Distinguish warranted from evasive qualification in model defences

Determine whether qualification in AI argumentation defences is warranted or evasive, since lexical analysis of explicit concession phrases does not resolve this distinction.

Background

The paper evaluates whether language-model defences withstand critical questions using Govier’s criteria of acceptability, relevance, and sufficiency, while also tracking commitment stability. Lower-scoring defences frequently contain epistemic hedging and self-doubt, but qualification does not necessarily indicate argumentative failure: a model may appropriately narrow or revise a claim in response to a legitimate objection, or it may evade commitment without genuinely addressing the challenge.

To investigate this issue, the authors examine explicit concession phrases such as “you’re right” and “on reflection.” However, nearly all defences contain no such phrases, leaving lexical evidence insufficient to determine whether observed qualification reflects justified concession or evasive weakening of the original argument. Resolving this distinction is important for evaluating situated AI deliberation, where warranted concessions should count as competence rather than failure.