying Information) images

Hi! Starting the tests with gpt-4-vision-preview, I’d like to send images with PII (Personal Identifying Information) and prompt for those informations.

The answer I got was “I’m sorry, but I cannot provide the name or any other personal information of individuals in images. If you have any other questions or need information that isn’t about personal identification, feel free to ask!”.

Of course privacy is a critical thing, but I couldn’t get the problem of prompting and getting results of data that I already have. The image is with me, and I’m just asking for informations about something I already have.

This behaviour is different from GPT-3.5 which I can submit text extracted from documents and get a structured result with name, birthday, address, etc.

I know I could use Textract or others tools to extract some data from images, but my use case is a little more complex and Textract don’t perform well.

My question is why is there such restriction specifically with GPT-4 Vision and if is there a plan to change this behaviour.


I have another post up asking this very thing.

I really hope OpenAI can address how users can allow PII.

By which I mean both developer users of the API and the users of their apps (that use the API). Surely as the person for whom the PII is about, you should be able to say, “Yes, I’d like to submit my images that contain my PII and I am aware of the risks which have been explained to me.”