AI transparency¶
Last updated: 1 November 2026
If a message told you "You are interacting with an AI system", this page explains what that means, where your messages go, and how to check whether content was generated by AI.
You are talking to software¶
Vornik is an AI agent platform. On Telegram, Slack, email, GitHub comment threads and the web chat, your messages go to a large language model, and the reply is generated by that model. No person reads or writes the replies as they happen. An operator configures the system, can review what it did and can intervene.
This notice is given under Article 50(1) of the EU AI Act (Regulation (EU) 2024/1689).
The AI acts and can be wrong. It runs tools, reads and writes files, opens pull requests, sends messages and searches a memory of past work. Language models write fluently whether or not a claim is true. Vornik scans output, detects some hallucinations and asks a person to approve sensitive actions; none of these checks is perfect. Do not treat its output as professional advice (legal, medical, financial or other); verify anything that matters.
Human oversight.
- Sensitive and irreversible actions need a person's explicit approval.
- Every tool call is recorded in an audit log.
- The operator can pause, cancel or steer any task while it runs.
- AI-authored content for public repositories passes a human review before it is merged.
Where your messages go¶
To generate a reply, the conversation is sent to a model provider chosen by the operator of the deployment:
| Provider | Role |
|---|---|
| Anthropic (Claude) | Language model inference |
| OpenAI | Language model inference |
| Amazon Web Services (Bedrock) | Language model inference |
| OpenRouter | Routing to further providers |
| Google Cloud (Vertex AI) | Language model inference |
| Ollama Cloud | Language model inference, including image input |
| Local models (for example Ollama) | Inference with nothing sent outside |
A deployment that uses only local models sends nothing outside. Ask the operator which configuration applies. This list is also the sub-processor list for data protection (compliance status).
How AI-generated output is marked¶
Article 50(2) of the EU AI Act requires the provider of an AI system to mark the text, audio and images it generates in a machine-readable form, and to let the people who receive it check that mark. This section says how Vornik does that, what it does not do, and how you can check a piece of content yourself.
What is marked¶
Natural-language output that a person reads or hears:
- chat and channel replies (web chat, Telegram, Slack, the chat API);
- emails;
- pull-request descriptions, reviews and comments, and issue comments;
- generated documents and reports (PDF, Word, HTML);
- files agents produce, including any generated image;
- synthetic voice replies, which are off unless the operator turns them on.
Not marked, because the law does not require it: source code and configuration, short labels and captions, messages from one AI agent to another, and an agent's intermediate steps.
How it is marked¶
Each mark is a small signed statement: the content with this SHA-256 was
generated by an AI system operated with Vornik, on this date. It carries the
IPTC source type trainedAlgorithmicMedia and is signed with an Ed25519 key
held by the deployment. Every mark is also recorded in the deployment's own
ledger.
| Where the output appears | Where the mark is |
|---|---|
| Chat API and Ollama-compatible API | AI-Disclosure: ai-generated and X-Vornik-AI-Mark response headers and the ai_generated field of the response; a streamed Ollama reply has no headers and carries the field on its final line |
| Web chat and file downloads | the same response headers |
Auto-Submitted and X-AI-Generated headers |
|
| GitHub | a hidden comment at the end of the text, <!-- vornik-ai-mark: … --> |
| Slack | message metadata (vornik_ai_mark) |
| Telegram | the ledger only: a Telegram message has nowhere to carry a mark |
| Documents | the file's own metadata: HTML meta tags, PDF Subject and Keywords, Word document properties |
| Agent files and voice replies | the ledger, recorded over the exact bytes |
Vornik does not follow the voluntary Code of Practice on marking and does not claim its presumption of conformity. The Code asks for an invisible watermark as well as metadata. The law does not require one, and Vornik cannot add one to text: a text watermark is applied by the model while it writes, and Vornik calls models through their providers' interfaces. None of the model providers Vornik uses documents a watermark that Vornik could detect or rely on. Vornik meets the law with signed metadata where the channel can carry it, and a ledger record where it cannot.
The limits¶
- Marks can be removed. Metadata is stripped by copying the text, forwarding a mail through some systems, or re-saving a file. The hidden GitHub comment disappears if someone edits it out.
- An edited copy is not recognised. The ledger matches the exact bytes. Changing a single character, including whitespace or line endings, means the lookup finds nothing. A signed mark still verifies, but it no longer matches the edited text.
- The ledger belongs to the deployment. Each organisation running Vornik keeps its own ledger and keys. The Vornik project cannot check content produced by someone else's deployment.
- Audio has no embedded watermark. That is why voice replies are off by default.
We will revisit this when a standard way to fingerprint or watermark text appears, when a model provider documents a watermark Vornik can rely on, or when a harmonised standard for Article 50(2) is published.
Check a piece of content yourself¶
The deployment that produced the content answers three public requests. None of them needs an account, and none returns the content, the project or the person it was for.
Was this exact text generated here? Hash it and look it up. Hash the
exact bytes you received: printf '%s' adds nothing, but echo adds a
newline, and a hash with one extra character finds nothing.
sha=$(printf '%s' "$TEXT" | sha256sum | cut -d' ' -f1)
curl -s "https://<vornik-host>/provenance/$sha"
# {"found":true,"statement":"Generated by an AI system operated with Vornik on 2026-11-01.","issued":"2026-11-01"}
# or, when there is no record:
# {"found":false,"statement":"No record of this content."}
Is this mark genuine? Send the mark from a header, a hidden comment or a document's metadata. Add the content's hash to check that the mark belongs to it:
curl -s -X POST "https://<vornik-host>/provenance/verify" \
-H 'Content-Type: application/json' \
-d '{"mark":"<the mark>","sha256":"<the content hash>"}'
# {"valid":true,"matches":true,"statement":"…","source_type":"trainedAlgorithmicMedia",…}
A mark that does not verify gets {"valid":false,"statement":"This mark does
not verify."}, with no detail of what failed.
Check it without asking the deployment. The public keys are at
https://<vornik-host>/.well-known/vornik-ai-mark-keys, each with its key id
and the period it signed in. A mark is base64url(payload).base64url(signature):
the signature is Ed25519 over the decoded payload bytes, and the payload's
kid names the key.
Your data and your rights¶
The organisation running the deployment is the data controller, and the only party that can answer a request about your data. The Vornik project does not hold, store, or have access to any data from a customer's deployment, so it cannot look you up or act on a request. If you do not know who runs the deployment, the message or document that brought you here usually says: the sending domain, the signature, or the person who contacted you.
Under the GDPR you can ask the controller for a copy of your data (Article 15), correction (16), erasure (17), restriction (18) and portability (20), and you can object to processing based on legitimate interests (21). If the controller does not respond satisfactorily, you can complain to your national supervisory authority directly.
If a deployment holds data about you that someone else put there (an email that names you, a CV about you, a photo you are in), GDPR Article 14 obliges the controller to tell you what they hold, why, on what legal basis, who else sees it, how long they keep it and where it came from, even though you never used their system. Two things Vornik does help them answer:
- it records where each piece of data came from, so a controller can answer "where did you get this about me?";
- its image analysis refuses to identify people, infer emotions, or deduce race, religion, politics, union membership, health, sex life or sexual orientation from appearance.
Operators: the Article 14 notice is yours to publish. A template is in the
project's docs/legal/art14-notice-template.md.
Questions¶
Contact the operator of the deployment you are using. For the Vornik project: [email protected].