Human–AI explanation

Use this profile when a person uses an AI system to prepare, discuss or revise an explanation, or seeks an explanation about an AI system. First distinguish these subjects. An AI’s role in producing text does not make its own operation the thing explained. A user asking why a deployed system produced an output may require evidence about that actual system, its inputs and operation.

The Preface’s account, expression, recipient and use conditions still govern the work. The human can contribute questions, corrections, constraints and preferences. AI-produced content and claimed causal grounds require qualification for the relevant subject. Fluent output, an expressed confidence and a system’s own story about its reasoning cannot establish an otherwise missing causal warrant.

In the pooled-mean exchange, an AI assistant can state the finite-list definition, recover the sums and counts, and answer 6.75 for equal individual weight. If the user changes to equal branch influence, EXD.1/.4 select the corresponding question, answer 5.5 with its label and retain the scope of the first answer. If the user asks only for the established calculation, further elicitation may add no value.

If the user instead asks, “Why did our deployed predictor choose this value?”, the arithmetic case supplies no causal explanation of that model. Recover the relevant model, input and available diagnostic or subject evidence. If that basis is unavailable, name what is missing and limit the answer accordingly. A more polished narrative or the model’s unsupported self-report cannot repair that lack.

For instructional use, the assistant can ask for a focused explanation and preserve the initial response, hint and retry as EXD.5 specifies. The resulting exchange shows what the person produced with that assistance. An AI answering the learner task itself can help inspect a public instruction or construct a candidate response; it is not an observation of the human learner. The profile likewise makes no inference that an AI recipient has human attention, memory or learning mechanisms.

Recognize a promising exchange when the question and subject are selected, available grounds constrain the content, and contributions can change the next move. Assure the actual claims and response at the relevant system, version, access and help conditions. When testing an explanatory interaction, distinguish conversational behavior, immediate objective task performance, confidence, later retention and decision quality.

Current social XAI work supplies joint question formation, incremental explanation and adaptation. Liao and colleagues and Schmude and colleagues inform elicitation anchored in a user’s task, without treating a fixed question bank as complete. Fichtel and colleagues’ 2025 study is a consequential limit: more co-constructive behavior under its prompting conditions did not yield a significant mean objective-understanding advantage. Neither the existence of an interactive system nor a more collaborative style settles the receiving outcome.

Adaptation can make the exchange more relevant while adding turns, reading burden and opportunities for unsupported content. Treating a human user as a fixed knowledge level can miss a changed purpose; treating every correction as a prompt to add detail can retain the wrong question. The useful result is a warranted answer, task-specific repair or honest subject return with interpretable interaction evidence. EXD.6 governs worthwhile local comparison; the instructional and technical/advisory profiles supply their respective stronger receiving questions.