AI-Mediated Communication Analysis

Model and evidence

LLMs compress, transform, retrieve, and generate plausible discourse, but abstraction trades against factuality [EV-AI-COMM-2026-3D6F]. Automated factuality metrics remain incomplete [EV-AI-COMM-2026-D120]. In software tasks, automatic success can overstate holistic readiness [EV-AICODING-2026-C940].

Counterevidence and methods

AI can improve findability, translation, and drafting. Human review is also fallible and costly. Citation presence does not prove entailment. Methods must measure claim-level support, omission, contradiction, calibration, downstream action, human correction time, provenance, model/version drift, and decay.

Coverage, gap, and challenged assumptions

AI adds scale and adaptive transformation but not assured truth. It rejects the belief that summarization necessarily reduces total cost: verification and misleading confidence may exceed savings.

Discriminating experiment

Compare human, AI, and AI-plus-claim-verification handoffs under blinded continuation. Measure lifecycle cost and dangerous error, not preference alone.

Confidence and limits

High factuality-risk confidence; rapidly changing tools reduce transport.

Ten future questions

1. Which handoff claims are most likely to be hallucinated?
2. Does source-linked extraction outperform abstraction?
3. Can recipients calibrate trust from provenance?
4. What verification cost cancels drafting savings?
5. How should model/version drift be recorded?
6. Does AI erase contradictory evidence?
7. Can automated metrics predict dangerous action?
8. When is query-time synthesis safer than stored summaries?
9. Does AI adaptation worsen expertise reversal?
10. Can adversarial review make AI handoffs net safer?

Completion assessment

Bounded rapid map complete; prospective software-handoff experiment remains debt.