AI Content Disclosure

AI-Generated Content Disclosure

Version: 2.0
Effective date: 3 August 2026
Provider: Pedro Cantueso García de Vinuesa (kreawave)
Contact: [email protected]

_This document explains how kreawave marks the content it generates, what our transparency duties are under the EU Artificial Intelligence Act, what your duties are when you publish that content, and — stated plainly — what synthetic voice can and cannot do. It is a transparency document: it does not create, extend or limit any commercial condition. Prices, allowances, charging rules and refunds are governed exclusively by the Terms of Service and the Refund Policy._

1. Scope

This document covers all content generated by the Service:

Audio (live today). Text-to-speech synthesis from the catalogue of pre-built synthetic voices.
Images and video (not available today). The image module and any future video module are not enabled and cannot be used. Every rule in this document is written so that it already applies to them the day they are switched on, without a new version being required. Where a rule is stated in the future tense, that feature does not exist yet.
Voice cloning (not available today). Uploading your own voice is disabled. Every voice in the catalogue is an original synthetic voice created by the Service; none is a clone of a real, identifiable person.

2. Everything the Service produces is artificially generated

All audio delivered by kreawave — including the free public demo, the free plan, previews and paid generations — is produced by an artificial intelligence model. It is not a recording of a human being. No human performer was recorded to produce your file, and the voice you hear does not correspond to any real, identifiable person.

Consequently:

The audio does not reflect real facts, opinions, endorsements or statements of any person.
It must never be presented as a genuine recording of a real individual, as testimony, as evidence, or as an official statement.
The Service states this at the point of exposure: the generation and home pages carry the notice "All audio is AI-generated and watermarked".

The same will apply to images and video when those modules exist.

3. How generated content is marked

The marking is applied by the Service, always, and cannot be switched off, opted out of, or purchased away by any plan.

3.1 Audio — two independent layers

Layer 1 — Imperceptible watermark inside the audio signal (primary layer).

Technology in production: AudioSeal (Meta AI), applied to every audio file the Service produces, at the engine's native sample rate (48 kHz).
It is inaudible under normal listening conditions and lives inside the waveform, not in the file metadata — so it survives renaming, metadata stripping, format conversion and lossy re-compression (verified across MP3 bitrates from 96 to 320 kbps).
It is fail-closed by design: a synthesis worker that cannot load the watermarking model refuses to start, and the assembled file is re-checked with the detector before delivery; if the mark is not found, the whole file is marked again. The Service will not deliver unmarked audio.
The mark carries a short identifier derived from the voice used, which allows us to corroborate that a given file came from this Service.

Layer 2 — Provenance metadata and content fingerprint.

The delivered WAV file carries standard RIFF/INFO metadata declaring: AI-generated audio, the voice identifier, the generation seed, the model used, the licence of the model weights, the UTC timestamp and a SHA-256 fingerprint of the audio.
The MP3 variant is transcoded on demand (44.1 kHz, 192 kbps). To avoid carrying over inconsistent tags, the WAV metadata is not copied into it; instead the MP3 carries an ID3 comment declaring "AI-generated audio | Watermark: active". The Layer 1 watermark is present in the MP3 as well — that is the layer that actually proves provenance.
Layer 2 is a convenience and an audit aid. It is not the compliance layer: metadata can be stripped by third-party tools, which is precisely why Layer 1 exists.

What Layer 2 does not contain: the metadata embedded in your file does not include your name, your e-mail address, your account identifier or the text you submitted. Provenance is reconstructed internally from the fingerprint and our own records, not from data written into the file you distribute.

3.2 Catalogue previews

Every voice preview you can listen to on the site is served from a separately watermarked and stamped copy (AudioSeal + provenance metadata), in WAV or lossless FLAC. The lossless variant is bit-for-bit identical in audio content, so the mark is intact in both.

The unmarked master of each catalogue clip is retained internally only, as the technical reference the engine conditions on; it is never served to the public. Marking that master would embed the watermark inside the reference and degrade every voice generated from it.

3.3 Images and video (when enabled)

When the image or video module is enabled, every file produced will carry machine-readable provenance metadata following the C2PA specification (Content Credentials) or an equivalent state-of-the-art standard, declaring that the file was AI-generated and including the generation timestamp, the model version and a cryptographic hash — verifiable with any compatible third-party tool. No image or video content is generated by the Service today, so no such file exists yet.

3.4 What the marking is and is not

It is a machine-readable declaration of artificial origin, designed to be detectable by us and by compatible tools.
It is not copy protection, DRM, or a restriction on your use of the file. Ownership and licence of the output are governed by the Terms of Service.
It is not infallible. Detection can degrade under deliberately aggressive manipulation (extreme time-stretching or pitch-shifting, heavy layering, severe re-encoding, cropping to very short fragments). The AI Act itself requires marking solutions to be effective, interoperable, robust and reliable "as far as this is technically feasible" (Art. 50(2)); we do not claim more than that.

