Week 2: The anatomy of a tell

Extending the deck's five tells with three more, and the mechanics of why each one is the specific shape it is

1 March 2027

Week 1’s deck gave you five tells: the uncanny hand, the morphing logo, the affectless narrator, physics that almost holds, the infinite zoom into nothing. All five share a mechanism worth naming explicitly: the model has no object permanence, only a frame-by-frame best guess informed by everything similar it saw in training, and a tell is the visible seam where that guess resets. This screening adds three more tells to the taxonomy, and asks the question the deck left implicit — why these failures, and not others.

Tell #6 — The Overcorrected Smile#

Count the teeth . AI-generated grins tend toward too many, too even, and too uniformly lit — no gum recession, no slight overlap, no incisor that sits a fraction behind its neighbour. The model has seen a hundred thousand dental photographs and stock-smile close-ups and almost no candid, badly-lit human mouths, so its prior on “smile” converges on the photogenic average rather than on any specific person’s teeth. A face that’s otherwise well-rendered will still hand you a smile that reads as a smile from across the room and as an orthodontic diagram up close.

Tell #7 — Temporal Identity Drift#

A face, a logo, or an object that is subtly not the same object between two frames a second apart — not a glitch, a reassignment. The earring moves to the other ear. The birthmark relocates. Nothing jumps or corrupts the way a damaged file would; the model simply re-derives the frame from its prior rather than from the frame before it, and the two derivations don’t agree. This is the tell hardest to point at in a still image, because it only exists across time — which is exactly why the specimen shelf insists on links, not screenshots.

Tell #8 — The Too-Smooth Cadence#

The affectless narrator (deck, Tell #3) is about what the voice fails to do — perform stakes. This is about how: text-to-speech models trained on audiobook and podcast corpora produce speech with no breath sounds between clauses, near-constant syllable timing regardless of clause length, and pitch contours that rise and fall on a schedule rather than in response to meaning. A human reader speeds up when excited and runs out of air on a long sentence. The model does neither, because it isn’t reading — it’s sampling a distribution over “narrator voice” that has no lungs to run out of.

Case study: Radiant Smile Dental Co.#

A recurring specimen on past semesters’ shelves: a thirty-second ad for a dental clinic that does not, as far as anyone has established, exist at the address given. The clinic’s own logo resolves correctly in the opening shot and drifts by the closing shot — three letters becoming a plausible four-letter shape that isn’t a word (Tell #2). Every patient in the before/after montage has Tell #6’s smile. The narrator reads “same-day appointments available” with the identical rising-then-falling pitch it used for “financing options,” because the model has no reason to know one clause matters more to a listener than the other.

Why the taxonomy keeps growing#

Eight tells now, not five — and next year’s cohort will find some of these eight fixed and have to name new ones, because each tell is a specific, patchable gap between what the training data emphasised and what a viewer actually checks. That’s the argument for treating this as a taxonomy exercise rather than a fixed checklist: the skill this week builds is how to notice a seam, which outlasts any specific seam this semester happens to catalogue.