The AI Act: who must do what, and when.
The AI Act is the EU's law on artificial intelligence. It is not the GDPR, which still governs your personal data. It splits duties between the companies that make the models, the services built on them — Evanesca is one — and the people and companies that use them at work. Here is each share, where Evanesca stands, and the dates.
Texts read on EUR-Lex on 6 Oct 2026
Three roles, three sets of duties
Document their models
The companies that make the models — Mistral AI, DeepSeek, Zhipu / Z.ai, Qwen, Swiss AI Initiative, Google, OpenAI and Meta — are providers of general-purpose AI models. They document their models, follow a copyright policy and publish a summary of their training content (art. 53).
Say it is AI, and mark it
Evanesca puts these models — and those that make its images — in an agent and a chat under its own name: in the Act's words, it is the provider of an AI system built on models made by others (art. 3(3) and 3(68)). Article 50 is its own duty.
It depends on the use
For yourself, outside any professional activity, the Act asks nothing of you (art. 3(4)). At work, you are a “deployer”, with a few duties, set out below.
Article 50: Evanesca's duty, and where it stands
Since 2 August 2026, two transparency rules apply to services like Evanesca. People must be told that they are interacting with an AI system, at the latest at their first interaction (art. 50(1) and 50(5)). And the outputs — text, images, audio or video — must be “marked in a machine-readable format and detectable as artificially generated or manipulated” (art. 50(2)).
Regulation (EU) 2026/1744, the “Digital Omnibus on AI”, gave services already on the market until 2 December 2026 to apply the second rule. Evanesca opens after 2 August 2026: both rules apply to it from its first day.
| The rule | What Evanesca does | Status |
|---|---|---|
| Tell people they are talking to an AI (art. 50(1), 50(5)) | The chat and Agent mode say they are an AI system before your first message, and every answer is labelled as AI-generated. | Built in |
| Mark generated text in a machine-readable way (art. 50(2)) | The route set out in the Commission's Code of Practice on AI-generated content, published on 10 June 2026: an imperceptible watermark on answers longer than 200 tokens, and signed, time-stamped metadata in exported files. No log of answers: the Code does not consider logging alone sufficient, and a log would break “no copy”. The Word, PDF and Excel documents the agent writes already carry “AI-generated” in their document properties; the .md and .csv files are not marked yet. | In progress |
| Mark generated images in a machine-readable way (art. 50(2)) | The images made by FLUX.1 [schnell] or PhotoMaker V2 — in the chat, Agent mode and the API — carry IPTC metadata (XMP, digital source type “trainedAlgorithmicMedia”, description “AI-generated with Evanesca”) marking them as AI-generated. It is metadata, not a watermark: a screenshot or a re-encode can remove it. | Built in |
Text you select and copy yourself cannot carry a reliable mark with today's techniques; the copy button and exports will carry it.
Image generation also meets two bans that Regulation (EU) 2026/1744 added to article 5, applying from 2 December 2026: AI that generates intimate images of an identifiable person without their consent, or child sexual abuse material (art. 5(1)(ba) and (bb)). Evanesca's terms forbid both uses.
If you use Evanesca at work
Then you are a deployer. Your duties are few, but real:
- Train your team. Article 4 asks providers and deployers to “take measures to support the development of AI literacy of their staff”, without requiring any particular level (wording of Regulation (EU) 2026/1744; the duty has applied since 2 February 2025).
- Say when you publish AI text. If you publish text generated with Evanesca “with the purpose of informing the public on matters of public interest”, say that it was generated by AI — unless a person reviewed it and someone holds editorial responsibility for it (art. 50(4)).
- Say when you publish a deep fake. If you publish an image made with Evanesca that resembles existing people, objects, places or events and could falsely pass for authentic — a “deep fake” (art. 3(60)) —, disclose that it was generated by AI (art. 50(4)). For an evidently artistic, creative, satirical or fictional work, saying so in a way that does not spoil the work is enough.
- Keep it away from high-risk uses. Evanesca is not intended for recruiting or selecting people, for decisions about workers — promotion, dismissal, allocating tasks, monitoring or evaluating performance —, for assessing creditworthiness, or for pricing life and health insurance (Annex III, points 4 and 5). The rules for those uses apply from 2 December 2027; turning a general-purpose system to one of them can make you the provider of a high-risk system (art. 25(1)(c)). Our terms say so (art. 25(2)).
- Never for a banned practice. The practices article 5 bans — among them manipulating people's behaviour, social scoring and inferring emotions at work or at school — have been banned since 2 February 2025.
- The GDPR still applies. If you paste other people's personal data, you remain responsible for it; a Business workspace gives you a data processing agreement (DPA).
What the model makers must do
Since 2 August 2025, the makers of general-purpose AI models must (art. 53(1)): keep technical documentation of the model for the authorities; give the providers that build on it — Evanesca among them — the information they need; put in place a policy to comply with EU copyright law; and publish a sufficiently detailed summary of the content used to train it. Models already on the market before that date have until 2 August 2027 (art. 111(3)).