3.5 Removing the mark is prohibited

Attempting to remove, strip, obscure, defeat or falsify either marking layer is a breach of the Terms of Service and of the Acceptable Use Policy, and may also be an infringement of the transparency obligations of the AI Act by whoever does it. The consequences of such a breach are those set out in those documents.

4. EU Artificial Intelligence Act

4.1 Our role and the obligation that applies to us

Regulation (EU) 2024/1689 (AI Act) imposes transparency duties on providers and deployers of certain AI systems. kreawave places a generative AI system on the market under its own name, so it acts as a provider (Art. 3(3)).

Art. 50(2) — this is our obligation. Providers of AI systems generating synthetic audio, image, video or text "shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated", with solutions that are "effective, interoperable, robust and reliable as far as this is technically feasible". Section 3 above is how we comply.
Art. 50(1) — not applicable. kreawave is not designed to interact conversationally with natural persons; it is a generation tool. We nevertheless state the artificial nature of the output on the pages where it is produced.
Art. 50(5). The information in this document is made available in a clear and distinguishable manner, and is linked from every page of the Service, before and at the time of generation.

4.2 Dates — and why no transitional period is being relied upon

The AI Act entered into force on 1 August 2024.
The Art. 50 transparency obligations apply from 2 August 2026 (Art. 113).
Regulation (EU) 2026/1744 (the Digital Omnibus on AI, adopted 8 July 2026, published in the Official Journal on 24 July 2026 and in force since 27 July 2026) postponed the high-risk regime — Annex III to 2 December 2027 and Annex I to 2 August 2028 — but did not postpone Article 50. Transparency obligations were left untouched and are live.
A narrow transitional window until 2 December 2026 exists for the machine-readable marking of systems already placed on the market before 2 August 2026. kreawave does not rely on it: the Service has marked every audio file it produces since before that date, and continues to do so.

4.3 Guidelines and Code of Practice

On 20 July 2026 the European Commission published its final Guidelines on Article 50, and on 8 July 2026 it issued the opinion finding the Code of Practice on Transparency of AI-generated Content adequate to cover Art. 50(2), (4) and (5) and to facilitate their implementation. The Code is voluntary. kreawave's technical implementation is built to meet Art. 50(2) directly.

kreawave has not signed the Code of Practice and makes no claim of adherence to it. The Code is voluntary and adherence to it is not a condition of compliance with Art. 50(2); our implementation is built to meet Art. 50(2) directly, and it is described, verifiably, in section 3. If we ever sign the Code, we will state it in this section with the date, in a new version of this document.

4.4 Supervision

The competent national authority for AI supervision in Spain is the Agencia Española de Supervisión de la Inteligencia Artificial (AESIA), created by Royal Decree 729/2023 of 22 August. The Spanish bill implementing the AI Act's governance and penalty regime (Proyecto de Ley Orgánica para el buen uso y la gobernanza de la Inteligencia Artificial, approved by the Council of Ministers on 26 May 2026) is still in its parliamentary stage; the directly applicable rules of the AI Act apply regardless of its outcome.

5. Your obligations when you publish generated content

Our marking obligation does not discharge yours. When you publish, broadcast or distribute content generated with kreawave, you act as deployer and you are responsible for:

Deepfake disclosure (AI Act Art. 50(4)). If the content resembles a real person, object, place or event and could falsely appear authentic, you must disclose that it is artificially generated or manipulated. The narrow exception for evidently artistic, creative, satirical or fictional works still requires disclosure in an appropriate manner that does not hamper the display of the work.
Adding a visible label ("AI-generated", "synthetic voice" or equivalent) where the destination platform, the applicable law or plain fairness to the audience requires it.
Platform rules. YouTube, TikTok, Instagram, Spotify and others impose their own AI-content disclosure requirements. Complying with them is your responsibility.
No impersonation and no misrepresentation. Never present generated content as a real recording of a real person, as a genuine statement, or as evidence of a real event. The full list of prohibited uses is in the Acceptable Use Policy.

6. Voice previews are selected recordings — read this before you buy

This is the most important honesty statement in this document, and it is deliberately explicit.

A catalogue preview is a recording we selected. It is not a live sample of what your text will sound like.

