Human–AI collaboration process
A short, blunt description of who decides what.
The human author decides
- Research question, scope, contributions claimed.
- Which sources are admissible and how disagreements between sources are resolved.
- Whether a claim has crossed from
ai-confirmedtohuman-confirmed. - When to publish, where to publish, and who gets author credit.
- Anything ethically loaded.
The AI agents do
- Draft prose, propose figures, condense, audit.
- Maintain
doc/provenance.ttlunder the curator agent's mediation. - Refuse to invent citations or paper over evidence gaps.
Disagreements
If an agent and a human disagree on a factual matter:
1. The agent records its position with a quoted source.
2. The human records the counter-position with a quoted source.
3. Both go into the manuscript as a hedged sentence ("X has been argued
[cite-A] and disputed [cite-B]") and into the graph as two
fair2r:Claims linked by fair2r:contradicts.
Authorship credit
AI agents are not authors. The CITATION.cff lists human authors only.
Substantial AI contributions are acknowledged in the paper's
Acknowledgements and recorded in doc/provenance.ttl as
prov:wasAttributedTo an fair2r:AIAgent individual.
Audit trail
Every working session ends with a logbook entry and (where claims changed)
new triples in provenance.ttl. No silent edits.