Can AI edit videos like a human editor? Not yet. I spent hours teaching GPT-6 Astra, OpenAI's computer-use agent, to cut in Premiere Pro, and it got 80% or 90% of the way there. Captions, B-roll, cuts and basic pacing are solved. Senior-editor judgment isn't, and as of September 10, 2026, training a human is cheaper and faster.
This post expands an 80-second video I recorded on September 10, 2026 and published on September 16. It's embedded below.
Key takeaways
- AI gets you to 80% or 90%. After hours of teaching GPT-6 Astra to edit in Premiere Pro, that's where the output landed, and the missing stretch is the part that matters for a serious channel.
- The mechanical layer is solved. Captions, B-roll placement, cuts and a basic sense of pacing are things AI already handles well enough to ship.
- Nobody big publishes 100% AI-edited videos. Every major personal brand and media company I know of runs human editors who use AI tools, not AI running on its own.
- Training a human is still the cheaper, faster path. As of September 10, 2026, getting a person to edit like a good editor costs less time and money than teaching a model to do it. That will flip eventually.
Watch it on YouTube · Instagram · LinkedIn · X
A few days after OpenAI put GPT-6 Astra out, I gave it a job I'd normally hand to a new editor: open Premiere Pro and cut a video the way we cut videos. I expected it to embarrass my editors. It didn't. It got close, and "close" turned out to be the whole story.
What happened when I taught GPT-6 Astra to edit in Premiere Pro?
It got to roughly 80% or 90% of what I'd accept from a person, and it took hours of teaching to get there.
Astra is OpenAI's model that can use a computer the way a person does: it looks at the screen, moves the mouse, clicks through menus and types. OpenAI's own system card is the official description; it was published September 3, 2026. So instead of exporting a rough cut from some AI-editor app, I put it inside Premiere and taught it our process: where captions go, how we cut, where B-roll drops in, how tight the pacing should be.
The result was good. Actually good, not "good for a robot." I'd just walked in expecting better, and the final stretch is where it fell apart. I'm a perfectionist, so my 100% is hard to reach. Be generous with what you call 100% and the gap looks smaller. It's still there.
What can AI handle in video editing today, and what still needs a senior editor?
Captions, B-roll, cuts and a bit of pacing are solved. Everything that makes a senior editor a senior editor is not.
| Task | AI today (September 2026) | Still needs a senior editor |
|---|---|---|
| Captions | Yes. Timed, styled, accurate enough to ship. | Only the style decision. |
| B-roll | Yes. Finds relevant footage and drops it in. | Knowing when a cutaway hurts more than it helps. |
| Cuts | Yes. Removes dead air and tightens the take. | Which take carries the point, and where the clip starts and ends so the payoff lands. |
| Pacing | Basic. It knows a video should move. | Rhythm: when a pause is tension and when it's dead air, when a line needs room. |
| Taste and judgment | No. | All of it. Matching the creator's style, knowing what the audience will sit through, calling the edit done. |
If 80% good is the bar, AI alone gets you there today. That's a real shift and I'm super bullish on where it goes. The gap between "mostly there" and done is the part a senior editor gets paid for, and right now nothing closes it except a senior editor.
Do major creators and media companies publish AI-edited videos?
No. Not one major personal brand or media company I know of is putting out videos that are 100% AI-edited. They all have editors, and those editors are empowered by AI.
That's not caution on their part. Their audience is the business, and the audience feels the difference between a mostly-there edit and a finished one even when it can't name what's off. At VPPA that's exactly how we run it: human editors cutting highlight clips and shorts with AI tools doing the mechanical layer. The tools keep getting faster. The person making the calls stays.
Is it cheaper to train a human editor or to teach AI to edit?
As of September 10, 2026, it's cheaper and faster to train a human to edit like a good human than to teach AI to do the same thing. I'm writing the date down on purpose so we can all see when I'm wrong.
Here's why it plays out that way right now. A human editor learns your taste from a handful of examples and then generalizes: show them how you'd cut one clip and they carry it into every clip after that. With Astra, I spent hours teaching and still landed at 80% or 90%, and every new edge case is another teaching session. The human compounds. The model, at least today, plateaus.
Obviously this changes. The technology gets more powerful on a daily basis, and I think it'll be on par with a junior human soon. But "soon" isn't "now," and the hiring decision you make this quarter should be based on now.
Will AI replace video editors?
Eventually it will take over a big chunk of what editors do today; it's a matter of time. It won't replace the judgment layer any time soon, and that's the layer that separates a video people finish from one they scroll past.
In practice, the editors who get squeezed are the ones whose whole value is captions, B-roll and cuts, because that's exactly what the model already does. The ones who stay valuable are the ones who make the calls. If you hire editors, hire for judgment and hand them the AI. If you are an editor, that's your roadmap.
Who this is for (and who it isn't for)
This matters if you're deciding whether to hire an editor, train one, or go AI-only for a channel where the edit is part of the product: a personal brand, a podcast, a media company, an agency.
It matters a lot less if 80% good is your bar. Internal updates, quick screen recordings, a test channel, a Loom you'd otherwise never edit: use AI, ship it, move on. Being a perfectionist about videos nobody will re-watch is a waste of time, and I say that as one.
Common mistakes when you test AI for video editing
- Grading it on the parts it does well. Captions and cuts look great, and that's where you get fooled. Grade the judgment calls, not the captions.
- Expecting it to replace an editor instead of speeding one up. Every big brand is on the second model. Run the same play.
- Teaching the model what a person would learn faster. That's where my hours went. If your editor could pick it up in an afternoon, you're solving the wrong problem.
- Using a lenient 100%. If your 100% is "it uploaded," AI wins today. Define 100% as what a senior editor would ship, then compare.
Frequently asked questions
Can AI edit videos in Premiere Pro?
Yes. GPT-6 Astra can operate Premiere Pro directly, the way a person does, and after hours of teaching it produced edits that were 80% or 90% of what I'd accept from a human editor. The remaining gap is judgment, not mechanics.
Will AI replace video editors?
Not as of September 2026. It already handles captions, B-roll, cuts and basic pacing, but no major personal brand or media company publishes 100% AI-edited videos. They run human editors who use AI. Expect it to reach a junior editor's level before it reaches a senior one.
What can AI do in video editing right now?
Captions, B-roll placement, cuts and a basic sense of pacing. It can't yet make the taste and judgment calls a senior editor makes, like which take carries the point or when a pause is tension instead of dead air.
Is it cheaper to hire a video editor or to use AI?
For 80%-good videos, AI is cheaper. For videos edited to a senior editor's standard, as of September 10, 2026 it's cheaper and faster to train a human than to teach AI the same thing. That will change; the date is there so you can check.