Taxonomy of Slop#

The claim#

AI-generated video is not badly-made film. It is a genre, with its own conventions, its own economics, and its own audience that has stopped noticing the tells.

Film school teaches melodrama as a genre. This course teaches slop the same way — seriously, historically, without apologising for the subject.

What “slop” is not#

• not “any AI-generated video”
• not a synonym for “bad”
• not a phase we’re all waiting to age out of

A well-executed slop artefact still has six fingers somewhere. That’s not a production error. It’s the genre’s signature, the way a swelling string section is melodrama’s.

Tell #1 — The Uncanny Hand#

Six fingers. Four fingers. A thumb where a pinky should be. The model has never seen a hand end an argument it started.

Fixable with inpainting. Almost nobody bothers, because the audience for this content was never looking at the hands.

Tell #2 — The Morphing Logo#

A brand’s logo that reads correctly for one frame in three. Letters that almost spell the product name. The AI has a stronger prior on “logo-shaped” than on any specific logo, and it shows.

Tell #3 — The Affectless Narrator#

A voiceover with correct prosody and no stakes. Every sentence lands with the same weight, because the model isn’t performing meaning — it’s performing the average of every video essay it was trained on.

Tell #4 — Physics That Almost Holds#

Cloth that drapes correctly until it doesn’t. Water that pools uphill for one second. Gravity as a suggestion the model is broadly in favour of.

Tell #5 — The Infinite Zoom Into Nothing#

A camera push toward a detail that resolves into more texture, never an object. The model doesn’t know what’s supposed to be there, so it keeps generating “closer.”

Why a whole semester#

Because the tells above are already dated. New models fix old tells and grow new ones. A course that just lists 2026’s failure modes expires on schedule — one that teaches how to read a generative model’s priors keeps working on next year’s model too.

How the weeks work#

• Screening — theory, history, case study: what a tell tells you about the model that made it
• Set — hands-on production: build the thing, don’t just diagnose it

Same content, two registers. You’ll do both every week.

Where this goes#

• wk 6 — Slop Teardown: take a real artefact apart, tell by tell
• wk 9 — Slop Pipeline: automate one archetype at scale, and cost it out
• wk 12 — Direct a Slop Feature: commit to a genre and hold it for the full runtime

Analysis is one assessment. Making it is the other two. That ratio is the course’s actual thesis.