When can an AI image detector be wrong?
AI image detectors produce probabilities, not verdicts. A "Likely AI" score can be wrong. Knowing when false positives and false negatives occur helps you interpret results correctly instead of over-relying on a single number.
What a detector score actually means
AiPicDetect's detector returns a percentage and a confidence band. A score of 82% means the model assigns roughly 82% probability that the image is AI-generated — it does not mean 82% of pixels are "fake". The how-accurate page explains the full interpretation. In practice:
- High confidence (>90%): almost certainly right, but not infallible.
- Medium (50–90%): treat as one data point. Check metadata and apply visual inspection.
- Uncertain (<50%): the model genuinely cannot tell. Do not conclude either way from the score alone.
Common false positive scenarios (real photo scored as AI)
- Heavy retouching and skin smoothing
- Professional portrait retouching removes skin texture in ways that resemble AI synthesis. Heavy Photoshop retouching is one of the most frequent false-positive triggers.
- Composite and product photography
- Commercial product shots with uniform studio lighting, clean backgrounds, and perfectly matched colour grades score higher than snapshots because they share composition patterns with AI outputs.
- HDR and tone-mapped photos
- Aggressive HDR processing adds an "over-rendered" quality that looks similar to early diffusion model outputs.
- Shallow depth of field from long lenses
- Extreme bokeh from a 600mm telephoto or macro lens can resemble the soft, dreamy backgrounds Midjourney v5 produces.
- Screenshots and re-saves
- JPEG compression artefacts added by a screenshot tool can push scores in either direction by masking the original pixel distribution.
Common false negative scenarios (AI image scored as real)
- AI images passed through photo editors
- Running a Midjourney output through Lightroom, adding film grain, and sharpening edges adds "real photo" signal that moves the score toward Uncertain.
- Very recent model checkpoints
- The training set lags the newest generators. Flux, SDXL turbo, and Midjourney v7 are harder for the underlying model because it has seen fewer examples.
- Low-information crops
- Cropping to a small texture patch (sky, fabric, grass) strips context that the model uses to make its prediction. Scores become unreliable on tight crops.
- Anime and illustration styles
- The training distribution is skewed toward photorealistic images. Stylized anime-style AI art is more likely to score Uncertain even though it is clearly AI-generated.
What to do when the score surprises you
- Check the metadata panel in AiPicDetect. Metadata often tells you more than the score.
- Run the four-step check from the main guide: visual inspection, metadata, reverse image search, C2PA.
- If the original file is available, upload that rather than a screenshot.
- Treat any score between 40% and 85% as "uncertain" for high-stakes decisions.
Ultimo aggiornamento 2026-09-13 · codice sorgente su GitHub