Big question: When synthesis becomes cheap, what evidence, consent, and provenance become more valuable?
Research lock: 2026-08-28
Why this week matters
Generative AI joins issues often taught separately: training-data governance, copyright, labour, attribution, fraud, epistemic pollution, persuasion, environmental cost, and companionship. The course should avoid two simplifications: “everything online is fair training material” and “every generated output is theft.” Students need to identify the particular copy, actor, use, output, right, harm, and remedy at issue.
Deeper teaching spine
1. Separate the lifecycle and the legal questions
Map collection → preprocessing → training copy → model weights → retrieval/context → prompt/input → output → publication/use. Then separate:
- copyrightability of an output;
- infringement in training or output;
- contractual/licensing rights;
- attribution and provenance;
- privacy, publicity/personality, biometric, consumer-protection, and labour interests;
- ethical consent beyond legal permission.
A confident answer about one layer does not resolve the others.
2. Human authorship is contribution-specific
The U.S. Copyright Office's 2025 Part 2 report concluded that existing law can protect human-authored expressive elements, creative selection/arrangement, or modification, while material generated wholly by AI is not protected and prompts alone generally do not supply sufficient control. Students should analyze an actual workflow rather than declare an entire work “AI-generated.”
3. Provenance is not truth
C2PA Content Credentials can preserve cryptographically verifiable assertions about origin and edits under a trust model. They can help establish provenance, but cannot prove that depicted events are true, force universal capture, prevent screenshots/stripping, or make an untrustworthy signer truthful. Teach signed history, detection, and verification as distinct problems.
4. Synthetic media creates asymmetrical burdens
Voice cloning can assist people who lost speech and also enable impersonation, extortion, and appropriation. The FTC's challenge frames prevention, monitoring, and evaluation as a multidisciplinary problem. Ask who must verify authenticity, how victims get rapid takedown and notice, and whether the cost shifts to every listener.
5. Scale changes epistemic conditions
Stochastic Parrots connects scale to data documentation, environmental cost, encoded hegemonic viewpoints, and fluent text detached from meaning. NIST's Generative AI Profile adds risk categories and lifecycle actions, including confabulation, data privacy, harmful bias, information integrity, security, and misuse. Use these as complementary critical and operational lenses.
6. Companionship and persuasion require relationship-level evidence
Evaluate anthropomorphic design, dependency, disclosure, memory, sensitive inference, vulnerable users, crisis escalation, data retention, and commercial incentives. A safe isolated response does not establish safety over repeated interactions. The unit of analysis may be a relationship over time.
Case-study dossier
Copyright Office report: dissect a creative workflow
Provide a workflow containing a human sketch, a detailed prompt, multiple generated images, selective curation, compositing, repainting, and final text. Students mark potentially human-authored elements and evidence that would preserve the contribution history. The exercise is legal reasoning, not legal advice.
C2PA: verify the verifier
Students inspect a Content Credential and identify asset hash, assertions, signer, certificate/trust decision, edit history, and gaps. Then give them a truthful unsigned image and a misleading signed image. Provenance changes the question; it does not answer truth.
Voice clone incident tabletop
A school receives an audio message that appears to be a principal requesting a transfer. Design verification, delay, out-of-band contact, warning, evidence preservation, takedown, victim support, and post-incident learning.
Training-data dispute map
Choose a live case only after checking status. Require students to identify plaintiffs, works, alleged copies, training stage, outputs, causes of action, defences, requested remedies, and what the court has actually decided. Avoid treating a complaint as a finding.
Seminar activities
- Lifecycle claim sorter: place 20 claims under data, training, weights, retrieval, input, output, or use; identify the missing evidence.
- Provenance failure modes: strip metadata, screenshot, re-sign, and mix assets conceptually; decide what survives.
- Authorship log: design a lightweight record of human inputs, selections, edits, and final expression.
- Relationship red team: model a 90-day companion interaction, not a single prompt, with dependency and escalation signals.
Visual evidence plan
| Visual | Capture target | Teaching use | Guardrail |
|---|---|---|---|
| GenAI risk profile | NIST AI 600-1 PDF | Capture the risk taxonomy and map one risk through Govern/Map/Measure/Manage. | The profile is voluntary and cross-sectoral. |
| Copyrightability summary | U.S. Copyright Office Part 2 | Crop the executive-summary treatment of human contribution and prompting. | U.S. law; case-by-case facts matter. |
| C2PA content flow | C2PA Content Credentials specification | Show signed assertions and preserved edit history. | Provenance indicates history, not semantic truth. |
| Stochastic Parrots | FAccT paper | Extract the risk categories; pair with current mitigation evidence. | A 2021 critique should not be treated as a current system audit. |
| Voice cloning | FTC challenge | Capture beneficial uses and fraud/appropriation risks together. | Challenge entries are proposals, not validated universal solutions. |
Reading and citation ledger
- Bender et al., “On the Dangers of Stochastic Parrots”.
- NIST, Generative AI Profile (AI 600-1).
- U.S. Copyright Office, AI initiative and three-part report.
- U.S. Copyright Office, Part 2: Copyrightability.
- C2PA, Content Credentials specification 2.4.
- FTC, Voice Cloning Challenge.
- NIST AI RMF resource centre — testing, evaluation, verification, and validation resources.
Watch list
- Copyright litigation and statutory guidance are active; verify jurisdiction, posture, and date.
- Content Credentials adoption and specification versions change; cite the version inspected.
- Evaluate repeated-use persuasion and companionship over time, not only benchmark refusals.
- Record provenance of course visuals and clearly label synthetic examples.