Can We Outsource Imagination? (justinpaulson.com)

I tested OpenAI’s Sora 2 — a combined video-generation model and social platform where every clip is 100% AI‑generated — and found a useful but worrying pattern. The app makes stunning, surreal short scenes easily, and social remixing lets users iteratively alter clips with tiny prompts (e.g., “now make it a bunny”), producing increasingly absurd variants. That casual remix culture can rapidly degrade output quality and coherence, and the next step OpenAI could take is even more ambitious: on‑demand, full‑length, recastable IP (imagine inserting yourself into Star Wars or a bespoke Netflix film generated instantly for your family). That trend matters because it exposes a core limitation: large language and generation models are superb pattern‑matchers optimized for predictability, not for radical novelty. In practice that shows up as homogenized art styles, smoothing artifacts, and conservative idea generation — the “most likely” five to ten answers rather than genuinely new concepts. Technical implications include quality being tightly coupled to prompt craft, risk of cultural convergence when entertainment is machine‑generated at scale, and thorny IP/licensing questions as models recreate existing works. The upside is real: AI can amplify human creativity, but to avoid bland uniformity we’ll need human-driven imagination and intentional curation, not wholesale outsourcing of storytelling to models.