AI Transparency Statement

1. What this Statement covers

We publish this Statement so that anyone using Voistral, or exposed to text produced with it, can understand where AI is involved and what that means for them. It also serves as our public record of how we approach the AI Act, and it is the document we would provide to a market surveillance authority on request.

This Statement covers the Voistral desktop applications for macOS and Windows, our websites and web dashboards, and the related services (together, the “Services”). It does not cover third-party products you choose to use alongside Voistral, including the applications you dictate into.

2. Our role under the AI Act

2.1 We are a provider of an AI system. We develop Voistral and make it available in the European Union under our own name and trademark. That makes us a provider of an AI system within the meaning of Article 3(3) of the AI Act, and it is the basis on which the obligations described in this Statement apply to us.

2.2 We are not a provider of a general-purpose AI model. We do not train, develop or publish AI models. Voistral is built on models operated by the third-party providers listed on our Sub-processors page — currently Deepgram for speech-to-text, and Anthropic and OpenAI for text generation. The obligations that the AI Act places on providers of general-purpose AI models sit with those providers, not with us.

2.3 Voistral is not a high-risk AI system. Voistral is a dictation and writing tool. It is not used as a safety component of a regulated product, and it does not perform any of the functions listed in Annex III to the AI Act — it makes no decisions and no assessments about any person. We therefore do not have the obligations that apply to high-risk AI systems. We reassess this whenever we plan a material new feature.

2.4 Voistral is not an emotion recognition or biometric categorisation system. Voistral converts speech into text. It does not identify or verify who is speaking, infer emotions or states of mind, or sort people into categories from their voice or any other biometric characteristic. The transparency obligation in Article 50(3) of the AI Act therefore does not apply to us, and our Acceptable Use Policy prohibits using the Services for those purposes.

2.5 Your role. If you use Voistral in the course of a professional, business, trade or freelance activity, you are a deployer of an AI system under Article 3(4) of the AI Act, and the duties in Section 7 may apply to you. If you use Voistral purely for personal, non-professional purposes, those deployer duties do not apply to you under Article 2(10) of the AI Act. Our own obligations as provider apply either way.

3. How Voistral uses AI

Voistral is a cloud service and requires an internet connection. Transcription and AI processing are not performed on your device. The Services use AI in the following ways.

• Transcription. While you run a dictation session, your microphone audio is streamed to our speech-to-text provider and converted into text.
• Polish. A language model cleans up the transcript. What it does depends on the tone you choose: the Natural tone removes filler words, false starts and stutters and fixes grammar and punctuation while keeping your wording, whereas the other tones (Formal, Casual, Concise, Technical, Bullets and Custom) deliberately change register, structure or length.
• Enhance. A language model rewrites and restructures your polished text into a finished piece suited to the application you are dictating into, and can expand a short dictation substantially. You control its behaviour through your persona settings.
• Expansions. When you speak a trigger phrase you have defined, Voistral substitutes longer text in its place. Static expansions insert text you wrote yourself and involve no model. AI and Hybrid expansions have a language model generate the replacement text from your instructions.
• Supporting features. Optional features such as custom vocabulary, entity detection and the redaction filter shape how the above operate. They do not create new content.

Which models are available to you depends on your plan, and you can see and choose them in your settings. Voistral never listens outside a dictation session you start, and it does not read your screen, your window titles or the contents of other applications — see Section 3.10 of our Privacy Policy.

4. Which outputs are AI-generated

Article 50(2) of the AI Act applies to AI systems that generate or manipulate synthetic text. It does not apply where the system performs an assistive function for standard editing, or does not substantially alter the input or its meaning. We have assessed each Voistral feature against that test, following the European Commission’s guidelines on Article 50 of 20 July 2026.

4.1 Outputs we do not treat as AI-generated content.

• Your raw transcript. This is a transcription of words you actually spoke. Transcription does not create synthetic content, and the Commission’s guidelines list transcriptions of conversations as outside the scope of the marking obligation.
• The Natural polish tone. This removes filler words and corrects grammar and punctuation while preserving your vocabulary, tone, meaning and approximate length. That is standard editing, which the AI Act expressly excludes.
• Static expansions. These substitute text you wrote yourself. No model is involved, so no content is AI-generated.
• Very short outputs. Short sequences of words, and source code, fall outside the marking obligation under the Commission’s guidelines.

