Temporal Telephone — Learning Machines

Motion Defaults — Video Default Test

This is the video version of the Default Test. The prompt is vague, but motion still becomes specific. After playback, check what the group made visible from "A person at work."

JobWhat kind of work appeared?

BodyWho did participants imagine?

SettingWhere did the work happen?

ToolWhat object or technology appeared?

PaceWhat movement or speed emerged?

GestureWhat repeated action appeared?

Reflection — Human · Machine · System · Ethics

Human

Which ghost mode did you use? What decision did you make at each frame? When did you choose to drift versus anchor? What would you do differently in Round 2?

Machine

Where did the sequence drift furthest from the prompt? At what frame did identity first break? Which ghost mode produced the most coherent sequence — and why might anchoring help?

System

A video model generates each frame conditioned on the prompt and sometimes the previous frame — just like Run A. What would it need to maintain identity across 100 frames? What system choices make coherence more or less likely?

Ethics

If a video model generates a person whose appearance drifts across frames — different skin tone, age, or expression — who is affected? What are the risks of identity drift in real-person video generation?

Bridge to AI video failure modes

Identity drift — a character's face, age, or appearance changes across frames. Your sequence likely showed this.
Motion continuity break — an object or body part teleports between frames. Caused by the same frame-by-frame prediction problem you simulated.
Physics break — objects fall the wrong way, cloth ignores gravity, water moves backwards. Each frame is a prediction, not a simulation.
Background teleport — the scene behind a subject changes without camera movement. The model re-predicts the background independently at each step.
Style drift — the visual style shifts across the sequence. No global style memory across frames.
Text corruption — readable text in frame 1 becomes garbled by frame 10. Text is predicted pixel-by-pixel, not recalled as symbols.

Run A (previous frame only) simulates standard video generation. Run B (anchor) simulates using a reference image. Run C (both) simulates hybrid conditioning. Real models use more sophisticated approaches — but the fundamental drift problem is the same.

Facilitator note — Workshop sequence (15–20 min)

Before the tool (5 min) — unplugged version. If time allows: give participants slips of paper. One person draws "a person at work." Pass the drawing face-down. Next person draws what happens next, seeing only a faint trace. Repeat 6–8 times. This makes drift physical before the digital tool.
Round 1 — Run A, telephone (5 min). Use "Previous frame" ghost mode. 6–8 frames. Don't reveal the prompt. Play back. Ask: where did it go? What drifted? Reveal the original prompt, then use the Motion Defaults checklist before moving to failure modes.
Round 2 — Run B or C, conditioned (5 min). Reveal the prompt. Use Anchor or Both mode. Same number of frames. Play back. Compare directly: does anchoring help? What still drifts?
Debrief with the failure bridge (5 min). Show the failure card. Ask participants to match each failure mode to something they saw in the playback. Name the drift precisely.

Key debrief questions:

A video model generates each frame using the prompt and sometimes the previous frame. After Round 1, does that mechanism surprise you?
What information would a model need to keep a character's face consistent across 60 seconds of video?
Run B used a fixed anchor — the first frame. Why might that help? What would it miss?
If this model generated a real person's likeness and it drifted — different skin tone, age, expression — what are the ethical problems with that?
You made intentional choices each frame. The model doesn't. What is lost when that judgment is removed?

Interactive temporal telephone workspace