Why Real Photos Get Flagged: Our Thresholds Are Public

Ask any AI image detector how accurate it is and you will get a number: 98%, 99%, "industry-leading". Ask to see the test set, the decision thresholds, or the changelog — and the answer is usually silence. We think that is backwards. Here is exactly how the AI picture detector scores images, when it fails, and why a false accusation bothers us more than a missed fake.

The five verdict levels, published

Every scan produces a 0–100 score that maps to five public bands:

• 0–14 — very likely a real photo
• 15–39 — likely a real photo
• 40–59 — uncertain (treat it as inconclusive, not as an accusation)
• 60–84 — likely AI-generated
• 85–100 — very likely AI-generated

The bands are versioned with the analyzer (v1.1.0). If we ever change them, the change gets a date and an announcement on the tool page. A score you cannot reproduce is not evidence — it is marketing.

Four ways a real photo gets flagged

Our detector reads sensor noise, texture flatness, and color statistics. Real photos can drift toward the AI side of those statistics in four common situations:

1. Heavy compression. Social platforms re-encode uploads aggressively; that destroys the fine sensor noise our noise signal looks for. A screenshot of an Instagram photo can land in the uncertain band.
2. Filters. Denoise, sharpen, and AI-enhance filters literally rewrite the noise pattern. Phone night mode and beauty filters push real photos toward smooth, AI-like statistics.
3. Computational photography. Modern phones already blend synthetic content into real photos — HDR merging, portrait-mode synthetic bokeh, AI-assisted sky replacement.
4. Flat graphics. Screenshots, documents, and charts are smooth by nature. Statistically they resemble renders, no matter how they were made.

Three ways an AI image slips through

Detection cuts both ways:

1. Re-compression after generation can mask the smooth statistics we flag.
2. Added grain. Film emulation and editing apps can paint sensor-like noise on top of a render.
3. Older and real-time models sometimes produce noise that reads as photo-like to pixel signals.

This is why embedded metadata — a Midjourney tag, Stable Diffusion parameters, a camera EXIF block — is treated as near-certain evidence when present, while pixel statistics alone are always presented as probability.

The feedback loop

Under every verdict there are two buttons: correct and wrong. Clicking one stores an anonymous record — the score, the verdict level, the detected generator if any — and never the image (which never left your device in the first place). Those records feed threshold calibration. If real photos are drifting into the "likely AI" band, that is visible in aggregate, and it gets fixed in the next version.

The principle

A false "AI" verdict on a real photo is our worst failure. It can cost a photographer their reputation, a claimant their payout, or a student their standing. Missed fakes are bad; false accusations are worse. Every threshold choice is made with that asymmetry in mind.

Run a few of your own photos through the AI picture detector — the per-signal breakdown shows you exactly which evidence moved the score, so you can judge the judge.