4.2 Outputs we do treat as AI-generated or AI-manipulated content.

• Enhanced text. Enhance rewrites, restructures and can substantially expand what you dictated — producing, for example, a complete email with a subject line and sign-off, or a structured document from a couple of sentences. That goes beyond standard editing.
• AI and Hybrid expansions. A model generates the substituted text, so the substituted text is AI-generated.
• Polish tones other than Natural. Formal, Casual, Concise, Technical, Bullets and Custom change the register, structure or length of your text rather than merely correcting it, so we treat their output as AI-manipulated.

4.3 How we apply this. The assessment is made per output, not per feature, because the same feature can produce results on either side of the line — Enhance applied to a long, already-written passage makes only minor corrections, for example. Where an output could reasonably be characterised either way, we take the conservative view and treat it as AI-generated content. Whichever way a particular output falls, you own it, as set out in Section 5.2 of the Terms of Service.

5. Marking and detection of AI-generated text

5.1 The obligation. For the outputs identified in Section 4.2, Article 50(2) of the AI Act requires us to mark the output in a machine-readable format and to make it detectable as artificially generated or manipulated, using technical solutions that are effective, interoperable, robust and reliable. Article 50(2) requires this insofar as it is technically feasible, taking into account the specificities and limitations of the type of content, the cost of implementation, and the generally acknowledged state of the art.

5.2 When it applies to us. Voistral was placed on the Union market on 22 June 2026, before 2 August 2026. Under Article 111(4) of the AI Act, as amended by Regulation (EU) 2026/1744, the marking and detection obligation therefore applies to us from 2 December 2026. Every other obligation described in this Statement applies from 2 August 2026. Content generated before 2 August 2026 is not marked retroactively.

5.3 The constraint we have to work within. We want to be straightforward about this rather than imply more than we do. Voistral inserts free-form text at your cursor into whichever application you are using. Plain text carries no metadata, so there is nothing for a metadata mark to attach to once the text leaves Voistral. The remaining technique for text is an imperceptible watermark, which for text is most reliably applied inside the model that generates it. We are a downstream provider building on third-party models, and we cannot apply a model-level watermark unless the model provider makes one available to us.

5.4 What we are doing about it. By 2 December 2026 we will:

• record AI-generation status in digitally signed, time-stamped metadata wherever an output is delivered in a form that can carry metadata — including stored transcription history records, exports and API responses;
• keep provider-side records that let us confirm, on a substantiated request, whether a given piece of text was generated by Voistral and by which feature;
• adopt imperceptible text watermarking, and the corresponding means of detection, as and when our model providers make it available to downstream providers such as us; and
• maintain a written technical-feasibility and proportionality assessment recording what is available, what we have implemented, what we have not, and why — reviewed at least annually and whenever the state of the art changes.

5.5 Code of Practice. We are assessing whether to adhere to the Code of Practice on Transparency of AI-generated Content, which the European Commission and the AI Board assessed as adequate in July 2026. Adherence is voluntary; the obligations in Article 50 apply to us either way, and this Statement describes how we meet them.

5.6 Questions and requests. If you need to establish whether particular text was produced with Voistral, contact legal@voistral.com.

6. Knowing that you are using an AI system

6.1 We tell you plainly. Voistral is presented as an AI dictation product throughout our marketing, our documentation, our checkout, the desktop application and this Statement. You always start a dictation session yourself, you choose which AI features to apply, and the features that use AI are labelled as such in the product.

6.2 Voistral is not a chatbot or an assistant. Article 50(1) of the AI Act addresses AI systems intended to interact directly with people in a genuine two-way exchange, such as chatbots, voice assistants, avatars and AI agents. Voistral is not one of those. It transcribes and rewrites your own words: it does not answer you, hold a conversation with you, act on your behalf, or reply to what you dictate. If your dictation is phrased as a question, Voistral turns it into a better-written version of that question rather than answering it. We consider that Article 50(1) does not apply to the Services, and we provide the information in this Section regardless.

7. Your obligations when you publish AI-assisted text

Some AI Act duties fall on you as the deployer rather than on us. We summarise the ones that can arise from using Voistral so you can meet them. This is not legal advice, and you are responsible for your own compliance.

7.1 Publishing text on matters of public interest. Under Article 50(4) of the AI Act, if you publish AI-generated or AI-manipulated text for the purpose of informing the public on matters of public interest, you must disclose that the text has been artificially generated or manipulated. Matters of public interest include politics and democratic processes, public administration and services, the administration of justice and law enforcement, fundamental rights, public security, public health, environmental protection, consumer safety, and economic, financial, political, scientific or cultural developments that may be a relevant subject of public debate.

7.2 The human review exemption. That disclosure is not required where the publication has undergone human review or editorial control and a person holds editorial responsibility for it. Human review means a person with relevant knowledge and professional judgement deliberately examining the substance of the text. A spell-check, a grammar pass or another purely formal check is not enough.

