Fake Images Detector is a decision-support tool, not an arbiter of truth. Its output is an indicative forensic signal. It is not proof, not an expert opinion, and not admissible evidence of authenticity or manipulation.
1. What the score actually means
The authenticity score is a weighted composite of four heuristic measurements: metadata integrity, noise character, compression response and error level distribution. It expresses how closely an image resembles an unmodified capture according to those four tests — nothing more.
A high score means no strong manipulation artefacts were detected. It does not mean the image is genuine. A low score means artefacts consistent with editing or generation were detected. It does not mean the image is fake.
2. Known causes of false positives
An authentic photograph can score poorly when it has been:
- Exported, cropped or resized in any editor, including phone gallery apps.
- Re-saved multiple times, or downloaded from a platform that re-compresses uploads.
- Captured with aggressive computational photography — modern smartphone HDR, night mode and portrait modes all apply heavy in-camera processing.
- Colour-graded, denoised, sharpened or filtered for legitimate editorial reasons.
- Screenshotted, or converted between formats such as PNG to JPG.
3. Known causes of false negatives
A manipulated or generated image can score well when it has been:
- Laundered through a screenshot, re-encode or social-media upload that destroys the artefacts we measure.
- Generated by a model whose artefact signature is not yet characterised.
- Edited so slightly — a removed blemish, a small clone — that the change is statistically invisible at whole-image scale.
- Deliberately post-processed with noise injection or re-compression to defeat forensic analysis.
Adversarial evasion is a real and active field. Anyone motivated to defeat automated detection can often do so.
4. Not a substitute for verification
Forensic signals are one input among several. Responsible verification also involves establishing provenance and chain of custody, reverse image searching, examining original files and camera metadata, corroborating with independent sources, and — where the stakes justify it — commissioning a qualified forensic examiner.
5. Not professional advice
Nothing produced by the Service constitutes legal, journalistic, investigative, insurance, medical or other professional advice. Results should not be the sole basis for publishing an allegation, denying a claim, disciplining or dismissing a person, initiating or defending litigation, or making any other consequential decision.
6. Reports are informational
The downloadable report is a convenience record of what the tool computed in your browser at a given moment. It is not a certificate, is not notarised or signed, carries no chain of custody, and should not be presented as authenticated evidence.
7. Results may change over time
Detection heuristics are refined as manipulation techniques evolve. The same image analysed at different times, in different browsers, or at different resolutions may produce different scores. Treat any single run as a snapshot rather than a fixed fact about the file.
8. No warranty and no liability
The Service is provided "as is", without warranty of accuracy, completeness or fitness for any purpose. To the fullest extent permitted by law we accept no liability for any loss, damage, reputational harm or adverse outcome arising from reliance on its output. Your use of the results is entirely at your own risk.
9. Your responsibility
You are responsible for having the legal right to analyse any image you process, for interpreting results competently and in context, and for any decision, publication or accusation you base upon them.
10. Questions
If a result seems clearly wrong — a known-authentic image scoring low, or an obvious fake scoring high — we genuinely want to hear about it. Please tell us; those reports are how the heuristics improve.