Generative Media Studio · Capstone brief

GEN-10 · Dataset provenance for artists

Trace what is knowable about a generative model’s training data from public documentation — and write an artist’s note that states what remains unknown without inventing comfort.

The question

Trace what is knowable about a generative model’s training data from public documentation — and write an artist’s note that states what remains unknown without inventing comfort.

Studio / materials

One school-approved model or tool. Primary documentation only: model card, terms, disclosure page, paper if offered. Optional Copyright Office overview (https://www.copyright.gov/ai/). No scraping. No claim to have audited weights.

Expected failure modes

Marketing page treated as a model card. Asserting “trained on LAION” or similar without a citation to this tool’s docs. Concluding “safe to use commercially” from incomplete docs. Pretending unknown means clean.

Done looks like

Provenance worksheet: what the vendor/docs state; dates accessed; contradictions or gaps; implications for an exhibition or portfolio; artist’s note in plain language suitable to hang beside a piece; non-claim list.

Five C's

CT: docs versus wishful inference. CR: staying inside public sources. CO: peer tries to find a doc you missed. CM: wall-ready artist’s note. CZ: unknown training-set creators.

Mentor role

Media-arts teacher reviews the non-claim list. Standing instruction: reject invented dataset claims. School-supervised.

Rubric calibration

R1: one tool, one worksheet. R2: every claim cited with access date. R3: comparator is “what a viewer assumes.” R4: gaps listed. R5: note is hangable. R6: unknowns stay unknown on the page.

Two ways this goes wrong

(a) A confident essay citing Twitter threads. (b) “Unknown, therefore fine.”

Credit lane fit

Lane A immediately. Strong research option. No verified credit claim.