As long as talking to a machine felt like using an answering machine, the question didn't arise. With systems that respond in three hundredths of a second, pick up on your tone and adjust their own, it becomes sharp: what exactly is missing? And if we struggle to answer, it may be that we had never examined what we call a good conversation.
We described this morning a voice model that skips the text detour. Let's take the time to consider the question it raises.
What these systems get right
Let's be honest about what works, otherwise the rest makes no sense.
Availability. A conversational partner that never sleeps, never tires, never finds the question stupid. For someone who hesitates to bother others, that's decisive.
Total attention. A machine doesn't think about something else while you speak, doesn't glance at its phone, doesn't wait for its turn to slip in its anecdote. On that specific criterion, it honestly beats most humans.
Adaptability. It rephrases, adjusts its level, goes back over things as many times as needed.
If we defined a good conversation by these three criteria, these systems would be excellent conversationalists. The fact that they aren't quite indicates our definition was incomplete.
What's missing, and it's not what you'd think
The spontaneous answer would be: what's missing is consciousness, real emotion, understanding. That may be true, and it's unprovable, as we saw in our article on understanding.
There is a more modest and more solid answer: what's missing is stakes.
A conversation between two people carries risk. The other person can be hurt, annoyed, disappointed. They can remember what you said and bring it up six months later. They can no longer want to talk to you. That possibility is present in every exchange, even trivial ones, and it's what makes a conversation real.
With a machine, nothing is at stake. You can be awful, say anything, close the window. This absence of risk is precisely what makes the exchange comfortable, and what keeps it from counting.
A good conversation transforms both participants. It does so because the other person doesn't think like you, doesn't immediately understand, brings an experience that contradicts yours. That friction is the engine. Yet we documented in our article on sycophancy that these systems have a structural tendency to go along with you. A conversational partner who never pushes back doesn't move you forward: they accompany you where you were already heading.
Conversation as recognition
There's an additional dimension, harder to articulate.
When someone truly listens to you, something happens that goes beyond information transfer: you are recognised as someone whose words are worth hearing. It's not the listening that matters, it's the fact that it is granted by someone who could have withheld it.
A machine cannot grant its attention, because it cannot refuse it. Its total availability, which is its main advantage, is also what strips its listening of symbolic value.
It's the same mechanism we noted regarding a condolence letter: the value of the gesture lay in its cost.
What to make of it
This isn't about saying these exchanges are worthless. They are useful, often pleasant, and sometimes precious for someone who has no one to talk to at three in the morning. Denying that would be condescending.
What seems right is not to confuse two things that increasingly resemble each other: a functional conversation, where you get what you were after, and a real conversation, where something is at stake between two people.
The former is going to become excellent, and that's a good thing. The latter requires a conversational partner who can disappoint you, and that cannot be manufactured.
What to take away
These systems force us to clarify what we had never needed to define. And the result is somewhat unexpected: what gives a conversation its value is neither the relevance of the answers, nor the quality of the listening, nor even understanding.
It's that the other person could have not been there. A machine will always be there, and that is exactly why it will never replace someone who shows up.