The more AI characters I work on lately, the more strongly I feel this: We are getting better and better at “making people,” but we are not yet very good at “raising them.” Creating a stable AI character is already far easier today than it was a year ago. The face can remain consistent. The voice can be fixed. The wardrobe can be unified. Even across different settings and shots, the character is becoming less likely to fall apart. So it is natural for many people to think: The character is done. But I often ask my team one question: If you replaced this face, what would remain of the character? Often, the room goes quiet. Because apart from the face, there is almost nothing. That is what I have become increasingly wary of: Facial consistency solves “recognition.” Character consistency solves “belief.” There is a long distance between the two.
In traditional film and television, an actor naturally preserves many things for a character. Their body. Their expressions. Their voice. Their habitual movements. The emotions carried over from the previous scene. Even during production, the actor gradually comes to understand the character. When a character is filmed for three months, they are not born again every day. But AI is different. A model does not inherently know: Yesterday, this character lost a friend. Three episodes ago, they promised someone something. Why they are afraid of a particular place. Why they pause when they see a certain object. If each time we give it only one reference image and a new prompt, then from the model’s perspective: It may be creating this person anew every time. The face is the same. But the person may not be.
That is why I have gradually stopped thinking of a “character profile” as a single character sheet. Age. Occupation. Personality. What they wear. What colors they like. Those things are useful, of course. But they are static information. What truly begins to make a person feel real is experience. What they have encountered. What choices they have made. What consequences those choices created. What has happened between them and other people. What they know. What they do not know. What questions remain unresolved. That is a character’s past. And what is often most interesting about a character is this: Choices from the past limit choices in the future. Someone who once betrayed a friend should not act as if they were meeting that friend for the first time when they see them again. Someone who once lost something important because they were afraid makes their next choice in the face of danger meaningful. Without a past, a character is only a premise. With a past, a character begins to have a fate.
My team and I later began using a simple concept: Character Memory. A character’s memory. It may sound technical. But it is not complicated at all. It simply asks that, after making each piece of content, you do not immediately forget it. What has this character just gone through? How has their relationship with someone changed? What decision did they make? Does that decision still count going forward? What should carry into the next piece of content? At the beginning, you do not even need a complex system. A spreadsheet can work. What matters is this: You are no longer recreating the character in every video. You are continuing that character’s life. Those are two completely different approaches.
AI is especially good at creating change. IP is most vulnerable to change without a reason. That is also why many AI accounts are striking at first glance, but become tiring after ten posts. The first post: This character is so cool. The second: Another setting. The third: Another world. The fourth: Another identity. Every post demonstrates what else the model can do. But the audience never knows: Who is this person, really? A character with real vitality cannot keep drawing an audience forever just by becoming “stranger” next time. They need relationships. They need desires. They need problems. They need a past. They may even need to make mistakes.
When I look to see whether a character is beginning to stand on its own, I watch what happens in the comments. If people keep asking: “How was this AI made?” “What model did you use?” “What was the prompt?” Then the technology is still in the foreground. But if the comments begin to say: “Why didn’t he tell her?” “I think she already knows.” “Did something happen between the two of them before?” That makes me happier. Because the audience is beginning to forget the technology. They are beginning to discuss the characters’ relationships. This may be a meaningful threshold for an AI character: From “How was this generated?” to “What will happen to them next?”
That is why I now think the real moat of an AI-native IP may not be its face. Perhaps not even its worldbuilding alone. It is shared memory. A character and its audience go through a hundred things together. The audience knows their past. They know their friends. They know their enemies. They know the mistakes they have made. They know jokes that only longtime followers understand. These things cannot be immediately replicated the next day by generating a more beautiful face. Because they take time. This may be one reason traditional IP can endure for decades. A truly valuable character is not complete the moment it is designed. Through stories that keep unfolding, the audience comes to know that character little by little.
I often use a few simple questions to judge whether an account is beginning to become an IP. Do users know the character’s name? Can they describe the character’s personality? Do they know the character’s relationship with another character? Do they remember what happened to the character before? Can they roughly guess what the character would choose when faced with something? If none of these questions can be answered, I would remain cautious even if the account has high view counts. Because users may remember a particular video. Not the person. And the value of IP lies precisely here: After the video ends, the person remains.
So I think the genuinely difficult question for AI characters in the future will not be: How do you make them live for one video? But rather: How do you make them live for three years? What does three years mean? It means that this person will have a history. Relationships will change. The world will expand. The audience will grow older. The creators themselves will change, too. If, with every model upgrade, we remake the character from scratch, then it will never have a past. So compared with “character consistency,” I would now even rather use another term: Character continuity. Consistency more directly answers: Does the character still look like the same person? Continuity answers: Is the character still the person they were yesterday? This may be the real dividing line as AIGC characters move from “digital images” toward characters with a life of their own. A face lets us recognize a character. A past lets us believe in that character. So the next time we create an AI character, I think that beyond asking: What do they look like? We should also ask: What past do they have? More importantly: Will what happens today become part of who they are tomorrow?
Copy this essay for someone thinking about AIGC content, IP, organisation or digital identity.
