A while ago, Doubao and Qianwen successively announced the discontinuation of their AI agent features. The timing of these two products' decisions coincided with the official implementation of the "Interim Measures for the Management of Anthropomorphic AI Interactive Services" on July 15, 2026.
The new regulations primarily oversee AI services that simulate human personalities, thoughts, and communication styles, engaging in continuous emotional interaction with users. Neither Doubao nor Qianwen publicly confirmed that the discontinuation was entirely due to the new regulations, only stating it was a product feature adjustment. However, judging from the actions of leading products, product forms such as role-playing, emotional companionship, and user-created personalities are clearly entering a phase of contraction.
Many may therefore conclude that AI's emotional interaction is disappearing. But I believe it is only an old product form that is vanishing.
AI agents like Doubao and Qianwen, which mainly sell personas and sustained relationships, are essentially a type of role-playing game. Their biggest problem is guiding users into their scripts, satisfying users' transient fantasies.
On the other hand, the recently launched GPT-Live demonstrates a more advanced interaction design approach. It does not equip the AI with a more complex personality but instead responds to users' emotional need of "longing to be seen" by reacting to voice, pauses, and conversational rhythm.
Character-based AI is a performance
Over the past few years, most anthropomorphic AI products have adopted similar designs. First, assign the AI a predetermined role—it could be a sweet girl, a psychological counselor, a historical figure, or an anime character, or even let users create their own character template.
Platforms also add long-term memory so it remembers past conversations, maintaining a continuous relationship. The core of this design is to make the AI sustain a performance.
To maintain its persona, the AI must speak in a way consistent with its character, display the emotions expected of that role, and address and treat users according to a pre-set relationship.
The AI constantly needs to prove who it is.
A gentle person must always be gentle; a sharp-tongued person must always be sharp. A lover must constantly express care; a psychologist must continuously perform understanding of you.
When users enter the product, they are also drawn into this performance. Choosing a psychologist role means accepting its pre-defined counseling relationship, so you have to cooperate in this fictional relationship.
This only superficially accompanies the user.
They keep proving who the AI is, yet rarely assess what the person on the other side of the screen truly needs at that moment.
GPT-Live does not have a more complex personality
In my experience with GPT-Live, I saw a different design approach for interaction.
During my recent trial of GPT-Live, the first thing I noticed was not that it had become smarter, but that it actively used filler words.
When I kept speaking, it produced sounds like "Mm-hmm," "Right," and "I'm listening." Sometimes it actively interrupted me; other times it tried to interject. When it realized my silence was just me organizing my thoughts, it paused and waited.
Even more surprisingly, when I was thinking and instinctively said "Hold on," it actually responded with "Okay, take your time to think."
From a text interaction perspective, these behaviors are hard to understand. Even by traditional large model evaluation standards, those filler words like "uh" and "um" would be considered unnecessary fluff.
But in human conversation, these "fluff" words play an important role. Their underlying message is, I am listening.
It tells users that your words haven't fallen into an input box waiting to be processed; "I and you" are having a conversation in the same factual space.
GPT-Live neither adds a complex identity story nor repeatedly tells users what personality it has; it simply focuses more attention on the user.
Humans' deepest emotional need is to be seen
In experiencing GPT-Live, one can see that the model begins to receive and respond to signals that were easily ignored by systems in the past.
Your speaking speed, your pauses, changes in your tone—these are "subtext" that people keenly capture in real communication. Now, the model also learns to use these signals.
In the past, traditional AI interactions filtered out such subtext as noise and then generated a superficially correct answer.
But from an experiential standpoint, what truly makes interaction feel human is not how human-like the AI's lines are or how sultry its voice sounds, but whether it can understand those "fluff" words that are never meaningless in human communication.
Humans' deepest emotional need is to be seen.
Information systems try to filter out noise, but human relationships often rely on that noise to confirm each other's presence.
When this noise—which should have been filtered out—can be captured and responded to by the model, the user's emotional needs are truly addressed. Only then does anthropomorphic interaction truly come into being.
Personalized "agents" will disappear, but interaction design based on users' emotional needs will only become more "authentic."