anyone else dealing with this? im a creative marketer and two of my clients in the last month have asked me to verify that stock images im using in their campaigns arent AI generated. one of them specifically put it in the brief: “no AI-generated imagery.”
the problem is that i have no reliable way to do this at scale. i buy maybe 60-80 stock images a month across projects. reverse image search catches some obvious ones but its not practical for everything.
are there any image authenticity tools that actually work for this use case? or is this one of those things where the detection technology is still way behind the generation tech?
dealing with the exact same thing. i manage content across multiple sites and the image verification question keeps coming up.
currently my approach is a combination of things:
- use established stock libraries that have explicit AI policies (Adobe Stock now labels AI-generated content)
- reverse search anything that looks too perfect
- check EXIF data for camera metadata - real photos usually have lens info, GPS coordinates, etc
- look for common artifacts: fingers, text in background, impossible reflections
its not foolproof but it catches the obvious stuff. the harder problem is images that were partially AI-enhanced. like real photos with AI-generated backgrounds.
From a publishing perspective, this is becoming a real editorial challenge. We updated our contributor guidelines six months ago to require that all submitted images include provenance information. Camera model, date taken, or the original stock library receipt.
The problem is enforcement. We receive hundreds of submissions monthly and we simply cannot manually verify every image. The industry needs standardized image provenance metadata baked into the file format itself - something like C2PA is trying to solve but adoption is painfully slow.
from a technical perspective the detection problem for images is fundamentally different from text. with text you’re analyzing statistical patterns in token sequences. with images you’re looking for artifacts at the pixel level, and those artifacts change completely with every new model generation.
GAN-generated images had specific frequency domain artifacts that were fairly detectable. diffusion models are much harder. the most promising research is looking at noise pattern analysis and semantic consistency checking rather than pixel-level artifacts.
short answer: reliable automated detection at scale doesnt really exist yet for current-gen image models.
I consult for several B2B companies on content strategy and this exact issue came up in three separate client meetings this quarter. The concern is not just aesthetic - it is legal. Some jurisdictions are starting to require disclosure of AI-generated media in commercial use.
My recommendation to clients is to build provenance requirements into vendor contracts now rather than trying to detect after the fact. Require your stock photo providers and designers to certify the origin of their deliverables. Prevention is more practical than detection at this stage.