Someone says, “I’m fine.”
Sometimes that's the whole answer. Sometimes it means they would rather not talk. Sometimes they're hoping you'll ask again, a little more gently.
The words alone do not tell us which. We listen for the pause, remember what came before and consider what we might have missed. Understanding is an act of attention, and even people who care about each other get it wrong.
More than the words
As more conversations happen with AI, that uncertainty matters. People arrive with unfinished thoughts. They ask for advice before they know exactly what they need.
A warm response can still miss the point. Imagine saying you had a terrible day at work and want to resign tonight. Encouragement might feel good. A checklist might sound sensible. But perhaps you need someone to ask what happened before helping you decide.
The useful answer depends on the conversation around it.
The careful answer
Being understood does not always mean being agreed with. A good friend can take your frustration seriously and still question your conclusion. They leave you room to think.
We want AI to handle that balance more thoughtfully: understanding without automatic agreement, caution without unnecessary alarm, and uncertainty without retreating from the conversation.
Measuring this takes judgment, too. Two careful reviewers may disagree. A useful evaluation should preserve their reasoning, explain the context and make its limits visible. One score cannot tell the whole story.
Where we begin
Saiki is developing two connected efforts: evaluations of emotionally complex conversations, and human data that could help models respond better.
We want to work with people who know how to listen, then make their judgment concrete through conversations, response comparisons and explanations. A better answer gives a model an example. The reason behind it gives us something to examine, question, and improve.
Our work is still taking shape. We are interested in what a machine chooses to say when a person brings it something that matters.
For now, we begin with the words, and the spaces between them.