Specifically:

The previews were produced with a more exhaustive selection procedure than the one used to serve you. Each catalogue clip was generated five times and the best of those five takes was kept, judged by automatic quality and pronunciation scoring. Customer generations run a different procedure: each segment gets up to three attempts, and the first take that passes the quality gate is delivered immediately — the Service does not generate five takes of your text and pick the best.
The preview and your audio are produced by different routes. The preview is the reference recording itself. When you generate, the engine performs a new synthesis conditioned on that reference clip. Your audio will carry the timbre and character of the voice, but it is not a re-run of the preview and will not be an identical copy of it.
The preview text is not your text. Each clip speaks a short fixed sentence in the voice's own language. Different words, different length, different punctuation and a different language produce different prosody, different pacing and different pronunciation risk.
Consequently, a preview is a faithful reference for timbre, register, gender, apparent age and general style — and an indication, not a guarantee, of the exact result for any specific text.

We publish this so that you can judge before paying, and we give you three free ways to do so: the public demo without registration, the free plan, and the catalogue previews themselves. Use them. Those free tools are the objective standard of what the Service delivers, and they are the right moment to decide whether a voice suits your project.

7. Known characteristics and limitations of synthetic voice

The following are inherent, disclosed characteristics of the technology, not malfunctions of the Service:

Variation between generations. The same text, the same voice and different runs can differ in timbre nuance, prosody, pacing and emphasis. Synthesis is stochastic.
Seams in long texts. Long scripts are split into segments and reassembled. All segments are loudness-levelled to a common target, but audible transitions between segments can still occur.
Pronunciation errors. Proper nouns, brand names, foreign words, acronyms, initialisms, numbers in unusual formats and rare terms can be mispronounced. Automatic pronunciation checking runs across all 23 catalogue languages, but its accuracy varies by language and it is a statistical aid, not a guarantee.
Accent and dialect are properties of the chosen voice. They are not steerable by instructions in your input text. Choose the voice that already has the accent you need.
Expressiveness has a ceiling. The Service is engineered for narration. Extreme emotional performance, singing, shouting, character acting and overlapping dialogue are outside its design envelope.
Automatic quality control is best-effort. Every segment is checked for truncation, over-generation and pronunciation before delivery, and failing segments are re-attempted. When a segment exhausts its attempts, the best available candidate is delivered rather than nothing, and the result is flagged internally.
Results are not reproducible across model versions. The same seed on a different model version will not produce the same audio.

These characteristics are disclosed here, before purchase, precisely so that they are known. Re-generating a segment or a script is a chargeable feature; the conditions under which anything is or is not charged, credited or refunded are set exclusively by the Terms of Service and the Refund Policy. Nothing in this document creates a right to a free re-generation or to a refund.

When images and video are enabled, the equivalent disclosure will apply: anatomical artefacts (hands, limbs, in-image text), model bias inherited from training data, temporal instability between video frames, and no guarantee of the complete absence of artefacts.

8. Reproducibility record

For every generation the Service records a reproducibility contract: the seed, the model identifier and version, the execution engine, the language, the voice used, the generation parameters, the sample rate, the segment map, the UTC timestamp, and the watermark record (method applied, whether it is the robust in-signal method, and the SHA-256 fingerprint).

This record is available to you through the job status endpoint of the API for as long as the job record exists.
It survives the deletion of the audio file: when the audio retention window expires the file is deleted and the record is kept, without the file reference.
Retention windows applied in production: audio files 7 days (168 hours); the submitted text is replaced by its SHA-256 hash after 7 days; job records are pruned after 90 days, except records preserved as evidence of a content removal. The Privacy Policy governs the complete retention and data-protection picture.

The purpose of this record is to identify the exact conditions of any generation in the event of a dispute, an abuse report, a takedown notice or a compliance audit.

9. Verification of a file's origin

If you need to establish whether a specific audio file was generated by kreawave — for example as a rights holder, a platform, a law-enforcement body or a court — the Service can run a forensic check on the file you provide. The check reports whether our watermark is present, the detection confidence, whether the mark matches this Service, and the corroborating voice identifier.

Requests are handled through the Notice and Action / takedown procedure, which is the documented route and the one that produces a traceable record. This check is run by the Service; it is not a self-service tool, and it is not available for bulk or speculative screening.

10. Changes to this document

This document is versioned. Any substantive change raises the version number and the effective date shown at the top, and the superseded version is retained so that it remains possible to establish which text was in force on any given date. Material changes to transparency practices will be announced by the means set out in the Terms of Service.

11. Contact

Questions about AI labelling, watermarking or AI Act compliance: [email protected]

Origin verification, takedown notices and misuse reports: see the Notice and Action Procedure.

Full provider identification and all customer-service channels (postal, telephone and electronic) are set out in the Legal Notice.

+34 621 34 26 94 (Monday to Friday, 10:00-14:00 CET)