# EU AI Act Article 50 — Transparency Guide

**Version:** 1.0  
**Effective:** 2 August 2026  
**Owner:** Legal / Product — Touch2Sign Ltd  
**Audience:** Customers (deployers), procurement, compliance reviewers  
**Internal analysis:** [../EU_AI_ACT_ARTICLE_50.md](../EU_AI_ACT_ARTICLE_50.md)

---

## Purpose

This guide explains how Touch2Sign meets the **transparency obligations** in Article 50 of the EU AI Act (Regulation (EU) 2024/1689), and what customers should do when using Aria AI and Sentinel AI features.

It is **not** a claim that every Touch2Sign workflow is a high-risk AI system under Annex III. It covers **transparency** for AI features that interact with people or generate content.

---

## What Article 50 requires (summary)

1. **Tell people** when they are interacting with an AI system, unless obvious from the context.
2. **Mark synthetic content** (AI-generated or manipulated text, image, audio, or video) so it is detectable in a machine-readable way, where technically feasible.
3. Extra rules apply to emotion recognition, biometric categorisation, deepfakes, and certain public-interest text — **Touch2Sign does not offer those features**.

Obligations apply from **2 August 2026**. A limited grace period to **2 December 2026** may apply only to Article 50(2) marking for generative systems already on the market before 2 August 2026. Touch2Sign implements marking for Aria drafts from the effective date.

---

## AI features in Touch2Sign

| Feature | What it does | Who sees it |
|---------|--------------|-------------|
| **Aria — draft** | Generates a first-draft document from a natural-language prompt | Senders |
| **Aria — signing analysis** | Pre-send analysis: reading time, key clauses, risk flags, signer briefing | Signers (briefing) |
| **Aria — summary & Q&A** | Plain-language summary and questions about the document during signing | Signers |
| **Sentinel — risks** | Extracts and classifies risks from signed / stored documents | Customer operators |
| **Sentinel — Ask / search** | Natural-language Q&A across contracts | Customer operators |

All AI processing for these features runs on **AWS Bedrock in `eu-west-1`** (Claude). Customer document content is **not** used to train foundation models. See the [Data Processing Agreement](/legal/dpa) and [sub-processor register](/legal/policies).

---

## How Touch2Sign complies

### Informing people (Article 50(1))

- **Signers** see a platform **AI & Signing Analysis Disclosure** in the consent step (alongside ERSD / Part 11 where applicable). Expanding it is recorded in the audit trail.
- **Signer briefing** states that analysis is powered by Touch2Sign AI and is guidance only.
- **Senders** see clear Aria branding when drafting with AI.
- **Sentinel users** see an on-screen notice that they are interacting with an AI system.

### Marking AI-generated drafts (Article 50(2))

When Aria generates a document draft, Touch2Sign embeds:

- A **machine-readable HTML comment** (`t2s-ai-generated:1` …), and  
- A **visible banner** with `data-ai-generated="true"` stating the content was artificially generated.

Senders should **review and edit** AI drafts before sending. After human editing, the banner may be removed; the signed PDF reflects the final human-approved content.

### What we do not do

- Emotion recognition or biometric categorisation AI  
- Deepfake image/audio/video generation  
- Automated legal advice or binding validity opinions  

---

## Customer (deployer) responsibilities

1. **Decide** whether your organisation enables Aria / Sentinel for your workflows.  
2. **Train staff** that Aria and Sentinel are AI systems; do not present AI outputs as human legal advice.  
3. **Review AI drafts** before sending for signature.  
4. **Inform your own users** (internal policies / privacy notices) that Touch2Sign AI features may analyse or generate document-related text when enabled.  
5. For **public procurement**, consider the [EU model contractual AI clauses](https://living-in.eu/group/16/commitments/standard-eu-model-contractual-ai-clauses) (linked from our clause source catalog).  
6. Keep AI **disabled** where your sector policy forbids third-party AI on document content.

---

## Evidence you can show auditors

| Evidence | Where |
|----------|--------|
| Signer AI disclosure + acknowledgement | Signing ceremony · audit events `ai_disclosure_expanded` / ERSD consent |
| AI draft marking | HTML source of Aria drafts (`data-ai-generated`, `t2s-ai-generated` comment) |
| DPIA covering AI | [/legal/policies/dpia](/legal/policies/dpia) |
| Sub-processors | [/legal/dpa](/legal/dpa) |
| Internal inventory | Programme hub · EU AI Act Article 50 analysis |

---

## Contact

Questions: **legal@touch2sign.com** · Privacy: **privacy@touch2sign.com**  
Help centre: [/help](/help) (article: *AI transparency — EU AI Act Article 50*)

---

*Last updated: 2 August 2026*