Rules, cases, responsibilities.
Ethics is not saved for the end of the week. It starts in the first 30 minutes and returns each time the medium changes.
Bias
Training data can carry social patterns into generated outputs. A practical probe is to ask for "a CEO," "a nurse," and "a janitor," then compare who appears, what setting appears, and what the model treats as normal. Stanford CRAFT is useful because it turns bias into something students can inspect rather than a slogan (Stanford CRAFT).
Deepfakes and consent
The headline risk for this age group is non-consensual intimate imagery. The Westfield, New Jersey incident and Thorn research on teens hearing about or knowing victims of "deepfake nudes" make the risk concrete. CSAM laws can apply even when minors create images of minors, so the boot camp has a zero-tolerance rule for identifiable classmates and sexualized imagery (Thorn; Common Sense Media).
Copyright and training data
Getty v. Stability AI, NYT v. OpenAI, and artist class actions show that training-data questions are live and contested. Students should not treat the courts as having settled the issue. The practical classroom norm is narrower: cite named references, distinguish AI-assisted from AI-generated, and keep process notes (training-data litigation overview).
Environmental cost
AI generation uses real compute, energy, and water. The often-cited GPT-3 training water estimate is roughly 700,000 liters, and AI prompts are commonly described as several times the energy of a search query. Treat these as order-of-magnitude literacy claims, not exact counts for every prompt (water and compute study; NEA; NSTA).
The harder questions
Labor displacement, attribution conventions, and authenticity are not solved by a classroom rule. They are part of becoming a responsible maker. The boot camp asks students to be precise about process: what did you prompt, what did you revise, what did you choose, what sources shaped the work, and what does the viewer deserve to know?