Prior art and evidence
What already exists, where it stops, and the evidence against our position.
What exists
| Area | What exists | What is missing |
|---|---|---|
| Survey methods | One method per question, such as best-worst scaling and card sorting | Built for surveys, not for work |
| Task theory | McGrath's task types and Foley's task taxonomy, read through secondary sources | Not built for AI, and not tied to interfaces |
| Choice architecture | Tools for structuring and describing choices | Consumer and policy settings |
| Human-AI research | Levels of automation, 18 guidelines, Shi et al., a survey of 154 papers whose control methods include option selection | They type the automation, the flow or the controls, not the choice |
| AI UX patterns | Shape of AI | Sorts patterns, with no rule for which one fits |
| Agent protocols | MCP elicitation, AG-UI interrupts, A2UI, ACP | They carry the message, not the kind of choice |
| Law | DSA Article 25, AI Act Article 14, GDPR consent | They say what is forbidden or required, not what is good |
| Records | C2PA, CAWG identity, ISO/IEC TS 27560 for consent | Named actors and consent are recorded. C2PA 2.2's listed actions show none for choosing or approving. Version 2.4 is unchecked |
AP2 separates advance intent from final approval, in signed records. We read that through a secondary source.
Evidence against our position
- Choosing alone may not count in law. See the table above.
- Approval prompts can lose attention. Anthropic reports 93% acceptance of permission prompts and says attention fades. Many of those approvals may also be correct.
- People may fail at oversight. Green argues that people cannot perform the oversight functions that policies want, and he proposes oversight by institutions.
- More options is not always better. The evidence is mixed. Parallel prototypes helped human designers, in a small study. Other work found that small panels of AI answers helped.
- People may dislike what works. In one study, the designs that reduced overreliance were rated least favorably.
What the law says about each type
This wiki separates being proud of a work from owning it in law. The US Copyright Office report, gives this map. It is the Office's view, not a court's, and a Chinese court has gone the other way.
| Type | The Office's view |
|---|---|
| Steer | Weak. Page 20 calls revising prompts "re-rolling" the dice |
| Pick one | Weak. Page 21 says selecting an output is not authorship |
| Mark up by comment | Weak, by the same reasoning |
| Edit | Stronger, case by case (page 27) |
| Combine and Arrange | Stronger at scale. Choosing two or three elements generally fails (pages 24-25) |
So the Shape family matters most for law. Our thesis is about pride and meaning, and that is a wider claim than copyright.
A rule on pre-selection
NN/g, GOV.UK and EU law seem to conflict. They answer three different questions: how a radio button works, whether people miss a question, and what counts as consent.
Our rule: pre-select only what is reversible. Never pre-select a removal unless the item can be recovered. Never pre-select consent.
Where this leads
We found no standard for this, so the wiki could propose one in three parts:
- A vocabulary: these choice types, open for any product or agent protocol to use.
- A bar: the test for a real choice, and the six parts of a moment.
- A record: an open format saying which choices a person made in a work. Existing records name actors and consent, and we found none that records this. It could support the claim that a work is truly co-made.