Skip to main content

EU AI Act — Article 50 transparency statement

EU AI Act Self-Classification

Regulation (EU) 2024/1689 applies to AI systems placed on the EU market. Kajo Voice contains AI components that process voice input on the user's device. This page gives the transparency disclosures that apply to it.

Article 50 transparency obligations have applied since 2 August 2026. The Digital Omnibus on AI (Regulation (EU) 2026/1744) gave only systems already on the market before that date extra time: until 2 December 2026, for the Article 50(2) marking duty.

Classification: not high-risk

Kajo Voice is not high-risk under the EU AI Act: it is not a prohibited practice (Article 5) and does not fall under an Annex III high-risk use case. It turns audio you already have into text, summaries, and cited answers on your own device and makes no consequential decisions about anyone. Article 50 asks that people know when they are dealing with an AI system, and that synthetic audio, images, video, and text be marked in a machine-readable way, unless the system only assists standard editing or does not substantially alter the input it was given or its meaning. Kajo names every model it runs, on this site and in the app's Settings → Models, presents transcripts and translations as editable text, backs chat answers with citations and a reminder to review them, and generates no synthetic audio, image, or video.

Transparency disclosures

Risk classification Not high-risk; Article 50 transparency applies

Not a prohibited practice (Article 5) and not an Annex III high-risk use case: Kajo makes no consequential decisions about anyone. Article 50's transparency duties are what apply, and the rows below set out how Kajo meets them.

AI models (all on-device) Whisper large-v3 turbo (MIT); Silero VAD v6.2 (MIT); NVIDIA Nemotron 3 Diarization (OpenMDW-1.1); Gemma 4 (Apache-2.0); Qwen3 Embedding 0.6B and BGE Reranker v2 m3 (Apache-2.0)

Whisper transcribes, the Nemotron model labels speakers, the embedding and reranking models power search, and one Gemma 4 model translates, summarizes, and answers cited chat — all on your Mac, Windows, or Linux machine, with no server-side inference on any path. Model weights download once, after installation; your audio and transcripts never leave the device for processing.

Output retention Local only — controlled entirely by the user

Transcripts, translations, and chats are stored only on the user's device. Kajo's servers never receive transcript content — there is no cloud-processing path. Delete library data in the Library; request deletion of your licence record via privacy@kajovoice.com.

Source attribution Users are told which models run

Every model, with its licence, is named on /what-runs-on-your-machine, and in the app Settings → Models lists each one by name. Chat answers cite the recording and moment each claim came from.

Human oversight Output is reviewable before it is used

Transcripts and translations are editable text; chat answers cite their sources, and the chat reminds the user, under the question box and in every exported answer, to review them before anything leaves the machine. The system takes no action on the user’s behalf and makes no consequential decisions automatically.

Logging & auditability Local only — user-controlled retention

Transcripts and library data live on your device under your control. Delete library items at any time from the Library. Export any single file on Free; Lifetime adds batch export. For account-side documentation, contact privacy@kajovoice.com.

What this means for you

  • You know it is AI: Kajo is an AI app from end to end, and every model it runs is named on What runs on your machine and in Settings → Models. All processing runs locally on your machine.
  • No cloud processing: There is nothing to opt out of — transcription, speaker labels, translation, summaries, and chat all run on-device by construction. Kajo sends no audio-derived text to any cloud model.
  • You can check and correct the output: Transcripts and translations are editable, and chat answers cite the recording and moment they came from. Nothing Kajo produces is acted on without you.
  • You can erase it: Delete a recording in the Library and its transcript, index, and the chats pinned only to it go with it. To delete your licence record, email privacy@kajovoice.com; we answer within one month.

Contact

Questions about Kajo's EU AI Act compliance: privacy@kajovoice.com . For GDPR inquiries, see our EU Privacy page and Subprocessor list .