"AI video editing" has become a label stuck on three genuinely different technologies: models that generate footage from text, traditional editors with AI features bolted on, and editors where you describe changes to real footage in plain English. They solve different problems, fail in different ways, and get conflated in almost every product page you will read. Here is the sober map: what each one actually is, what AI reliably does well in an edit today, and the questions that cut through the marketing.
Generative video models (the Sora and Veo family, and their successors) synthesize new footage from a text description. Nothing existed before the prompt; the model dreams the pixels. This is creation, not editing, and it is the technology behind both the impressive demo reels and the "AI slop" flooding feeds. AI-assisted editing is a traditional timeline editor with machine-learning features embedded: auto-transcription, text-based cutting, background removal, reframing suggestions, filler-word removal. You still operate the editor; AI accelerates specific tasks inside it. Prompt-driven editing inverts the interface. You state the outcome in plain English ("trim the first 10 seconds, crop to 9:16, add captions") and the system plans and executes real operations on your real footage. Supercut is built this way: an AI turns your sentence into a structured edit plan, and a deterministic engine maps that plan to exact FFmpeg commands that run in your browser. The AI never touches pixels; it plans, and conventional video processing executes.
Strip away the hype and a solid core of capabilities has become dependable.
AI has no taste. It cannot know that the third take is the one where your delivery landed, that a pause should breathe instead of being trimmed, or that your brand would never use that transition. Editing judgment, what makes a video good rather than merely finished, remains entirely yours, and any product implying otherwise is overselling. The failure modes differ by category, and knowing them is most of the defense. Generative video hallucinates: physics, hands, text, and continuity can silently go wrong, and using it means reviewing every frame. AI-assisted features misfire locally: a transcript mangles your product name, auto-reframe follows the wrong face. Recoverable, if you check. Prompt-driven editing’s risk is misinterpretation: the plan does not match what you meant. The honest mitigations are architectural: deterministic execution (the same plan always produces the same result, so nothing drifts mid-render), showing you the plan before it runs, and previews before export. That determinism is why Supercut separates the AI that plans from the engine that executes.
AI features need compute, and where that compute runs decides where your footage goes. Most cloud AI editing sends your video to servers for processing, which means copies of your raw footage exist outside your control, governed by retention policies and terms you should actually read, including whether uploads can be used to improve or train models. It does not have to work that way. Modern browsers run real models locally: Whisper transcribes on your device, face tracking runs on your device, and WebAssembly executes FFmpeg on your device. Supercut’s split is explicit: the language model that plans your edit sees the text prompt and basic metadata (duration, dimensions, filename), never the pixels, and every pixel operation happens in your browser. When evaluating any AI editor, ask the one-sentence version: does the AI see my footage, or only my instructions?
Five questions separate genuine utility from a demo reel. Does it save minutes on jobs you actually do every week, or only on contrived showcase tasks? Can you see and correct what the AI decided, or is it a black box between you and the export? Where does processing run, and what leaves your device? Is execution deterministic, or can the same request produce different results on different days? And what does failure cost: a wasted preview, or a published mistake? Tools built for real workflows answer these plainly. Marketing built on vibes changes the subject.
Add auto-captions to any video in your browser. On-device transcription with viral caption styles, no upload, your footage and audio stay private.
Crop and reframe any video to 9:16 for TikTok, Reels and Shorts. Subject-tracking keeps the action in frame, all in your browser, no upload.
Clean up muddy audio and reduce background noise in a video, in your browser. No upload, your footage and audio stay on your device.
Blur faces in a video to protect privacy, right in your browser. Face detection runs on your device, your footage is never uploaded.
Most online editors upload your footage to their servers. Here is how to edit video privately in your browser, so your files never leave your device.
Add subtitles or captions to any video by typing a plain-English prompt. On-device transcription times each word to your speech, and nothing uploads.
One of three things: generating new footage from text (Sora-style models), AI features inside a traditional timeline editor (auto-captions, reframing), or prompt-driven editing where you describe changes to real footage in plain English and the system executes them. They are different technologies with different strengths.
It can execute well-defined operations automatically: transcribe and caption, track a face and reframe, cut at scene boundaries, clean audio. It cannot decide what makes your video good. Treat AI as a very fast operator and keep the judgment calls yourself.
Often, but not necessarily. Many tools upload footage to cloud GPUs, while others run models on your device. Supercut sends only your text prompt and file metadata to the AI that plans the edit; transcription, face tracking, and all pixel processing run locally in your browser.
It is replacing the mechanical layer: transcription, rote cutting, format conversion, first-pass cleanup. The scarce skill was never dragging clips on a timeline, it was knowing what to keep and why. People with taste get faster; taste itself is not automated.
Generative video creates footage that never existed, from text. AI editing operates on footage you shot: trimming, reframing, captioning, cleaning. If your starting point is real video you need improved rather than imagined, you want an editor, not a generator.