How Flinzo Uses AI
This page explains in plain English how Flinzo works today. If anything here is unclear, write to support@flinzoai.com and we will answer.
The short version: AI never changes the pixels of your footage. A model reads a transcript, scene boundaries and a handful of still frames, and it writes a plan; an editing assistant turns your messages into edits you can undo. Ordinary, deterministic video software executes that plan against your original footage. That is why the same plan produces the same video twice, why you can see what will happen before it happens, and why running Flinzo does not cost what a model generating video would cost. The one thing a model may generate is a sound effect — section 1 says when.
1. Which features use AI, and which provider powers each
| Feature | Provider | What the provider receives | What it returns |
|---|---|---|---|
| Transcription and word-level timing | ElevenLabs Speech-to-Text | The extracted audio track only — not the video | Text with a timing for each word |
| Understanding what is in the footage | Google Gemini API | The transcript, the scene boundaries, and 20 to 40 sampled still frames | A description of what each part of the footage contains |
| Writing the edit plan | Google Gemini API | The same transcript, boundaries and sampled frames | A plan in JSON: which segments to keep, in what order, where captions go, how to reframe |
| Motion graphics, hooks and inserts for a new clip | Google Gemini API | The transcript of the clip and a short summary of it. No footage | Suggestions for the hook, highlighted words, punch-ins, B-roll, music and sound effects. Our own rules decide what is applied, and one Undo removes all of it |
| The editing assistant — changes from your messages | Anthropic Claude | Your message, the current edit and the transcript. When you ask about a specific moment, a few still frames of that moment go to Google Gemini | Edit operations, each of which you can undo |
| Sound effects, only when no library sound fits | ElevenLabs Sound Effects | A short text description of the sound. Never your footage or audio | A short audio clip, marked as generated in the export’s rights record |
The chat panel in the editor is an AI assistant, not a person. Every change it makes is an edit you can undo, it spends credits only at the price it shows you first, and it does not learn anything about you between sessions beyond what your account settings record.
Not AI: cutting, reframing, audio mixing, silence removal, encoding, caption and motion-graphics rendering, and compositing. Those are performed by conventional media software following the plan. No model generates or alters a frame of your footage, a face or a voice. Generated sound effects are the only exception, and music is generated only if you turn that on; both are labelled in the export’s rights record.
2. What a model sees, and what it never sees
We never send your whole video file to an AI model. For a three-minute clip, what leaves our servers for analysis is the transcript, a list of scene boundary timestamps, and between 20 and 40 still frames sampled across those scenes. That is enough for a model to understand structure and content, and it is a fraction of the data — and the cost — of sending the video.
Sampled frames are derived from your footage, stored privately with the project, used for analysis, and deleted with the project. Face detection for framing is not an AI service call: it runs on our own workers, keeps only crop geometry and salience scores, and creates no biometric identifier. The privacy notice has the detail.
Your footage is never used to answer another user's request, and no model output is shared between accounts.
3. Training on your content
We do not use your content to train AI models. We have no model of our own, and we never send your content to a provider in order to improve one.
Whether each provider trains on what we send it is a separate question, and the honest answer differs per provider. The subprocessors page states it per provider and says "pending confirmation" wherever we do not yet hold the evidence — either the provider's terms in writing, or our own executed opt-out with a stored confirmation. We will not claim a provider does not train until that evidence exists.
Two details that "nothing is stored" would hide. Providers retain requests for a limited period to detect abuse, even where they do not train on them; one of ours documents roughly 55 days for that purpose. And content that a provider's trust-and-safety system flags can be retained considerably longer — one provider documents up to two years. We would rather tell you that than write a sentence that reads better and is false.
4. Limitations, and what AI output cannot promise
- Models are not deterministic. The same footage can produce a different plan on a different day. The renderers are deterministic: the same plan always produces the same video.
- A model can pick the wrong moment as your strongest, misjudge pacing, or mislabel who is speaking.
- Transcription makes mistakes, especially with names, jargon, accents, overlapping speech and noisy audio. Captions are machine-generated and need your review before you publish.
- Reframing decides where to crop from motion, audio direction and where faces are. It can crop something you wanted in frame.
- Quality checks run on the rendered output and reject a render that fails its profile rules, but they do not judge whether the edit is a good edit. That is your call.
- Nothing about Flinzo predicts how a video will perform. We make no claim about views, followers, revenue or reach, and we will not.
Every plan is shown to you before anything is rendered, as readable cards rather than as data, and nothing renders until you approve it. You can mark a moment, describe the change in your own words, and get a targeted revision. You are the editor; Flinzo is the assistant that does the tedious parts.
Provider models change, and a provider can retire the version we use. When we change which model powers a feature, this page changes with it.