How HMNPROOF reaches a risk score
We scan images, videos, and URLs using a layered detection system — then give you a clear risk report you can actually use. No PhD required.
1. Upload or paste
Drop an image or video, or paste a link. We never ask for personal information to run a scan.
2. Layered analysis
Multiple independent checks run in parallel, looking at metadata, provenance, AI artifacts, and where the content came from.
3. Clear report
You get a single percentage score, plain-language summary, and a breakdown of what supported the verdict.
What we check
Every layer contributes a signal. No single test decides the result.
Metadata & provenance
EXIF, XMP, and C2PA Content Credentials reveal when, where, and how a file was created.
AI-generated artifacts
Vision models and byte-level forensics look for fingerprints left by generative tools.
Source & context
We trace where the content appeared, assess the platform's trust signals, and run reverse-image searches.
Face & identity checks
Optional face-region forensics and 1:1 identity matching against a photo you provide.
Video coherence
Frame-level analysis catches temporal inconsistencies common in AI-generated or manipulated video.
Continuous learning
A feedback loop re-weights signals based on real-world outcomes and user flagging.
What we cannot detect
Hiding these limits would be the opposite of trust. Here is where the technology stops and human judgment still matters.
- Re-compressed or screenshotted media that destroys watermark and forensic signal.
- Hybrid edits where a real photo is changed by AI in only a small area.
- Generators that intentionally suppress watermarks, metadata, and AI fingerprints.
- Context-level deception — a real image used to misrepresent what happened.
- Biometric-scale identity matching. We offer opt-in 1:1 comparison, not mass scraping.
HMNPROOF outputs guidance, not legal verdicts. Use it alongside human judgement and, when needed, professional advice.
Technical details for engineers & journalists
Layer names, detectors, and how confidence is derived.
Technical details for engineers & journalists
Layer names, detectors, and how confidence is derived.
HMNPROOF combines the following independent signals. Each signal returns a normalized score and a confidence weight. The final risk score is a weighted aggregate, not a single model output.
- EXIF / XMP metadata parsing via exifr
- C2PA Content Credential byte probe
- Shannon entropy + JPEG DQT analysis
- Gemini 2.5 Flash vision forensics
- Per-face artifact scoring
- Temporal video-frame coherence
- Reverse-image deep links (Google, TinEye, Bing)
- Opt-in 1:1 identity match descriptor
- Source-host reputation and context scoring
- Bayesian learning-engine re-weighting from feedback
Confidence intervals are derived from the spread of individual signal outputs. A wide interval means the layers disagree; a narrow interval means they agree. We surface low-signal cases explicitly rather than hiding them behind a single number.
Detector versions are pinned per scan so reports remain reproducible. New generators are added as pluggable modules without rewriting the core pipeline.
Why can't I just upload this to ChatGPT?
It's the most common question we get. The short answer: a general-purpose chatbot can describe a picture, but it cannot forensically authenticate one. Here is what HMNPROOF does that a chat model on its own does not.
ChatGPT looks at pixels. We look at provenance.
Every file we receive is parsed for EXIF, XMP, and C2PA Content Credentials — the cryptographic provenance markers that OpenAI, Adobe, Google, Microsoft, and major camera makers embed in real and synthetic media. A chat model does not read these byte-level signatures; it only sees the rendered image.
Our verdict is an ensemble, not a single opinion.
HMNPROOF runs ten independent detectors in parallel — metadata, C2PA, Shannon entropy, JPEG quantization tables, vision forensics, per-face artifact scoring, temporal frame coherence for video, reverse-image lookup, source-host reputation, and an opt-in 1:1 identity descriptor. Each returns its own score and confidence weight. The final risk number is a Bayesian aggregate where the layers cross-check each other. One model guessing is not the same as ten detectors agreeing.
We get smarter with every scan. A chatbot does not.
