OpenAI said, “thanks for the thoughts, closing the topic with no promises”
I have a good working theory of why API is getting no transparency generation feature.
- gpt-image-2 has SynthID watermarking embedded in the image
- OpenAI does not disclose the metadata traceability this could include, back to a specific API organization, such that they might discover who made an image bringing negative publicity
- Transparency on prior models was a programmatic technique, marking AI generated pixels of a certain type or color intensity transparent by the backend
- The background surface that could otherwise be cut out and removed may also be an image destination loaded full of logit watermarking
- You shall not be allowed an image where the watermarking and traceability can fail.
ChatGPT transparency: no watermark
My prior GIF destination posted here was not supported for upload to the verification tool). When I go back to get the transparency generated in the original file produced by ChatGPT, post gpt-image-2 launch (and new challenges and an internal tool method the AI must use):
Content credentials can be stripped with just a re-save.
“transparent” on the API - solely by prompting a background color as key.
WebP at compression 57/100 and quality:medium was requested for a bit of “lossy” in the delivery of a image - that I cut out myself:
A single API asset that might then be creatively integrated elsewhere with some Photoshopping then makes your whole final product marked as AI and an API product based on assets essentially useless.
Imagine someone takes a screenshot of your video game, with your AI sprite, and the whole image is flagged as a deception? Your stickers on your store web site serve as an ad for OpenAI.
No transparency by request on API, I’m sure, until they can guarantee similar forced identifying uselessness and origin traceability is ensured.