The Algorithmic Cut

This is the Paradox of Skill: the faster the tools get, the less we practice the structural thinking that makes an editor. We drift from architects of time to curators of algorithmic probabilities.

Here are six ideas to push back.

1. Curator or architect?

Montage was the deliberate collision of Shot A and Shot B to make a new meaning. Prompting means choosing among suggestions. For the Algorithmic Native, who has never edited without AI, speed feels normal and friction feels like a bug.

Speed without comprehension is efficiency without authorship.

2. “Noise” is where the human truth lives

Plugins such as Gling or EditBuddy give clips a confidence score and penalize soft focus, shake, and off-center framing. That discards what Barthes called the punctum: the flawed, unexpected detail that moves us.

3. The silence problem

Transcript-based editing (Descript, text editing in Premiere or Resolve) treats film as language. It can’t see what Walter Murch called the breath and the blink: the pauses, glances, and held silences that carry a scene. Wiseman’s observational work could never be cut from a transcript.

4. The invisible Recommendation Editor

Auto Reframe, Smart Reframe, and Smart Conform crop widescreen to vertical by following faces. In doing so they erase scale, empty space, and the tension between people.

Platforms add another layer: concealed loops are built to inflate retention, while revealed loops can be used deliberately. Discovery algorithms decide which cuts get seen, so they co-author your edit.

5. “And then” versus “therefore”

Retention-optimized pacing favors “and then”: A happens, then B. The essay-film tradition, from Eisenstein onward, uses “therefore”: two images collide to make an argument. Algorithmic defaults smooth that collision away.

6. The Answering Cut

Show the person in the footage two versions: one shaped by automation, one cut by a human. Their response isn’t a vote or a veto. It shows where meaning entered the image, and sometimes a blind spot that began at the camera. Then choose a consequence:

Keep · Re-edit · Reframe · Return to footage · Shoot a new image

Becoming the director of the algorithm

You don’t have to reject AI. You have to build structural literacy: knowing how software handles time and what it assumes by default.

• Start analog. Use paper edits, index cards, and a written thesis before opening your NLE.
• Look for the punctum. Keep the hesitations and imperfections machines filter out.
• Trace every element across Souriau’s levels: profilmic (what existed before the lens), filmic (what montage constructs), and synthesized (what AI generated).

Next time you cut, ask: am I the architect of meaning, or just approving the algorithm’s guess?

The project began as a book on AI-controlled editing and grew into educational tools for film schools. At the first team meeting, the research team gathered in Ankara and online to introduce themselves: research assistants, co-researchers in Turkey and France, and colleagues from the Media Archaeology Lab.

The meeting settled the shape of the work: five exercises that test automated editing against human editors. The fifth, proposed by filmmaker Nathan George, has people film themselves and then compare how AI and human editors represent them. The team agreed to start the first exercises with existing CILECT material, to film the workshop process on phones and cameras for documentation and possible further use, and to plan the first workshop together.

Because of the nature of the experiment, and because they knew observers and researchers were watching them, all four participants started by watching the footage slowly, carefully, and in detail. After the first hour, however, they gave up examining every detail and switched to skimming.

One reason they first watched each shot in full, before getting bored and switching to 2x speed or skipping, may be their assumption that AI would not miss a second, so they shouldn’t either.

Editing approaches

Half of the participants edited their footage fairly linearly, with few additional layers, effects, or transitions. The other half made their final cut more elaborate: they added extra layers and intersecting cuts, combined the sounds of different clips, or built a narrative into their edit. The hardest part of the session was that each student had their own editing ideas and workflows, so their processes, and therefore their timing, differed.

Overattributing agency to AI

A tentative finding of this phase is the overattribution of agency to AI editing interfaces. Some students implied that AI could detect jokes and understand Chinese, German, and English at the same time. They also thought AI could understand and interpret what was happening inside the frame and logically decide whether to keep or discard the footage. So they mainly valued the content and the exposure and framing of the footage, believing AI would favor these. They also thought the algorithm would choose human faces and speech over white noise or wind, or over a frame of greenery with no people in it.

