Does compression affect AI image detection?

The image you need to check almost never arrives as the file the camera or the generator produced. It reaches you after a trip through a messenger app, a screenshot, a downscale, and a couple more re-uploads. Every one of those steps quietly rewrites the pixels, and detection lives or dies on those pixels. So the short answer to the title is yes: compression affects detection, more than almost anything else does, and a detector that pretends otherwise is quoting you a number from a world you don't live in.

The chain an image survives before it reaches you

Our benchmark degrades every test image through the same abuses real evidence goes through, because scoring on clean files would flatter the detector and mislead you. The conditions, roughly in order of how much they hurt:

• Clean. The original file, no re-encoding detected. The best case, and the rarest.
• JPEG recompression. Saving a JPEG again, moderately or heavily. Nearly every "Save As" and most uploads do this.
• Downscaled to half resolution. Resizing discards pixels, and with them the fine detail a detector reads.
• Messenger re-encode. Apps like WhatsApp re-compress and resize on a roughly 1600px dual pass, aggressively.
• Screenshot. A viewport rescale, a brightness shift, and another re-encode, all in one.
• Social chain. A screenshot of a WhatsApp image: the compound insult that actually walks into a lab.

Most images you care about have been through several of these, not one.

Why degradation erases the signal

AI generators leave statistical traces in the high-frequency part of an image: fine-grained pixel relationships no eye can see but a trained detector can score. That is also, precisely, the part of an image that lossy compression and downscaling throw away first. JPEG exists to discard high-frequency detail you won't consciously miss; a screenshot resamples the pixel grid; a messenger re-encode does both, twice. The signal a detector depends on and the signal these processes destroy are the same signal.

That produces two separate failures, and honest detection has to answer for both.

Failure one: detection power drops

With less signal surviving, more AI images fall below any given threshold and slip through unflagged. This failure is unavoidable; the only honest response is to measure it per condition and say so.

Failure two: false positives inflate, if nobody calibrated for it

An un-calibrated detector reads compression artifacts as evidence of synthesis and starts flagging real photos. A public detector at its default cutoff flagged about 27% of genuine photos in our benchmark once they had been through a messenger re-encode. This failure is avoidable, and a tool that suffers it was never calibrated for the images it actually sees. There is more on that in what false positives mean.

What the numbers actually do

Here is the part vendors skip. Ranking quality holds up better than you might fear: even under the hardest social-chain condition, the detector's AUC is 0.933, so it can still mostly sort AI from real. But you can't act on a ranking; you act on a threshold, and once a threshold is fixed to keep false positives capped, degradation shows up as lost power:

• At the standard (≤5% false-positive) cap, detection runs about 64% on a clean file and falls to roughly 55% on a screenshot of a WhatsApp image, landing around 55-64% across conditions.
• At the strict (≤1%) cap, it runs about 45% clean and around 43% on the social chain, roughly 36-45% across conditions.

Read two things out of that. First, the drop is real: a screenshot costs you detection a clean file wouldn't. Second, the operating rate stays capped and honest across conditions only because the threshold is recalibrated for each one. Hold a single threshold across all of them and the false-positive rate blows past its cap on the noisy conditions. The power you give up to degradation is the price of keeping the alarm trustworthy.

The honest fix: estimate the condition, then calibrate for it

You can't undo compression, but you can stop pretending it didn't happen. Our pipeline estimates an image's degradation history from its encoding fingerprint, selects the threshold calibrated for that condition, and prints both on the report: the estimated condition, and the detection power measured under it. A verdict on a screenshot is judged against how the detector performs on screenshots, not on pristine files it will never see. The report even lists detection power for every condition, so you can see how the verdict would move if the image's real history differs from the estimate.

This is the whole reason the benchmark scores 8,064 images across 8 conditions instead of one clean set. A single accuracy figure, quoted regardless of image history, isn't a measurement of the image in front of you; it's a measurement of a cleaner image you don't have.

What this means for the image on your desk

• Preserve the original; work from the least-degraded copy. If you can get the file as sent rather than a screenshot of it, do. Every extra hop strips signal you can't get back.
• Read "not flagged" even more cautiously on a battered image. Low surviving signal means a pass carries little weight, and the report's estimated condition tells you roughly how little.
• Ask any detector what it does about condition. "Does your accuracy number account for compression and screenshots, and is it measured per condition?" A tool that can't answer is quoting clean-image performance for real-world images.

What degradation can still defeat

• Extreme degradation leaves too little to read. A 100-pixel thumbnail or a tenth-generation meme may carry almost no signal, and no method, ours included, can conjure a verdict from what isn't there.
• The condition estimate is an estimate. The pipeline infers history from fingerprints; an unusual processing chain can be misjudged, which is exactly why per-condition power is printed rather than a single hidden number.
• A new generator is still a new generator. Degradation is only one axis. An image from a model released after the calibration run can be missed no matter how clean it arrives.

The takeaway is boring and load-bearing: the question is never "how accurate is the detector," but "how accurate is it on an image with this history." Compression doesn't only lower the score; it changes which score is the right one to ask for. A detector worth trusting measures that, states it, and prints the limits next to the verdict.