Draft pending counsel review — last updated 2026-09-29. This is a working draft published for transparency. It is not final legal advice and will change after review.
AI transparency
SlideDeck uses AI to draft, check, speak and listen. Here is exactly how, and where the limits are.
Models we use
| Role | Model route | Via | Used for |
|---|---|---|---|
| Typed classification | typesafe/jev-1.13 | OpenRouter | Brief classification, routing and schema-constrained decisions. Never the source of factual truth. |
| Main language model | openai/gpt-6-luna | OpenRouter | Argument construction, drafting, script writing and substantial verification. |
| Alternative language model | xiaomi/mimo-v2.6-pro | OpenRouter | Quality-tested alternative or independent verification route. |
| Narration and coach voice | google/gemini-3.8-flash-lite-tts | OpenRouter | Generated speech for narration and rehearsal questions. |
| Transcription | assemblyai/universal-3-5-pro | OpenRouter | Captions and your transcript during rehearsal (audio sent in short phrases, not stored), and transcribing uploaded recordings. |
| Transcription benchmark | openai/whisper-1 | OpenRouter | Quality comparison on consented sample audio only; not used for your recordings. |
No other models or fallbacks are used. Search, where you enable it, retrieves sources; it doesn’t generate content. Each assessment records the model and presentation versions used, so a model upgrade can’t silently rewrite past feedback.
Generated content and voice
- Slides, scripts, questions and feedback are AI-generated drafts for you to review and edit.
- The coach voice and narration are synthetic, generated by a text-to-speech model. They are not recordings of a person.
- We don’t clone voices, including yours.
- During rehearsal the interface shows honestly whether the coach is listening, processing or speaking.
- Generated images, where used, are never presented as evidence of a real event.
Practice feedback limits
- Feedback is private practice feedback, not an official grade or a prediction of one. Mapping to your rubric is labelled as practice.
- We do not infer emotions, personality or nervousness, and we do not grade accent, stutter, disability or communication style.
- Audio-only practice never claims to assess eye contact or body language.
- Where speech recognition is uncertain, that is reported separately from your presentation quality; unreliable metrics are marked unavailable.
- Our transcription model removes filler words such as “um”, so we don’t report a filler count rather than guess one.
- If you state a true fact that isn’t in your sources, it is marked for verification rather than called wrong.
- The coach never invents a “correct answer” to challenge you, and rewards calibrated uncertainty.
Source policy
Every substantive factual claim is linked to inspected evidence or visibly labelled with one of these states:
- Supported: Inspected evidence supports the bounded claim.
- Partially supported: Evidence supports only part, or a weaker wording.
- Conflicting: Credible sources disagree in a material way.
- Unsupported: No inspected evidence supports it.
- Stale / recheck: Evidence may no longer be current for this use.
- User assertion: Supplied by you, not independently verified.
- Calculated: Derived from identified inputs and a reproducible formula.
- Illustrative / opinion: Not offered as an empirical fact.
We don’t attach citations to opinions to make them look objective, and we don’t claim to have read a whole article from its abstract. Citation is not a licence to redistribute full documents: shared and public exports follow source visibility and excerpt permissions.
Your data and training
We don’t train models on your presentations, sources or recordings, and we don’t opt you in to provider training. Any future research or model improvement programme would be a separate, explicit opt-in. See subprocessors for where data goes.