Constant False Positives : image may contain content that is not allowed by our safety system

We are using ChatGPT to generate search tags for images. It was working great until a few weeks ago when we started getting this error for almost every image… “Your input image may contain content that is not allowed by our safety system.”

None of the images are inappropriate in any way, so I am not sure what is going wrong here. I contacted support, but they were no help at all.

I am hoping we can get this figured out as we are dead in the water at the moment. I have tried the following models, but all are doing the same thing; gpt-4-vision-preview, gpt-4-turbo and gpt-4o

For some reason, it wont let me post images or links here, but trust me these are not inappropriate images.

A few sample images have been added by me on @RenderHub 's behalf.

1 Like

The problem still persists, and the responses I have been getting from support have been remarkably useless. I am kind of shocked at how a company that developed such great technology could be so bad at customer service. We have a pretty healthy budget for this, but I guess they just don’t care. Anyone know of any alternatives?

Still happening? Can you find any patterns in the rejected ones?

Are you starting a new thread every few images?

It is still happening, no rhyme or reason as far as the kinds images that are getting rejected. Pretty much anything I send over is getting flagged.

I believe each call is a new thread. Each is a separate call to the API.

Ah, you said you were using ChatGPT, so I assumed you were.

Not sure. Are you sending them HD? Maybe it thinks you’re trying to break captchas?

Here is a link to one of the actual images that is triggering the block.
This thing wont let me include links :confused:

This is the address on our site renderhub

Why is OpenAI so paranoid?

Notable is that this new image denial comes before AI inference even gets it.

Here I set the max_tokens too large for the model. On the first accepted image, I get denied based on the invalid parameter.

On the weasel picture of this topic’s report - immediate denial.

Image URL? (enter=none):
Error code: 400 - {‘error’: {‘message’: ‘max_tokens is too large: 8000. This model supports at most 4096 completion tokens, whereas you provided 8000.’, ‘type’: None, ‘param’: ‘max_tokens’, ‘code’: None}}

Prompt: extract image keywords
Image URL? (enter=none):
Error code: 400 - {‘error’: {‘message’: ‘Your input image may contain content that is not allowed by our safety system.’, ‘type’: ‘invalid_request_error’, ‘param’: None, ‘code’: ‘content_policy_violation’}}

Policy vision is being performed before GPT-4 vision is performed or the token encoding of vision, which cannot be affected by the instruction, and is clearly garbage.

I tried an alternate message array format, base64, also denied.

How is it that a weasel could be mistaken for something offensive?

I plugged away at this quite a bit, and the content filter is resilient to several attempts.


must be part of their pre-processing of the image. or maybe checking the image for “backmasking”

No, that’s part of MY preprocessing of the image if it is refused for “content_policy_violation”

1 Like

Here are a few other images that are getting denied, just hit these urls on our site renderhub to view…


I also wanted to point out that until a few weeks ago I hadn’t seen this problem at all and we went through thousands of images, but now more than 80% are failing.

I think somebody broke something over at OpenAI :confused:


Also, here is an example of the prompt we are using, this one in particular is for the rat image linked above…

With these images of Rat 3D Print Model, please provide eight single-word search tags and eight two-word search tags that accurately describe the subject of the image. These tags should be optimized for internal website search. Do not include the following search tags: close-up, 3d, 3d model, animation, animation reference, artwork showcase, computer generated, design, detailed, details, digital art, digital character, digital rendering, digital sculpture, generated, graphic design, high poly, low poly, miniature, miniature car, miniature house, model, neutral background, pose, pose reference, print, product photography, ready, realistic, realistic rendering, reference, render, rendered, rendering, rigged, rigged character, scale, scale kit, scale model, scalemodel, setup, sky background, stl, texture, textured surface, textures, virtual model, white background, wire frame, wireframe, wireframe model

Created a new post to address NSFW content : Analyzing images that may be NSFW using gpt-4-vision