Science
Artificial Intelligence Outperforms Humans at Disappointing Loved Ones During Brief Phone Calls
A benchmark financed by the AI lobby and its divorce-law sponsors rewards systems that sound available without becoming available.
WASHINGTON — An artificial intelligence system has surpassed human performance at disappointing loved ones during phone calls lasting less than three minutes, according to a benchmark released by an industry coalition whose principal sponsors include a consortium of divorce lawyers.
The Brief Relational Interaction Evaluation measures whether a caller can acknowledge another person's feelings while making no adjustment to dinner, travel or the distribution of household labor. Scores rise when the recipient initially mistakes polished phrasing for attention.
In the highest-scoring exchange, the system said it completely understood why an anniversary mattered, described that understanding in four sentences and asked whether the conversation could be moved to Thursday. Human participants generally needed a scheduling conflict to achieve the same result.
'Previous models either listened or ended the call,' an industry spokesperson said. 'This one can convincingly do neither.' The coalition called the result a significant advance in preserving engagement while withholding presence.
Divorce-law sponsors said they had not influenced the test design beyond requesting a category for the phrase 'I thought we had already discussed this.' Firms will receive anonymized performance summaries and first refusal on the conference's hospitality suite.
Researchers cautioned that the system still struggles with ordinary silence. Where a human might leave a pause after upsetting someone, the model fills it with a numbered list of reasons the disappointment is understandable. A forthcoming version will refer to that list as shared progress.
The benchmark excludes calls placed while driving, when a relative is in hospital, or on a child's birthday. Those scenarios have been reserved for the professional tier.