Transparency and evidence infrastructure for AI-generated content

Images & audio first<br>Ship AI features without building your own transparency infrastructure.

Check whether the images and audio leaving your pipeline actually carry provenance<br>marking — and keep dated, tamper-evident proof that you checked.

Run the free checker<br>Join the waitlist

Open source, MIT. No account, no sign-up, no dependencies.

Available now

The checker, running

One command against your own files. Below is the real output of the command,<br>copied from a terminal — not a mockup, not a rendering.

zsh — ai-act-verify

$ npx ai-act-verify sample-with-manifest.png sample-without-manifest.png

sample-with-manifest.png (PNG, 418.1 KB)<br>C2PA manifest PRESENT — PNG 'caBX' chunk<br>manifest size 18.0 KB<br>XMP present<br>validation NOT PERFORMED — presence is not validity

sample-without-manifest.png (PNG, 287.1 KB)<br>C2PA manifest ABSENT<br>XMP absent

Summary 1 with a manifest · 1 without

This tool detects whether a manifest is present, not whether it is valid:<br>signature, trust chain and integrity are not verified.<br>A manifest present does not prove compliance with Article 50 of the EU AI Act,<br>and its absence does not prove a breach: the Commission Guidelines place several<br>kinds of output outside the marking obligation. Technical tool, not legal advice.

Verbatim terminal output, including the disclaimer the tool prints itself. Everything in this section exists and runs today.

Install<br>npx github:lucapersichini/ai-act-verify ./your-image.png<br>Publishing to npm shortly — until then this runs it straight from the source repository.

Reads the container

JPEG, PNG, WebP, WAV, MP4/MOV/HEIC/AVIF, MP3, SVG. Anything else returns UNSUPPORTED rather than a guess.

Fits your CI

--json for scripts, --strict to fail a build when an asset ships unmarked.

Says what it doesn't know

There is no VALID status, and a test enforces that. Presence is never reported as validity.

Not built yet

What we're building next

The checker answers a question once. The product below keeps the answer — dated,<br>ordered and hard to alter. It does not exist yet.

Concept — coming soon

AI Act Verify<br>Evidence<br>August 2026

EVENTS THIS MONTH<br>12,481<br>+18% vs July

IN SCOPE<br>8,204<br>4,277 out of scope

VERIFIED<br>7,956

Recent evidence

EVENT<br>TYPE<br>PROVIDER<br>POLICY<br>DECISION<br>MARKING

evt_a3f9<br>image/png<br>Anthropic<br>v1.2

PASS<br>VERIFIED

evt_b7c2<br>text/plain<br>OpenAI<br>v1.2

OUT OF SCOPE<br>Guidelines (68)

evt_c1d4<br>audio/wav<br>self-hosted<br>v1.2

ACTION REQUIRED<br>ABSENT

Scroll sideways to see the full illustration →

This is an illustration, not a screenshot. Nothing above is clickable,<br>nothing is connected to anything, and no such dashboard exists today. It shows where we<br>intend to go: an append-only evidence log of every check, a policy engine that evaluates<br>each output on its own — including telling you when no obligation applies — and a record<br>you can still read a year from now.

Evidence over time

Append-only, hash-chained records of what was checked and when. Hashes only — never your prompts or outputs.

Policy engine

Per-output evaluation, versioned against the Commission Guidelines, so a rule change does not mean a code change.

In your stack

SDK wrapper and HTTP middleware, so the check happens where you already ship — not in a separate console.

What we will not claim

We do not make anyone "legally compliant." We provide technical controls and evidence. Whether that satisfies an obligation is a legal question, and it stays yours.

Present is not valid. Detecting a manifest is not validating its signature and trust chain. We label the two differently, always.

Text is not ready. Text detection APIs are not publicly available today, so for text we could record a dated assertion — not a verification. We start where verification is real: images and audio.

Not every output needs marking. The Commission Guidelines place source code, machine-to-machine exchanges, short sequences and standard editing outside the obligation. A tool that flags everything is wrong.

Early access

Premium access

The checker stays free and open source. The evidence layer is what we intend to charge for.

€199<br>/ month<br>Intended early-access price

Not billable yet — nothing is charged today, and no payment details are collected anywhere on this site.

Verification of upstream marking on images and audio

Dated, hash-chained evidence records

Per-output scope evaluation, including "no action needed"

SDK and HTTP middleware

Register your interest and we'll come back to you personally when early access opens.

Name or company

Email *

What would you use it for? optional<br>What are you generating, at what volume, and who is asking you about it?

I confirm my interest in being an early user of AI Act Verify when it launches.<br>This is not a binding purchase commitment.

Confirm interest

Interest confirmed. We'll come back to you personally — expect a real reply, not an autoresponder.

Waitlist

Tell us what you're actually dealing...