AI video generation is the most overhyped and most underrated category in AI simultaneously. Overhyped because most demos you see online are cherry-picked best cases. Underrated because the real use cases — the ones that actually save time and money — barely get coverage.
I’ve spent significant time testing everything from Sora to Runway to Pika to Kling. Here’s what I actually think.
The Honest State of AI Video in 2026
What’s genuinely good:
- Short-form B-roll generation for YouTube and long-form content
- Style transfer and visual consistency across clips
- Background removal and replacement in motion
- Extending clips by a few seconds to hit a target length
- Creating product visualization from images (for e-commerce)
What’s still rough:
- Generating realistic human motion that holds up under scrutiny
- Maintaining consistent character appearance across scenes (getting better, not there yet)
- Lip sync that works reliably at 1080p+ quality
- Anything requiring text in the video (AI still can’t reliably generate legible text)
The Tools, Honestly
Runway Gen-3 Alpha
This is what I use most. The image-to-video and video-to-video tools are the most practical for creators who already have footage. The motion brush — where you paint which parts of an image should move — is genuinely creative and useful.
Best for: Extending existing footage, creating stylized B-roll, visual effects on real video. Not great for: Long-form coherent narrative, realistic human motion.
Sora (OpenAI)
The quality ceiling is the highest I’ve seen. When it works, it’s stunning. The problem is consistency and control — you don’t always get what you describe, and iterations can feel unpredictable.
Best for: Concept visualization, creative short-form, one-off impressive pieces. Not great for: Repeatable content workflows, anything requiring precise control.
Pika
Pika has found its lane in the consumer and social media creator space. It’s faster and cheaper than Runway, with good quality for the price point. The “pikaffects” (style modifications on existing video) are fun and useful for social content.
Best for: Quick social content, experimenting with styles, accessible entry point. Not great for: Professional production quality, nuanced control.
Kling (Kuaishou)
The Chinese entrant in this space has been surprising. The human motion is among the most realistic I’ve tested, and the longer generation times (up to 3 minutes) are technically impressive.
Best for: Realistic character motion, longer generated sequences. Not great for: Western-style content, it’s optimized for different aesthetics.
What I Actually Use It For
My real use is narrow: when I don’t have footage for a concept, I generate B-roll. When a clip is 2 seconds short, I extend it. When I want to show something that’s hard to film — like a visual metaphor for an abstract AI concept — I generate it.
I haven’t replaced any part of my production that involves me. My face, my voice, my presence — that’s still real. The AI fills in the gaps.
The Question Worth Asking
For any tool in this category, ask yourself: does this help me make content I couldn’t make before, or does it replace content I was already making well?
The first use case is where AI video shines. The second is where you’ll end up with something generic.
Use it to expand what’s possible, not to cut corners on what you already do.
— Deco