Our learning engine re-weights every signal based on aggregate scan outcomes and verified user feedback. New generators (Sora, Midjourney v-next, future diffusion models) are added as pluggable detector modules without retraining the core. A consumer chatbot's weights are frozen between releases — yours is not the scan that improves it.
We search the open web for where else this lives.
Every scan is fingerprinted with a SHA-256 content hash and run through reverse-image lookup. You see how many other places the same media appears online, with thumbnails and source links. A chat model has no live index of the visual web.
Our output is a defensible report, not a paragraph.
You get a probabilistic risk score with a confidence interval, signal-by-signal breakdown, source links, a PDF download, a public verification link, and — for low-risk originals — a serialized HMN Certified badge tied to the file's hash. That is something you can hand to an employer, a platform's trust & safety team, a journalist, or a lawyer. A chat transcript is not.
We handle video and URLs natively.
HMNPROOF samples video frames for temporal coherence and fetches URLs server-side with host-trust scoring. Pasting a TikTok or news link into a chatbot usually returns 'I can't access that.' We treat it as a first-class input.
Privacy and chain-of-custody.
Uploads run through automated safety screening, sit in a private bucket with row-level security, and are purged on a 24-hour cycle for anonymous scans. A rescan-by-hash cache lets you re-run a file against the latest model without re-uploading the original. Pasting sensitive media into a public chatbot has none of these guarantees.
HMNPROOF uses large vision models as one of its layers — not as the verdict. The verdict comes from the ensemble, the provenance signals, the web evidence, and the learning engine that tightens those weights over time.
Two models. One verdict. Honest disagreement.
Signed-in scans of still images are cross-checked by a second, independent vision model running a blind prompt. The agreement percentage is shown on every report. When the two models disagree meaningfully, we say so — we do not average disagreement into false confidence. This pattern is borrowed from peer-reviewed deepfake-detection literature, where ensemble agreement is consistently more predictive than any single classifier's confidence score.
≤15 point gap. Verdict is reinforced; confidence is raised.
16-35 point gap. Treat verdict as directional, not definitive.
>35 point gap. Confidence is reduced; seek the original.
Nonpartisan, transparent, accountable.
Nonpartisan by construction. HMNPROOF does not weight verdicts by who is depicted, which platform hosted the media, or what the content argues. The same image gets the same score regardless of the politics, religion, nationality, or viewpoint of the subject. Source-host reputation is a documented signal in the report — never a silent thumb on the scale.
Probabilistic, not declarative. Reports return a risk percentage with a confidence interval and a signal-by-signal breakdown. We do not say "this is fake" or "this is real." We say what the evidence suggests and how strongly the layers agree.
Disclosed limitations. Every report lists what HMNPROOF cannot detect (re-compressed media, partial AI edits on real photos, generators that suppress provenance markers, context-level deception). Hiding limits is the opposite of trust.
Pinned detector versions. Each scan records the exact detector versions used. Reports remain reproducible against the model state at the time of analysis, not the model state today.
Independent of generator vendors. HMNPROOF does not train, sell, or license generative models. There is no conflict-of-interest between selling the cloning and selling the detection.
Standards and prior art we build on
- C2PA — Coalition for Content Provenance and Authenticity. Cryptographic content credentials embedded by Adobe, Microsoft, OpenAI, Sony, Leica, and others. c2pa.org
- IPTC / EXIF / XMP metadata standards. Industry-standard provenance fields embedded by professional cameras and editing software.
- NIST SP 800-238 (draft) and the AI Safety Institute's media-authentication work. Government-aligned guidance on synthetic-media detection benchmarks.
- Partnership on AI — Synthetic Media Framework. Industry framework on responsible disclosure and detection. partnershiponai.org
- DARPA SemaFor program. Foundational research on semantic forensics that informs the ensemble-agreement pattern used in this report.
HMNPROOF is independent and not endorsed by the above organizations. References list the public standards and research areas that informed our methodology.