Has Some Utility
Not definitive
Response from developer
A photo checked straight off an iPhone camera roll, taken that same morning, doesn't come back 40-50. The only way we've seen that split happen is when a file reaches the detector with its capture metadata stripped or altered somewhere along the way, so whatever you tested had been through more than your wife's camera. The numbers themselves work like this: when the original capture metadata is there, the call is decisive, and when it isn't, we score the pixels alone using our SOTA Image detection model and print whatever probability that evidence actually supports. Plenty of detectors would round a 55 up to a 99 because it looks better in a screenshot. We think that amounts to lying with statistics. Share the photo directly from your camera roll and run it again and compare the results.
Garbage
Response from developer
Thanks for the review. Slop or Not checks images in layers. Files carrying AI provenance metadata, meaning C2PA, IPTC, or Google's SynthID tags, get flagged as AI, and in our internal tests that catch rate is 100%. Files without metadata go to OmniAID, our own detection model, which scores 90% on unwatermarked AI images in those same tests; results can shift as new generators appear. So yes: edit a real photo with an AI tool and the tool writes AI provenance into the file. The detector flags it because the file says so. That part of your review checks out. The other half doesn't. "Showed AI ones and says it's real" is disingenuous: it describes behavior we cannot reproduce. If it really happened, email the photo to support@slopornot.ai and we will fix it for every customer after you. The same detector runs free at slopornot.ai. Anyone reading this can test your claim themselves in seconds.
Did not detect reliably
2. Accuracy claims are misleading, given that confidence scores vary, and this score was not very high on an AI abstract I tested. People will falsely assume a 75% confidence score falls within the 95% accuracy threshold.
3. The possibility of false positives is not disclosed, which could lead to overconfidence and unfair penalization of students, if present at all.
No way to attach text, pdf, or doc files on iOS.
Response from developer
Quick note on the numbers since they get conflated a lot. Per-class accuracy on our test set and per-image confidence aren't measuring the same thing. 75% confidence on a sample doesn't contradict 95% per-class accuracy. False positive risk is called out in the app itself and on the store page, and PDF or document attachment isn't on the roadmap because the iOS Share Sheet already handles those flows. If you've got a sample you want us to look at, mail support@slopornot.ai and we'll dig in. EDIT: For anyone reading this, the review was written in bad faith. The Slop Or Not text detector scores 99.09% (@ 1% FPR) on the RAID benchmark. https://raid-bench.xyz/leaderboard
Not quite what I thought it was - yet.
Response from developer
You're right, iOS won't let an app scan things while you're inside another app. Apple's privacy rules need an explicit tap to hand content over. For now the quickest route is Share Sheet, then pick Slop Or Not. We're looking at an opt-in Photos auto-scan for a future build. Thanks for keeping the door open, please give it another go when that lands.
False positives
Response from developer
Sorry the first run was a miss. Logos and other hand-made artwork getting flagged is the failure mode we hate most. Since v1.0.4 the classifier has been retrained on a lot more design-tool exports, and there's now an EXIF/XMP shortcut that picks up genuine Adobe and Affinity metadata. If you've got a minute, email one of those logos to support@slopornot.ai. We'll run it through the current model, tell you what actually tripped it, and add it to our test set.
Extremely inaccurate + expensive
Actually useless
If I feed it any of my artwork prior to shading (so it has a lot of spaces that are just a solid colour) it will say my artwork is 99+% AI slop. Additionally feeding it any image that is a single solid colour, such as a screenshot of itself, will say that it is 100% ai generated. I tried generating afew images, locally too for even more obvious samples, all returned as real images.
This app does the opposite of what it is supposed to do and should not be worth your time, especially not worth your monthly subscription to be able to “check” more than 3 images per day. Furthermore, it includes a tool to help disguise AI generated slop which is completely against what the app advertises itself as. Your great grandmother is likely a more reliable tool than this app at detecting AI slop, do not recommend.
Response from developer
Thanks for the careful test. The flat-colour and low-entropy failure you described is real, and we've been chipping away at it. The classifier in v1.0.6 and later was calibrated against more solid-colour reference plates and handles single-fill regions differently, so a flat-colour screenshot shouldn't come back as 99% AI any more. On the Britishizer, that one's there on purpose. The whole point is to show how easy it is to dodge AI-detection tools, so people see where the limits sit. If you ever want to revisit, mail support@slopornot.ai and we'll walk through your samples.
I had high hopes
Response from developer
The "it's AI but it tricked me" samples are honestly the most useful feedback we get. Worth another look if you've got a moment, the latest builds ship with our new OmniAID model, which is a meaningful step up on the cases that were tripping the old classifier. The remaining weak spots are mostly heavily edited photos and a handful of Stable Diffusion fine-tunes, and if you can email a screenshot or even a hash to support@slopornot.ai we'll add similar examples to our evaluation set so the next pass narrows the gap further. Thanks for sticking with it.
