AI transparency notice
Applies to: Throughline 3.0.0 · Last updated: 2026-08-16
Written to satisfy Article 50 of the EU AI Act (Regulation 2024/1689), whose transparency obligations apply from 2 August 2026. It is written to be read by a user, not only by a regulator.
1 · You are interacting with an AI system
Throughline uses machine-learning models for four things:
| feature | what the model does | runs |
|---|---|---|
| Transcription | speech → text | on your Mac |
| Speaker diarization | marks who is speaking when | on your Mac |
| Story recommendations | suggests which takes support a narrative | on your Mac, or a cloud provider you choose |
| Translation | translates transcript text | on your Mac |
Output from all four is machine-generated and can be wrong. Transcripts contain errors. Diarization mislabels speakers. Story recommendations are suggestions, not judgements. Treat every one of them as a draft to check, not a result to rely on.
2 · Our role under the AI Act
Throughline is a deployer of general-purpose AI models, not a provider of them. We do not train the models we ship; we redistribute them unmodified, under their own licences, with attribution (§4). We do not use them to infer emotion, to categorise people biometrically, or for any purpose listed in Annex III of the Act.
Speaker diarization separates voices — it clusters audio segments by acoustic similarity. It does not identify anyone, match anyone against a database, or build a voiceprint that persists beyond the session. It is not biometric identification within the meaning of Article 3(35).
3 · Sync placement is not AI
The core function of this product — deciding where a clip belongs on a timeline — is signal processing, not machine learning: envelope correlation and GCC-PHAT cross-correlation over audio waveforms. It is deterministic and it explains itself. Every placement carries a plain-language account of the actual evidence behind it, and the application refuses to present a clip as placed when nothing measured it.
This matters for reading the accuracy claims: they describe a measurement, not a prediction.
4 · The models, and their licences
| model | purpose | licence |
|---|---|---|
| Parakeet TDT 0.6B v3 | transcription | CC-BY-4.0, © NVIDIA (attribution required — shown in About and embedded in exported transcripts) |
| Whisper (base / large-v3) | transcription | MIT, © OpenAI |
| pyannote community-1 | diarization | CC-BY-4.0 (attribution required) |
| wav2vec2 aligner | word alignment | Apache-2.0 / MIT |
| Phi-3.5-mini-instruct | story recommendations | MIT, © Microsoft |
| Qwen2.5-1.5B / 0.5B Instruct | story recommendations | Apache-2.0, © Alibaba |
| Apple on-device model | story recommendations | Apple system framework |
Models under a non-commercial licence are never offered. The catalogue carries a
redistributable flag and the picker cannot show a model that is not marked
redistributable — because a user choosing one would put themselves in breach
with no way of knowing.
5 · Known limits, stated plainly
- Transcription accuracy falls with overlapping speech, strong accents, heavy background noise and non-English audio.
- Diarization struggles when speakers overlap or sound alike.
- Story recommendations reflect patterns in the model’s training data. They can be bland, confidently wrong, or blind to what makes a particular interview work. They are a starting point for an editor, never a substitute.
- The smallest model (0.5B) is the weakest at producing valid structured output. It is offered for older Macs, and the validator’s repair retry matters most there.
6 · What is never sent anywhere
Your video and audio are never transmitted to any AI provider. On-device models receive audio locally. Cloud story providers receive transcript text only, after you have chosen that provider and given consent, using your own API key. See PRIVACY.md §3.1.
7 · Human oversight
Every AI output in Throughline is advisory and editable. Nothing is applied to your timeline without you. There is no automatic decision with legal or similarly significant effect — the product edits video.