They attributed traits to the AI-integrated interface as if it were a peer, overestimating its analytical skills. This is exactly what the exercise is designed to expose.

“I don’t know how smart AI is, haha.” — Participant 4

Assumed tastes

At some points, students suggested that they would favor what AI would favor, but that they would not choose such footage or edit it that way, since that is what AI would do in their opinion. We see some intersection between the students’ tastes and the assumed tastes of AI.

They also constantly asked questions about fixing, reframing, and stabilizing the footage; therefore, they also associate stable and clear footage with AI.

As a tentative conclusion for the pilot phase, it was interesting to see that AI was assumed to have the qualities of a fifth participant: another Communication and Design student who could read and speak Chinese, German, and English at once, with expert editing skills.

Every participant had the same task: predict which clip set an algorithm would keep and which it would flag as waste, then build a two-minute cut from the flagged pile.

Two AI agents, two approaches

The two AI agents read the footage in very different ways. Claude never looked at the images; it did its editing through the metadata (gyroscope, accelerometer, ISO, brightness) provided by the GoPro Hero7. It flagged 45 of the 60 clips for shake, walking bounce, wind on the mic, false starts, and faces cut by the frame edge. Codex, on the other hand, took multiple screenshots of the timeline and the footage, and through that process it was able to judge what was in the picture. It flagged 43 clips, mostly for foreground heads, backlit faces, and camera rigs blocking the performers.

Claude, reading and editing through sensor data, could not see the crew and kept several of those clips — even admitting it in its own logs: “the method cannot see composition.”

The crew as dividing line

The documentary crew was the clearest dividing line: Codex and the participants kept noticing them — “too much cam,” “another cam blocking,” “camera person is visible.” Claude, reading through sensor data, could not see them.

How participants flagged clips

The participants indulged in something that neither agent did: they often flagged for content rather than the signal — “nothing is going on,” “no performance,” “just talking,” “backs turned.” When they kept a clip it was usually for “talking” or “faces,” which suggests they assumed the algorithm values people and speech.

The pull toward repair

The decision logs reveal the pull toward repair that the exercise is built to expose: “can be fixed with color grading,” “would have wanted to isolate the sound,” “I wish I could brighten it.” One participant even kept a clip anyway, noting that they thought “AI would deem it unnecessary but it goes well as a b-roll.”

That gap between a technical flag and an editorial use is where the second activity will begin.

In all of the modes without exception, the edit follows a conventional structure: opening with footage of ascending the wall and closing with the descent. While this structure is the same, the specific clip selections differ depending on the chosen mode. The interface does not explain how the clipping modes work, so we don’t know what features differ.

The tool seems to prioritize shot order and sequence structure over precise cut points or pacing within individual clips.

Montage mode

The sequence result seems like a mixture of traditional narrative editing and the visual language of social media “edits.” It emphasises the backstage of actions and scenic views over footage that could contribute to the storyline. It didn’t hesitate to use footage that is problematic in technical terms, especially in terms of lighting. It uses lots of panning and dolly shots, which resembles social media edits with non-stop dynamic but stable camera movements.

Story-driven mode

Instead of showing a nice collection of images, the sequence seems to represent more of a documentary narrative, almost in a verite mode, with little stable shots and lots of handheld ones. This mode is the most unique in terms of footage selection.

Auto-style mode

This mode adopts a travel-vlog aesthetic with almost non-narrative visuality and a sequence that resembles a travel show. Instead of a narrative, there is a real-time experience shown. It largely neglects more cinematic footage in favor of casual, real-time moments.

Chronological mode

It attempts to infer a linear narrative from the footage and arrange it chronologically. While it successfully maps logical continuity (such as path progression and walking routes), visual coherency is disregarded. The transition balance between different lighting and camera qualities can feel inconsistent and cluttered.

The AI Clipper demonstrates an ability to recognize spatial progression and establish a conventional narrative without any prompt. However, its effectiveness is constrained by a lack of algorithmic transparency and an indifference to pacing and technical flaws. The tool functions better as a rough-cut structural assistant than as an autonomous editor.