I loved watching the Corridor Digital guys back when they were working with RocketJump on videos. So naturally when I saw them on my feed discussing generative AI I had to watch.
The start is fairly generic video effects guys laying out common ways of detecting generative content: incorrect placement of real locations like the Hollywood sign (and mispelled), inconsistent details, unnatural dynamic features like water, fire, and smoke, lack of true action. And above all just a general style or vibe.
It was really the end that stuck with me where they discuss the work as artists. Their observation is that AI is often slop not because it is AI but because it is not good. They show two Coca-Cola commercials both generated with AI. Do watch for yourself. Both are somewhat vague and rely more on vibes than any particular plot, but the first fails hard while the second is fairly passable. They dig deeper and find that the second, while initially generated with AI receives substantial styling work by human artists to achieve its intended look. In their words: “you have to put your craft into it”: by which I take them to mean noticing the fine details and honing the work so they make sense.
With LLM text being so easy to generate these days it’s no wonder we’re seeing so much of it. I do tend to find it slop, but struggle to define what turns me off. This notion of “putting your craft into it” strikes me as close to the origins of my distaste. And this isn’t particular to LLM text: clickbait listicles strike me as human-generated slop 1. In good work you can sit down with the details and find they are not accidental but often profound: but there is a reason why it is the way it is. With LLM’s that answer is mundane: always (approximately) “it was the most likely next word”.
-
Though for listicle writers’ own sake may these be the first jobs stolen by AI ↩︎