實體 AI：聽見説話與理解請求並不相同

Social capabilities across human development | Part 11

Listening to the end of someone's statement matters, but receiving words does not establish shared meaning. Listening also includes checking understanding, responding in a correctable form, and reflecting corrections in action.

The blue bowl, not the red one

In a kitchen, someone asks, "Could you pass that bowl?" The other person reaches toward a red bowl but pauses to ask, "This one?" The requester points to a blue bowl. It is handed over, and cooking continues.

In this fictional scene, merely hearing and executing the words would have produced the wrong object. Clarification jointly resolves ambiguity. The correction is used as information for the task rather than treated as defeat or blame.

The difficulty of applying understanding

Apperly and colleagues investigated adults' difficulties interpreting instructions in relation to another person's perspective. Their experiments provide material for questioning whether using others' information in action is always automatic. See Why are there limits on theory of mind use?.

A communication task cannot be generalized directly to every household request. This article distinguishes understanding a perspective from selecting a target on that basis, and considers the value of everyday clarification.

The adult experiments cited here do not establish how these abilities change during childhood. In this series, the practical question is how people can make their understanding correctable as situations and communication demands change.

Too much clarification can also impose a burden

Repeating every request at length can increase the other person's workload. Focus clarification on differences affecting the outcome. Ask about color when bowl color matters, or time when timing matters. Avoid repeatedly questioning information the person already made clear.

"Understood" alone does not show what was understood. A brief statement such as "I'll pass the blue bowl" makes a mistake correctable. Clarification coordinates joint action rather than testing the speaker's words.

A design hypothesis for physical AI

A possible design separates speech recognition, candidate meanings, execution targets, and permission to act. Correctly recognized words can still leave the target ambiguous. Defer contact with the target while clarifying, then update the planned target after correction.

Recording a correction in memory while continuing an old movement is not recovery. Test whether conversation and physical commands use the same confirmation result.

Evaluation could include similar objects, reformulation, cancellation, and correction after movement begins. Examine useful clarification and unnecessary clarification burden together. Correcting the target matters more than declaring that the person's mind was understood.

A decision in four panels

1. A blue bowl is requested.
2. The other person starts reaching toward a red bowl.
3. They pause and check the color.
4. They hand over the blue bowl.

This fictional comic depicts bringing understanding and execution into agreement rather than competing in fluency.

References

1. Why are there limits on theory of mind use?

About this series

This series uses published literature to examine everyday situations. Its 7 domains and 22 capabilities are an editorial framework, not a standard academic classification. Differences in age or culture are not rankings of people's social worth or capability. The proposed AI designs do not establish humanlike emotional experience or demonstrated product performance. The comics illustrate fictional examples.

LOGIHEART's research and development vision

LOGIHEART is researching how to implement social capabilities in physical AI that interacts with people.

Explore the research and development vision at the LOGIHEART website and in the English introduction video.