7.3 What this means in practice. Ordinary working use of Voistral — emails, chat messages, notes, documents, issue reports, code prompts — is not text published to inform the public on a matter of public interest, so this duty does not arise. It becomes relevant if, for example, you publish news, public-health guidance, or reporting on a matter of public debate that was substantially written by AI and never reviewed on its substance by a person.

7.4 Deepfakes. The deepfake labelling duty in Article 50(4) applies to AI-generated or manipulated image, audio and video content. Voistral produces text only and has no capability to generate or alter images, audio or video, so that duty cannot arise from your use of the Services. We do not synthesise, clone or alter anyone’s voice.

7.5 Recording other people. Separately from the AI Act, laws in many places restrict recording or transcribing other people and may require everyone’s consent. You are responsible for obtaining any consent you need — see Section 5.5 of the Terms of Service and Section 2 of the Acceptable Use Policy.

8. Prohibited AI practices

Article 5 of the AI Act bans certain AI practices outright. Our Acceptable Use Policy prohibits using the Services for any of them, including biometric identification or categorisation, emotion recognition, social scoring, predictive policing, and manipulative or exploitative techniques. We do not build any of those capabilities into Voistral.

Regulation (EU) 2026/1744 added two further prohibitions, which apply from 2 December 2026: AI systems that generate or manipulate realistic non-consensual intimate material depicting an identifiable person, and AI systems that generate or manipulate child sexual abuse material. Because Voistral produces text only and cannot generate images, audio or video, it is not capable of producing that material. Our Acceptable Use Policy separately prohibits unlawful content, content that depicts, promotes or facilitates child sexual abuse or the sexual exploitation of minors, and content that infringes anyone’s privacy rights, and we enforce those prohibitions under Section 15 of the Terms of Service. You can report suspected misuse through our Report Abuse page.

9. Accuracy, human oversight and limitations

AI output can be wrong. Speech recognition can mis-hear you, and language models can produce text that is inaccurate, incomplete or misleading even when it reads fluently and confidently. Enhance in particular may add structure or phrasing you did not dictate. Always read Voistral’s output before you send, publish or rely on it.

Output is not professional advice, and you must not use the Services or their output as the sole or decisive basis for decisions with legal or similarly significant effects on people. The full disclaimers are in Section 6 of the Terms of Service, and the corresponding use restrictions are in Section 5 of the Acceptable Use Policy.

10. We do not train models on your content

We do not use your Customer Content to train any AI or machine-learning model of ours, and we configure our providers so that your content is not used to train theirs. Your microphone audio is not stored — it is discarded as soon as the transcript is ready. The detail, including provider retention periods, is in Section 6 of our Privacy Policy and Section 7 of the Terms of Service.

11. AI literacy

Article 4 of the AI Act requires providers and deployers to take measures to ensure a sufficient level of AI literacy among their staff and other people who operate and use AI systems on their behalf. We maintain an internal AI literacy record covering everyone who builds and operates Voistral, and we review it at least annually and whenever we ship a material change to an AI feature.

We also support your own AI literacy: our documentation explains what each AI feature does, and this Statement sets out where AI is involved and what its limits are.

12. Supervision, complaints and contact

12.1 Talk to us first. For any question, concern or complaint about how Voistral uses AI, contact legal@voistral.com or use our contact form at voistral.com/support/ticket. We would like the chance to resolve it directly.

12.2 Supervisory authorities. Compliance with Article 50 of the AI Act is enforced by national market surveillance authorities. We are established in Sweden, where the coordinating market surveillance authority and single point of contact for the AI Act is currently the Swedish Post and Telecom Authority (Post- och telestyrelsen, PTS, www.pts.se). For data-protection matters, the Swedish authority is the Swedish Authority for Privacy Protection (Integritetsskyddsmyndigheten, IMY, www.imy.se), as described in our GDPR page.

ZeroEnd Technology AB
Company registration number: 559540-1224
Registered address: Järpvägen 41 b, 756 53, Uppsala, Sweden
VAT number: SE559540122401
AI Act and legal matters: legal@voistral.com · Privacy: privacy@voistral.com
Support: voistral.com/support/ticket

13. Changes to this Statement

We will update this Statement when we add or change AI features, when our assessment in Section 4 changes, when our marking and detection measures change, and when the law or official guidance changes. When we do, we will update the “Last updated” date above, and where a change materially affects your rights or obligations we will give notice as described in Section 19 of the Terms of Service. We review this Statement at least annually.