Live
C2PA JUMBF Manifest
Cryptographic signatures embedded directly in image data — Adobe, OpenAI, Microsoft, Google. We scan the bytes themselves, not just XMP strings.
Confidence
0.97
Weight
1.00
Cost
free
Detection Stack
Ten layers, one verdict.
No single layer is flawless. Stacked, they deliver the most comprehensive AI inventory on the EU market. Our fingerprint database improves with every scan — this is the compounding moat.
Fingerprint database · Live
Loading …
Live
EXIF · XMP · IPTC
22 generator patterns in Software, CreatorTool, and Description fields. Effective even for generators without C2PA (Stable Diffusion, Midjourney v6).
Confidence
0.90
Weight
0.85
Cost
free
Live
Perceptual Hash DB
DCT-based pHash for every image. Lookup against our fingerprint database — improves with every scan. Detects re-uploads and crops in under 50ms.
Confidence
0.85+
Weight
0.95
Cost
free
Live
URL · Filename Heuristics
Fast first-pass detection via asset URL and filename. Low weight, extremely low cost — catches obvious cases.
Confidence
0.45
Weight
0.35
Cost
free
API-key ready
Hive Moderation ML
Pixel-level forensics for unsigned generators. ~85% recall on JPEG roundtrip. Active from Starter tier via HIVE_API_KEY.
Confidence
0.40–0.95
Weight
0.90
Cost
$0.01
API-key ready
Sightengine Ensemble
Second ML opinion as ensemble partner. In case of disagreement with Hive → human review queue. SIGHTENGINE_USER + SECRET.
Confidence
0.40–0.95
Weight
0.85
Cost
$0.008
Live
Text Heuristics (local)
Vocabulary, em-dash frequency, transition word density, sentence length uniformity — detects ChatGPT/Claude/Gemini output without API.
Confidence
0.40–0.85
Weight
0.55
Cost
free
API-key ready
GPTZero · Originality
External text detection API. Invoked only if local heuristics have already triggered (cost optimisation). GPTZERO_API_KEY.
Confidence
0.30–0.97
Weight
0.90
Cost
$0.02
Phase 3
Audio Forensics
ElevenLabs detection API + spectral analysis for AI voice. Sora/Veo frame forensics for AI video. Phase 3.
Confidence
—
Weight
0.85
Cost
—
Phase 3
Human-in-the-Loop
If two ML models disagree → a human decides within 5–10 seconds. Enterprise tier opt-in for “Verified Accuracy”.
Confidence
0.99
Weight
1.00
Cost
$0.10
You will see actual confidence values and hit rates per layer after each scan under “Detection Layer Breakdown” in the report.
Asset Types
Four media classes, one pipeline.
We scan everything that can be embedded in HTML — images, text, video, audio. Each class passes through the detection layers relevant to its format.
Images
Product photos, hero visuals, blog illustrations from Midjourney, DALL·E, Firefly, Stable Diffusion, Flux.
C2PA · EXIF · pHash · ML
Text
Blog articles, product descriptions, about pages — generated with ChatGPT, Claude, Gemini.
Heuristics · GPTZero
Video
Hero loops, explainer videos, animations from Runway, Sora, Veo, Kling, Pika.
URL · Frame sampling (Phase 3)
Audio
Voice-overs, podcast voices, AI voices from ElevenLabs, OpenAI Voice, Resemble.
Spectral forensics (Phase 3)
The EU AI Act in 60 seconds
Four key dates. Three sanction levels.
A factual summary of the transparency obligations under Article 50 EU AI Act and the sanctions framework from Article 99. Source-based. Not legal advice.
Status
In force
Article 50 (transparency obligations) has applied since 2 August 2026. Sanctions from 2 August 2027.
- 13 June 2024
Adopted
PublishedRegulation (EU) 2024/1689 adopted by the European Parliament and Council. Enters into force as directly applicable EU law.
- 2 Aug 2026
Transparency active
In forceArticle 50 becomes applicable. Deepfakes and AI-generated texts on public topics must be labelled. Chatbots must disclose themselves.
- 2 Dec 2026
Watermark spec
In forceProviders of GenAI systems (OpenAI, Adobe, Google) must embed machine-readable markers. Specific technical standards via delegated act.
- 2 Aug 2027
Full sanctions
Sanction riskFines may be imposed — up to a maximum of EUR 15 million or 3% of global group annual turnover (Art. 99). Warnings under unfair competition law (UWG) may be possible from the date of applicability.
Fine framework
What the sanctions catalogue contains.
Article 99 EU AI Act defines three levels. The higher of the two values always applies: fixed amount or percentage of global group annual turnover.
Level / Article
Fixed amount
Group turnover
Relevance
Art. 99 (3)
Prohibited practices
Highest level — prohibited practices under Art. 5 (social scoring, biometric real-time surveillance).
€35 million
7%
Not relevant for transparency violations.
Art. 99 (4)
Deployer obligations
Breaches of deployer obligations — including Art. 50.4 (deepfake labelling) and Art. 50.1 (chatbot disclosure).
€15 million
3%
Directly relevant for you as a website operator.
Art. 99 (5)
False information
False or incomplete information provided to market surveillance authorities upon request.
€7.5 million
1%
Avoidable through proper documentation.
Important
Maximum ≠ likelihood.
Authorities determine fines on a case-by-case basis — considering severity, intent, turnover, and demonstrated diligence. SMEs are generally sanctioned well below the maximum thresholds.
We state upper limits because they are public and you should assess risk realistically — not to predict a specific penalty. Anyone who does so is being misleading.
Assessment factors (Art. 99(7))
Severity
Type and duration, intent vs. negligence.
History
Repeat offence vs. first-time occurrence.
Economic advantage
Did the breach measurably increase conversion or reach?
Duty of care
Were technical measures taken? This is precisely where Provifai documents.
Sources: Regulation (EU) 2024/1689 (AI Act), Articles 50 and 99. Sanctions are enforced nationally by market surveillance authorities — in Germany, expected to be BNetzA + BSI. Not legal advice.
The Result
You see immediately what matters.
Your content, clearly categorised: What is AI, what is authentic, what awaits your decision. Technical evidence is provided for every finding — but it appears in the second line, not the first.
example-shop.com
Checked today · 12 subpages · 18 seconds
- 37
- checked
- 13
- AI detected
- 6
- pending
LabelledProduct image “Summer Campaign”
Created with AI — evidenced within the image itself
PendingBlog cover image
Very likely AI — not yet labelled
- Pending
Text “About Us”
Likely written with AI
AuthenticRetail store, entrance area
Authentic photo — camera data present
- Deepfake
Video testimonial
Face artificially generated — labelling required
You can confirm or dispute each finding with a single click — your input is decisive, not our estimate.
Demo with sample data. All visuals AI-generated and declared in our certificate .
Ready?
Enough theory. Discover what’s running on your domain.
Article 50 has applied since 2 August. The audit takes less than a minute.