The evidence
behind every verdict.
Forensa detects AI-generated text and images and reports a likelihood, with a confidence interval and source-model attribution on every result. So your office can defend a finding, not just flag it.
Watch the instrument work.
A looping demo of a real scan: the document is read sentence by sentence, the signal bars settle as evidence accrues, and when the scan locks, the result speaks for itself, with every metric on the page.
The notion of authenticity in academic discourse has undergone a quiet but profound transformation in the past decade.
Where attribution once depended on stylometric instinct and a footnote convention, it now contends with the unbounded fluency of generative systems.
These systems do not so much imitate prose as approximate the statistical contours of it, producing artefacts that are recognisably literate and yet rarely surprising.
When one examines a body of suspected output, the absence of error often becomes its most telling signature.
Sentences arrive with a kind of evenness, neither hesitant nor exuberant, calibrated to a register that admits no idiolect.
It is this calibration, more than any single vocabulary choice, that the perplexity score is designed to detect.
—%
Awaiting
Demo content. The looping demo uses representative numbers — real scans use the production detector ensemble (roberta-v2.3 for text, image-forensics-v1 for image, audio-ensemble-mfcc and video-multi-signal-5ch in beta). Uploads are scanned in-memory and never retained.Run your own scan →
Evidence-first, by design.
Forensa is built for the moment a finding is challenged: every verdict comes with the uncertainty, the source, and the line-by-line evidence behind it.
Built for every file your office sees
Text and image detection today; audio and video in development. One pipeline, one verdict format, one PDF.
Likelihood, with uncertainty
Every finding ships with a confidence interval an officer can defend. ~85% on clean GPT text, ~60% on paraphrased text (2026-04-22 benchmark of record; methodology in docs).
Attribution, not just detection
Forensa indicates which model likely produced the content (GPT-class, Claude-class, image diffusion) so the finding earns context, not a bare yes/no.
An artifact you can defend
A signed PDF with the verdict, its interval, source attribution, sentence-level evidence, detector version, and chain-of-custody hash. Ready for the hearing.
Built for hearings, not headlines.
Text and image today.
Audio and video in development. Every modality returns a probability with a confidence interval, never a bare yes/no.
Evidence by the line.
Highlight the spans most likely to be AI-generated. Officers can read the same evidence the model used.
Which model wrote this.
Forensic fingerprinting indicates the likely generation source (GPT, Claude, Gemini, image diffusion families).
Uploads, in-memory only.
Student work is scanned in-memory and discarded. We never retain, log, or train on uploaded content. DPA available on request.
Defensible artifact.
Signed PDF with verdict, interval, attribution, signal breakdown, and detector version, ready for disciplinary review.
Fits the workflow.
POST /api/v1/analyze with a single call, or use the dashboard for ad-hoc reviews. SSO, audit log, and admin roles on Enterprise.
Transparent pricing.
Start free, upgrade when you need higher volume or institutional features.
Free
10 scans / month
- +Text + image detection
- +Forensic report (PDF)
- +Confidence intervals
- +Anonymous trial without account
Pro
300 scans / month
- +Text + image today; audio + video as released
- +Source-model attribution
- +Sentence-level highlighting
- +API access
- +Scan history
Enterprise
Unlimited scans
- +All Pro features
- +SSO + SAML
- +Audit log + admin roles
- +Data Processing Agreement (FERPA / SOC2 frameworks)
- +Dedicated support + SLA