The illustration: MIT Media Lab researchers reported lower neural engagement during AI-assisted essay writing in a small EEG-based preprint study, but the findings are preliminary. A datapoint I cannot forget is this: the average buyer of a multi-million dollar apartment is now around twenty. That is not a gossip detail. It is a signal that attention has become convertible currency, and the people who understand distribution can turn it into real assets. The new elite is not “someone who posts”. It is someone who can operationalize it without losing control of what that force does to others. Which is why AI-assisted video editing matters more than it seems. Short-form content is built from micro-decisions: what to show, what to hide, what to cut before it drifts. A single reel can cost hours because selection is a cognitive sport. Automate it, and you get scale. Automate it carelessly, and you get something else too – a culture where taste is outsourced and responsibility loses its owner. AI-assisted video editing AI-assisted video editing The Flood and the Filter The Flood and the Filter Marketing, at its simplest, is the fight for human attention and the attempt to shape a choice at the shelf, physical or digital. The constraint is plain and unforgiving: there are only twenty-four hours in a day. Now imagine content not produced by tired people with deadlines but produced by factories. Once generation becomes cheap, the pressure does not merely rise; it changes form. We will publish more than ever, while audiences will process less and less, and the gap between those two curves will define the next decade of media. That pressure is already visible in media consumption. Reuters Institute reports that social video use for news rose from 52% in 2020 to 65% in 2025, while overall trust in news remained flat at 40%. Reuters Institute Reuters Institute When people say, “AI will make content better,” I translate it into something more realistic. AI will make content louder. So the scarce resource will not be production. It will be filtering, taste, and trust. This is why paid newsletters got a second life. People are not paying for more information. They are paying for the right exclusions, for a mind they trust to decide what is worth their limited attention. Volume is cheap. Meaning has a cost. That willingness to pay remains limited overall, which makes trusted curation more valuable, not less – the proportion paying for any online news stayed at 18% across 20 richer countries in 2025. 18% 18% Why Reels Are Not a Button Why Reels Are Not a Button A short, high-quality Reel looks like a minute of effortless life. Under the surface, it can take far longer to produce than the finished runtime suggests. The manual workflow is rarely glamorous. It keeps repeating itself, and it keeps returning to you later, as unfinished mental residue. You review twenty to forty minutes of footage and mark what is visually usable. You cut and discard, then rewatch the same seconds until doubt fades. You shape pacing and emphasis until it matches a goal that is often challenging to phrase. You review twenty to forty minutes of footage and mark what is visually usable. You cut and discard, then rewatch the same seconds until doubt fades. You shape pacing and emphasis until it matches a goal that is often challenging to phrase. It is also not just time. Good editing requires a strange overlap of instincts, and teams usually split them across people, then pay for that split through feedback cycles. across people across people A visual eye that can sense framing, motion, and awkwardness. Domain sense, because “important” depends on what the video is actually about. Marketing intuition, because a clip is a promise, not only a record. A visual eye that can sense framing, motion, and awkwardness. Domain sense, because “important” depends on what the video is actually about. Marketing intuition, because a clip is a promise, not only a record. The irony is that the hardest part is not cutting. Deciding what should exist at all is the hard part. The Human Part of the Pipeline The Human Part of the Pipeline I have started treating automation as a contract with consequences. The system earns from repetitive watching. I keep the part where the meaning gets authored and signed. The arrangement is simple. Software watches everything, segments it, ranks it, and proposes options. A human decides what the clip implies, what it emphasizes, and what deserves a signature. If the choice resists explanation, it resists trust. To keep this workable, I separate measured work from authored work. Measured work behaves like physics. A system can find shot boundaries, track objects through time, estimate stability, flag moments where something abruptly dominates the frame. Those signals make raw footage navigable and shrink the search space. Authored work begins when signals end. Intent and context spill beyond timestamps and change the meaning of the same evidence. Publishing carries consequences that live outside bounding boxes. A human belongs in the loop as an author who stays accountable. Evidence Before Taste Evidence Before Taste Auto-editing is often sold as a single model that returns the best moments. Real footage breaks that fantasy fast. A safer approach uses layered decisions. layered decisions layered decisions First comes structure. Split the timeline into coherent units using visual cuts, audio changes, and transcript anchors. This produces a neutral map of the video. structure structure Then comes evidence. Track entities frame by frame, and compute motion from bounding boxes. One distinction matters more than people expect: separate camera motion from scene action. When many tracked objects share similar displacement, the shared component approximates camera motion. Subtract it, and the