Open weights are not a free pass. A maker that releases a model under a free and open-source licence, weights included, is spared the first two duties — never the last two, and never for a model with systemic risk (art. 53(2)). Systemic risk is presumed when training used more than 10²⁵ floating-point operations (art. 51(2)); such models carry extra duties: evaluation and adversarial testing, risk mitigation, reporting serious incidents and cybersecurity (art. 55).
These duties are the makers', not Evanesca's: Evanesca does not develop, train or modify the models. Whether a licence counts as “free and open-source”, and whether a model carries systemic risk, is for the maker to establish, under the Commission's supervision. What Evanesca does is lead you to each maker's own documentation.
| Model | Maker | Licence | Official card |
|---|---|---|---|
 Mistral Large 4 | Mistral AI | — | Read the card |
 DeepSeek V4.1 Flash | DeepSeek | MIT | Read the card |
 GLM-5.3 | Zhipu / Z.ai | GLM-5.3 License | Read the card |
 Qwen3.8 27B | Qwen | Apache-2.0 | Read the card |
| DeepSeek V4 Flash | DeepSeek | MIT | Read the card |
| Apertus v1.5 70B | Swiss AI Initiative | Apache-2.0 | Read the card |
| GLM-5.2 | Zhipu / Z.ai | MIT | Read the card |
| Qwen3.6 35B | Qwen | Apache-2.0 | Read the card |
 Gemma 4 26B | Apache-2.0 | Read the card | |
| Mistral Small 4 | Mistral AI | Apache-2.0 | Read the card |
| Qwen3.5 397B | Qwen | Apache-2.0 | Read the card |
 GPT-OSS 120B | OpenAI | Apache-2.0 | Read the card |
| Qwen3 Coder 30B | Qwen | Apache-2.0 | Read the card |
| Qwen3 235B | Qwen | Apache-2.0 | Read the card |
| Mistral Small 3.2 | Mistral AI | Apache-2.0 | Read the card |
 Llama 3.3 70B | Meta | Llama 3.3 Community License | Read the card |
| Pixtral 12B | Mistral AI | Apache-2.0 | Read the card |
The dates
| Date | What applies | Reference |
|---|---|---|
| The AI Act enters into force, after its publication in the Official Journal on 12 July 2024. | Art. 113 | |
| Banned practices (art. 5) and AI literacy (art. 4). | Art. 113(a) | |
| Rules for general-purpose AI models (chapter V), governance and penalties. | Art. 113(b) | |
| Regulation (EU) 2026/1744, the “Digital Omnibus on AI”, enters into force (published on 24 July 2026). | Reg. 2026/1744, art. 4 | |
| General application, including the transparency rules of article 50 — for Evanesca, from its first day. | Art. 113 | |
| Machine-readable marking (art. 50(2)) for systems placed on the market before 2 August 2026. Two new bans: AI that generates intimate images of an identifiable person without consent, or child sexual abuse material (art. 5(1)(ba) and (bb)). | Art. 111(4) and 113(a), as amended | |
| General-purpose AI models placed on the market before 2 August 2025 must comply. | Art. 111(3) | |
| High-risk rules for the uses listed in Annex III — employment, credit, insurance and education among them. | Art. 113(c), as amended | |
| High-risk rules for AI in the products covered by Annex I, such as toys and medical devices. | Art. 113(c), as amended |
Sources: Regulation (EU) 2024/1689 and Regulation (EU) 2026/1744, which amends it, read on EUR-Lex on 6 Oct 2026.
What Evanesca does not claim
Evanesca does not call itself “AI Act compliant”: compliance depends on the model, the service and the use, and a label we gave ourselves would prove nothing. This page says instead who must do what, what Evanesca does about its share, and what is still in progress. The checks run on each model are on the transparency page.
Questions about the AI Act
Do I have duties under the AI Act if I use Evanesca for myself?
No. The Act calls someone a “deployer” when they use an AI system under their authority, “except where the AI system is used in the course of a personal non-professional activity” (art. 3(4)). Personal use carries no duty; article 50 is Evanesca's.
Is Evanesca “AI Act compliant”?
Evanesca does not use that label: compliance depends on the model, the service and the use. Its own duty is article 50 — telling you that you are talking to an AI, and marking generated text and images in a machine-readable way. The table on this page shows where each stands.
Are the images Evanesca generates marked as made by AI?
Yes: generated images — in the chat, Agent mode and the API — carry IPTC metadata marking them as AI-generated (art. 50(2)). It is metadata, not a watermark: a screenshot or a re-encode can remove it. If, at work, you publish an image that could pass for real, article 50(4) asks you to say that it was generated by AI.
Are open-weight models exempt from the AI Act?
Partly, and not all of them. A maker that releases a model under a free and open-source licence, weights included, is spared two documentation duties, but still owes a copyright policy and a public summary of its training content — and a model with systemic risk gets no exemption at all (art. 53(2)).
Can I use Evanesca to screen job candidates?
No. Recruitment and decisions about workers are high-risk uses under the AI Act (Annex III, point 4), with rules applying from 2 December 2027. Evanesca is not designed or intended for them, and its terms say so.