How to Read an Authenticity Score Without Fooling Yourself
A number between 0 and 100 invites false confidence. Here is how to interpret the score, the bands, and the layer breakdown responsibly.
Our detector returns a single authenticity score. That number is useful and it is also dangerous, because a single number invites a binary reading that the underlying evidence does not support.
What the score is
It is the mean of four independent heuristic measurements: metadata integrity, noise character, compression response, and pixel consistency. Each returns 0–100, and the composite is their average.
That is all it is. It is not a probability that the image is fake. It is not a confidence interval. It is a summary statistic over four tests, each of which can be wrong in known ways.
The three bands
- 75 and above — likely authentic. No strong manipulation artefacts detected.
- 50 to 74 — inconclusive. Some layers returned anomalies. This is a genuine “we don’t know,” not a soft accusation.
- Below 50 — strong manipulation signals. Multiple layers flagged patterns typical of generation or heavy editing.
We deliberately kept the middle band wide. Tools that force a binary verdict produce confident-sounding errors, and confident-sounding errors are how people get defamed.
Always open the breakdown
The composite hides the interesting information. Consider two images that both score 62:
- Image A: metadata 85, noise 20, compression 78, pixels 65. The noise layer is screaming. That pattern is consistent with generated content.
- Image B: metadata 70, noise 65, compression 55, pixels 58. Everything is mediocre and nothing is alarming. That pattern is consistent with a normal photo that has been re-saved a few times.
Same score, completely different stories. The breakdown is not a detail — it is the actual result.
When layers disagree
Disagreement is informative, not a malfunction. A high metadata score with a very low noise score often means a generated image saved with plausible dimensions. A low compression score with a high noise score often means a genuine photo that has been through several export cycles.
Things the score cannot tell you
- Whether the content is true. A completely authentic, unedited photograph can be captioned with a lie about where and when it was taken. Forensics cannot detect a false caption.
- Who made the change. We measure that pixels are inconsistent, not who made them so.
- Whether an edit was deceptive. Cropping, colour grading and dust removal are normal editorial practice. The tool cannot distinguish routine processing from intent to mislead.
The discipline
Treat the score as a prompt for investigation, never as its conclusion. If it is low, ask what else corroborates that. If it is high, do not stop checking provenance. The number narrows where you look; it does not tell you what you have found.