Next generation of verification.

Case 0412 · GaussMass

profile_photo.jpg3024 × 4032 · sRGB

Evidence, in log-oddsauthentic ← 0 → synthetic
AI model
+1.10
Spectral
+0.85
Diffusion
+0.70
Temporal
n/a
EXIF
+0.25
Web intelligence
−0.15
Sum +2.75p = σ(sum) = 0.94
VerdictSynthetic0.94 calibrated likelihood of synthesis

Three evidence families agree.

Routed to
Case queue
Audit packet
Signed
An illustrative evidence card. Every scan returns the verdict and the evidence behind it.

Partnerships and support

What clients say.

“Plurall AI helped us show the creative upside of AI videos to our clients while clearly exposing the risks of deepfakes. Their team equipped us to teach non-technical stakeholders to recognize red flags and understand where manipulation can creep in.”
CEO, Lapta Media
“Plurall has designed its deepfake detection tool from the ground up.”
TechCrunch, December 29, 2025Named one of the 32 top enterprise tech startups from Startup Battlefield at Disrupt 2025.Read the article ↗

Evidence, not a score.

Synthetic media is cheap to make and hard to spot. We fuse forensic signals across media, identity, and behavior so a reviewer can see why something is flagged, not just that it was.

Our mission →
Evidence console
VerdictDeciding
authenticreviewsynthetic
0.50
A simulation of published principles, for illustration. It is not model output and shows no product parameter.

Measured by slice.

14.07 ms
to score an imageGaussMass on NPU, 90.24 MB model
512 / 512
real portraits passed as authenticOpen real-portrait guard, N = 512
93.89 %
recall on edited synthetic imagesMoves 0.64 pp after benign edits
24–48 h
from a new generator to a re-scored releaseFull-network retraining, one GPU
How the numbers are made →

Detection you can measure.

How the numbers are made, from the public technical papers. Read the papers ↗

One platform, three surfaces.

Deepfake detection
Image, video, audio, metadata, spectral, temporal, and web-intelligence signals fused into a clear verdict, with the evidence behind it.
Fraud defense
Escalate suspicious media, impersonation attempts, and account-level risk into cases your team can resolve.
Proof-of-humanhood
Verification workflows, liveness status, and audit trails for marketplaces, platforms, and trust teams.
Start with a pilot →

Deepfake detection for enterprise.

Client portal

Upload an image, a video or a recording and get the verdict with the evidence behind it. Review cases with your team and export an audit packet for each one.

Sign up for free ↗

API

Send media from your own product and receive the same verdict, evidence and audit trail as JSON, with webhooks when a scan settles.

Read the docs →

Chrome extension

Check an image or a video on the page you are already on. The verdict and the forensic analysis open beside it.

Add to Chrome ↗

Partner deployment

Managed security providers roll Plurall AI out to their own clients and route alerts into the monitoring they already run.

Partner with us →

Watch the demo.

See the Plurall AI platform at work in the client portal.

Try it for free ↗

Partners and recognition

Questions, answered.

What does Plurall AI do?

Plurall AI builds deepfake detection, fraud defense and proof-of-humanhood verification for trust, safety, security and compliance teams. Our models read forensic signals across image, video, audio and identity and return a verdict with the evidence behind it.

Which kinds of media can it check?

Four models cover four kinds of media: GaussMass for images, Influx for video, Sage for live camera feeds, and Timbre for audio, which is coming soon.

What do I get back from a scan?

A verdict, a calibrated likelihood, the evidence by category, the media hash and timestamps. It is decision support with the evidence attached, so a reviewer can see why something was flagged.

What happens when the model is not sure?

It says so. If the likelihood is too close to the middle, or the evidence pathways disagree, the system returns review-required instead of a verdict. Borderline items go to a reviewer, not to a guess.

How are the accuracy figures measured?

By named operational slice, never as one pooled number. Training and measurement never share a person, a generator family or a time period, and every reported rate carries a 95% Wilson interval.

How quickly do you adapt to a new generator?

The network is small enough to fine-tune end to end in four to eight hours on one GPU. When a new generator family appears, a slice is added, the model is retrained, every slice is re-scored, and the release ships within 24 to 48 hours.

How do I try it?

Start a free account, or tell us about your media and volumes and we will set up a pilot. Developers can go straight to the API reference.

Blog and press.

See what your team is missing.

Portrait of Carl Friedrich Gauss, oil on canvas, painted by Christian Albrecht Jensen in 1840.
Carl Friedrich GaussThe normal distribution and least squares. Our image model, GaussMass, is named for him.Christian Albrecht Jensen, 1840
Engraved portrait of Joseph Fourier by Julien-Léopold Boilly, early nineteenth century.
Joseph FourierFrequency analysis. The spectral evidence reads the periodic traces a generator leaves behind.Julien-Léopold Boilly, early 19th century
Portrait of Pierre-Simon Laplace, oil on canvas, painted by Paulin Guérin in 1838.
Pierre-Simon LaplaceThe probability of causes. Every signal updates one calibrated likelihood.Paulin Guérin, 1838
The full mathematical lineage →