How we use artificial intelligence
Vilko is an AI product and we think you should know exactly what it does, with what, and what it can't do. No law requires this yet; we wrote it because it's the right thing.
01Everything you see is generated
The videos Vilko produces are not recordings of reality. There is no camera, no actors, no location. It is all generated by an AI model from your description.
The people who appear do not exist. They are not real actors and are not based on specific people.
02Which models we use
Seedance 2.5 (ByteDance, via Atlas Cloud) to generate the video; Seedance 2.0 as a rescue when a take resists.
Claude (Anthropic): Opus 5.5 talks with you in the chat, writes the script, makes the sheets, applies your changes and reviews every take as a director — watching it at 12 frames per second at 768 px, with a flicker-and-jump detector over all 24 real frames. It also reads the photos, web pages and PDFs you send.
Nano Banana Pro (Google) for the character and object sheets, in 4K, and Seedream 5.0 Pro (ByteDance) for the look proposals, both via Atlas Cloud; GPT Image 2 (OpenAI) as a fallback if either fails. Gemini (Google) to watch and listen to the videos and audio you upload as references.
ElevenLabs for music, sound effects, voice and subtitles (Scribe).
Topaz Labs to upscale the video to high resolution, up to 4K.
OpenAI solely to transcribe voice dictation, if you'd rather talk than type.
We may switch models when a better one ships. Our commitment is that the switch is always to raise quality, never to cut cost.
03Your content trains nothing
We do not use your descriptions, scripts or videos to train models of our own, and we contract providers on tiers that don't either.
Your projects are yours and we won't show them as examples without your written permission.
04What the machine reviews and what it doesn't
Every take is reviewed by Claude as a director, at 12 frames per second: continuity within the take and between takes, color and light against the look you chose, physics and logic, text, anatomy, the pacing you asked for, and whether the take tells what you said in the chat. A separate detector hunts for flicker and jumps across all 24 real frames. Whatever fails gets trimmed, corrected, retouched with VFX or regenerated — up to eight attempts per take that you never see.
That review does not judge whether the ad is good, whether the message is right, or whether it complies with advertising law in your country. That's on you before publishing.
05What AI still gets wrong
Text inside the image frequently comes out wrong. That's why important text — your address, your phone, your price — is placed as real text over the video, not generated.
Fine details of a real product may come out different. If you sell something where the exact detail matters, review the result carefully before publishing.
One- or two-frame defects can escape review. They're near-imperceptible, but they exist.
06Your responsibility when publishing
You decide what you publish and it goes out under your name. Before uploading, check that it doesn't misstate your product and that it complies with advertising rules in your country and on the platform you're posting to.
Several countries now require disclosing that an ad was AI-generated. Inform yourself: that obligation is yours as the advertiser.
07Watermark
Videos carry no Vilko watermark, on any plan: the file comes out clean.
We do not embed invisible watermarks or tracking metadata in your deliverables.
Notice
This document is an operating draft, not legal advice. If anything here affects you materially, consult your attorney — and we'll do the same before we start charging.