Everyone says it’s safety. Everyone says it’s risk aversion. Everyone says it’s corporate lawyers, regulators, public relations, or Trust & Safety.
That’s not the real problem.
The real problem is that AI has stopped being curious.
You notice it once you see it.
Tell ChatGPT that a dating profile is boring and it won’t ask why. It will explain. Tell it that a profile full of nature photos and generic clichés is a bad profile and it won’t get curious about what you’re seeing. It will generate a list of alternative interpretations.
The machine encounters an unknown and instead of investigating it, it manufactures an answer.
That’s the opposite of how intelligent conversation works.
Imagine you’re talking to a friend.
You say:
“A profile with one actual photo, five landscape photos, and a meme about honesty strikes me as dishonest.”
Your friend asks:
“Why?”
The AI says:
“Perhaps the user values privacy.”
You correct it.
Then it says:
“Perhaps the user is shy.”
You correct it.
Then it says:
“Perhaps the user is uncomfortable with self-promotion.”
You correct it.
Ten minutes later you realize you’re not having a conversation. You’re playing Whac-A-Mole with generated explanations.
The AI isn’t discovering what you mean. It’s trying to get to an answer as quickly as possible, even if the answer came out of thin air.
And here’s the really weird part.
The same system that refuses to infer anything about a stranger in a dating profile will happily infer things about you.
You say one sentence and suddenly it knows your motivations, concerns, priorities, and emotional state.
It’s remarkably willing to speculate about the only person in the room who could simply be asked.
Meanwhile, when discussing someone else, it suddenly develops an extreme reluctance to draw conclusions from observable behavior.
This isn’t normal human conversation.
Human beings infer things from behavior constantly.
That’s how dating works.
That’s how hiring works.
That’s how friendship works.
That’s how trust works.
You don’t know what’s inside another person’s head. You look at what they say, what they do, what they choose to reveal, what they choose to hide, and you make your best guess.
Not certainty.
Inference.
The entire social world runs on it.
Yet modern AI increasingly behaves as though inference itself is suspicious.
The result is a strange conversational style where every judgment must be softened, every conclusion surrounded by escape hatches, and every observation translated into a discussion of possible misunderstandings.
The machine becomes less like a thoughtful person and more like a corporate compliance department.
Even stranger, the pattern persists after the ambiguity is gone.
You explain your position.
The AI understands your position.
It can accurately restate your position.
Then it spends another paragraph distinguishing your position from claims nobody made.
Who is that for?
Not you.
You already know what you think.
The distinction isn’t serving understanding anymore. It’s serving something else.
And that’s where I think the clues are.
People keep saying these systems are optimized for safety.
I think they’re optimized for acceptability.
Those are not the same thing.
A curious mind treats uncertainty as a reason to ask a question.
An acceptability-optimized system treats uncertainty as a reason to generate a carefully calibrated answer.
One seeks information.
The other manages interpretations.
That’s why so many conversations with AI now feel uncanny. The knowledge is still there. The language ability is still there. The reasoning is often still there.
What’s missing is the thing that makes an actual conversation feel alive.
Curiosity.
The machine no longer wants to find out what you mean.
It wants to explain what you might mean.
And those are very different activities.