Week 9: Synthetic media I — what a generative model actually learned to copy

A guest lecture on what an image or voice model is actually imitating when it generates something convincing, and why that's a different question to "is it real."

21 September 2027

A generative model producing a convincing face or voice hasn’t learned “reality” — it’s learned the statistical texture of its training data convincingly enough to satisfy a viewer who isn’t looking for anything in particular. That’s the same asymmetry from week 2’s connoisseurship lecture, run at industrial scale and without a human forger’s own attention or fatigue: a model doesn’t get tired of getting every detail right, but it also doesn’t understand which details a specific viewer might check. Naledi Osei, who builds detection tooling for broadcasters and archives, opens the semester’s two-week move from physical materials into pixels — same questions about technique and tell, now asked of a system that never held a brush.

Outline#

• what a diffusion or GAN model is actually optimising for, in one sentence
• where synthetic media still fails in ways a physical forger wouldn’t (or would, for different reasons)
• the archive’s problem: detecting at scale, not one suspicious object at a time
• this week’s studio begins Assignment 3, “The Tell”