Determine whether users infer intelligence from available rationales

Determine whether users infer an AI assistant’s cognitive capability from the presence of a rationale accompanying its output.

Background

The paper distinguishes between making an explanation available and making that explanation useful or diagnostically informative to users. Although a rationale may expose the apparent basis for an AI assistant’s output, users may not interpret it as evidence of intelligence unless it addresses the questions they care about. The study measured whether a rationale was available but did not resolve whether users actually made the proposed inference from the rationale to the assistant’s intelligence.