Why accuracy varies so much
A detector may do well on clean generated artwork and struggle with a compressed phone screenshot. It may flag a smoothed real portrait and miss a generated image that has been edited after export. That does not make the tool useless; it means the result needs context.
Accuracy also changes over time. Image models improve. Editing tools add new workflows. Platforms compress files differently. A detector score should be read as a current signal for this file, not a universal verdict about the image forever.
False positives: when a real image looks synthetic
A false positive happens when a real or mostly real image is flagged as likely AI-generated. Common reasons include heavy skin smoothing, product retouching, artificial studio light, HDR processing, low-resolution screenshots, or repeated texture after compression.
This is why public accusations based on one score are risky. A seller's product photo may be over-edited, not fake. A profile picture may use a beauty filter, not a generated face.
False negatives: when a generated image looks clean
A false negative happens when a generated image receives a low or uncertain score. This can happen when the image is realistic, post-processed, downscaled, screenshotted, or mixed with real photo elements.
For important images, a low score should not stop the review. It should simply lower the priority of the AI-generation question and move attention to source, caption, and usage.
How to read confidence wording
Think in three buckets. A high signal means "pause and verify." An uncertain signal means "the file is not enough." A low signal means "no strong AI signal found here." None of those means "case closed."
The more serious the consequence, the less you should rely on one detector result.
What improves your odds
- Upload the original file instead of a screenshot.
- Compare the score with metadata, not just visual impressions.
- Check whether the image has been resized or reposted many times.
- Use reverse image search for source history.
- Keep notes about uncertainty instead of forcing a yes/no answer.
FAQ
Is any AI image detector always right?
No. Detectors are probabilistic tools. Good ones can still produce false positives and false negatives.
Why did two tools disagree?
Different tools use different models, thresholds, and training data. Disagreement usually means the image deserves a broader review.
Does original file quality matter?
Yes. Original files usually preserve more detail and metadata than screenshots, thumbnails, or reposts.