residual motion belongs to the scene. Shaky footage stops impersonating energy. evidence evidence A second trap is dominance. A clip can look stable and still feel unusable when a foreground object suddenly fills the frame. Dominance can be measured as the area ratio of the largest entity plus the growth rate of that ratio. Sudden surges often correlate with visual aggression. Next comes candidate generation. Long scenes remain useless for short-form. Slide short windows through stable regions, and treat each window as a candidate. Score each candidate using separate sub-scores for camera stability, scene action, and dominance pressure, then combine them with weights you can inspect. Legibility matters, because a score that resists plain explanation becomes a liability at the moment of publishing. Only then does language enter. Intent arrives as text, from a prompt or brief. A language model can validate alignment by reading the intent plus structured summaries of candidates. Its job is comparison and reasoning, and it operates on metadata, summaries, and signals, never on raw pixels. language language What to Keep Human What to Keep Human Once selection becomes automatable, the tempting next step is automating presence. Digital avatars already speak in a familiar style. It sounds like democratization, publish while you sleep, publish while you travel, publish while you are sick. It also multiplies intensity, and intensity changes ethics. If someone interacts with an avatar and it harms them, the consequence spreads across a chain. The platform that hosts the interaction. The party that built the avatar and the surrounding workflow. The person whose likeness was used. The user who arrived in a fragile state. The platform that hosts the interaction. The party that built the avatar and the surrounding workflow. The person whose likeness was used. The user who arrived in a fragile state. Agency diffuses until accountability dissolves into ambiguity. In critical moments, the final okay belongs to a human who understands the domain and accepts the consequence. That balance is already becoming visible in support work, where voice AI can handle more of the exchange, while responsibility still rests with people. That principle is no longer only philosophical: article 14 of the EU AI Act requires human oversight for high-risk AI systems, and NIST’s AI Risk Management Framework treats human-AI interaction as part of trustworthy deployment. voice AI voice AI article 14 article 14 I learned this lesson in a way that started as comedy and turned serious fast. I have watched a model fabricate a convincing citation to a paper that never existed. When challenged, it sometimes corrected itself, and the risk still stayed. The model behaves like a top student who cannot bear saying, I cannot know. The danger is the user who forwards it upstream without understanding the field. This is not an edge case. Recent research and reporting continue to document fabricated or inaccurate AI-generated references across major language models. to document to document AI multiplies competence. It also multiplies incompetence. Cognitive Resistance Cognitive Resistance Selection is a cognitive sport, and outsourcing it carries a cognitive cost, especially in short-form video. A model can process long stretches of footage and propose likely highlights with steady confidence. That is useful for the brutal part, scanning, segmenting, flagging shaky shots, and surfacing clean two- to five-second candidates. The risk begins when I start treating those candidates as answers rather than raw material, and approval becomes my main skill. process process I notice two warning signs. One is reflex – I open the model before I have even written the brief. The other is impatience – five minutes into reviewing clips, I crave a shortcut, as if judgment were only friction. So I keep a low-tech constraint tied to the timeline. Thirty minutes on paper first. I define what the reel must communicate, what it must avoid, and what “good” means for this audience. Only then do I let automation compress the watching and propose options, and only then do I use a language model to challenge my brief, suggest alternatives, and stress-test intent. For Generation Beta, born from 2025 to 2039 and projected to make up 16% of the global population by 2035, this resistance will matter even more. The labor market is moving in the same direction. Check this – The World Economic Forum identifies AI and big data, analytical thinking, and creative thinking among the skills rising fastest in importance. Generation Beta Generation Beta skills rising fastest skills rising fastest Architecture Creates Ethics Architecture Creates Ethics The internet was built for transmission. Accountability never became a first-class feature. When there is no strong notion of personhood in the architecture, you lose constraints that usually shape behavior. You can watch incentives rewrite culture in real time. Open a fresh video account and scroll. The system often raises the temperature because it optimizes time spent, and emotional spikes stick. Negativity travels faster than achievement, less because people are broken and more because attention is hackable. If AI makes content cheaper and engagement remains the metric, the question is more than technical. What prevents the feed from becoming a casino for the nervous system? That is design, and it is ethics, because design decides what gets rewarded. Staying Human Staying Human Automation can shrink the watching, yet attention stays finite, and content factories will keep printing noise. In that economy, the real scarcity is discernment, and discernment requires ownership. Use machines for signals and drafts, keep authorship, intent, and responsibility in human hands, or the signature disappears.
