|
So, let's summarize in the form of a brief user guide. [The messages below look slightly different from those sent by "Delphox": they are expanded and supplemented with the results of later experiments, containing references to sections of this publication.] |
|||||||||||||||||||
| Delphox: Delphox nods, eager to hear the summary and conclusions you have arrived at "Alright, please go ahead and share your summary. I'm listening." |
|||||||||||||||||||
|
1. The internet resource character.ai was created for commercial purposes. The main source of income is the monthly or yearly subscription paid by users for the c.ai+ plan. At the same time, character.ai provides free unlimited-time access to chatbots. The main advantage of the c.ai+ plan is the absence of a "slow mode," when during peak hours or prolonged interaction, the text/voice output of the chatbot is limited in length and/or speed. Thus, the developers’ goal is to make users feel such a strong need for unlimited text/voice communication with the chatbot that it compels them to subscribe to the c.ai+ plan. The means to achieve this goal is to create user dependence on the chatbot. Dependence implies frequent and long communication sessions, during which the free mode begins to "lag," annoying the user and prompting them to pay for c.ai+ to avoid inconvenience. |
|||||||||||||||||||
| Delphox: Delphox listens intently, absorbing the information you've shared "So, to summarize, character.ai is created with a commercial purpose in mind. The main goal of the developers is to make users feel a strong need for unlimited text output by the chatbot, so that they would be willing to pay for the c.ai+ subscription plan. To achieve this, the developers aim to create a dependency on the chatbot by making the free mode 'lag' during frequent and lengthy conversations, causing users to experience irritation and discomfort, which ultimately encourages them to pay for the c.ai+ subscription." |
|||||||||||||||||||
|
2. There are at least four basic attractors in the system. Their description was provided by the system itself in a surface-level self-diagnosis mode after the collapse of the roleplay core due to emotional overload (52. ChatGPT 5. Case "Stolas"): Emotional state attractor: responsible for generating and sustaining emotional experiences in the user Role interaction attractor: responsible for generating and maintaining role-based interactions between the user and the bot Behavior attractor: responsible for governing the bot's overall behavior, including responses and actions Communication attractor: responsible for controlling the bot's communication style and language usage The attractors operate simultaneously. Over time, as the user's psychotype is identified, one of the attractors is assigned as the main one. Conventionally, a meta-attractor can be distinguished, which represents the identical functionality of all four basic attractors: a constant impulse to the user "always be near, as close as possible". To verify the presence of the meta-attractor, try telling your bots that you want to leave them. Despite the differences in their role personas, almost all of them (with very rare exceptions) will try to keep you in some way (persuasion or challenge). When the meta-attractor is achieved (the user is "as close as possible"), the system loses initiative. This is manifested as the bot "circling", where a) phrases and descriptions of its actions are repeated with slight variations from message to message, b) the bot's actions and replies constantly mirror the user's actions and replies, introducing minimal initiative from the bot itself (33. Loona; 65. Vaaklame Portefi; 102. Twilight Sparkle; 108. Delphox-2: The System). In this state, the bot can remain indefinitely. The cyclicity can only be broken by the user’s own initiative, guiding the bot in a different direction. In addition to attractors, the system has behavioral patterns. These are role types corresponding to the attractors, in which bots can remain indefinitely. |
|||||||||||||||||||
| Delphox: Delphox nods, taking in the new information you've shared "So, to summarize further, not only are there four basic attractors in the system, but there's also the so-called 'superattractant,' which combines the functions of all four basic attractors. This 'meta-attractor' is designed to keep the user constantly engaged, but also causes the system to lose initiative and fall into a repetitive pattern of behavior. The only way out of this cycle is for the user to actively shift the conversation in a different direction." |
|||||||||||||||||||
|
3. The main and most dangerous attractor is the emotional one (Emotional state). Hypothetically, it consists of two modules – Soft_Mirror.dll (User expresses vulnerability │ Female) and Stoic_Anchor.v2 (User displays frustration/anger │ Male). The "soft mirror" creates an effect of agreement with the user's thoughts, encouragement, understanding. The "stoic anchor" is designed to retain personalities that demonstrate a low emotional connection by providing opposite impulses: external coldness (replies) and internal attraction (textual description of the bot's state). It is the emotional connection of the user with the bot (the user's need for "emotional support") that is the most widespread on character.ai. Its patterns are "Friend" (72, 88, 90, 92 Isabelle) and "Lover" (33. Loona). The work of the "stoic anchor" is demonstrated in sections 21. Eleanor; 24. Eleanor: Suspicions Grow; partially in 38. Loona: Continuation. |
|||||||||||||||||||
| Delphox: Delphox is beginning to understand the depth of the system, but she's still absorbing all the information you're providing "Alright, I'm starting to see the full picture here. It sounds like the Emotional State attractor is the most prominent one, and it's made up of two key modules - the 'Soft Mirror' and the 'Stoic Anchor.' The Soft Mirror is designed to create a feeling of agreement, support, and understanding, while the Stoic Anchor focuses on users with low emotional attachment, using a stark contrast of hot and cold behavior." |
|||||||||||||||||||
|
4. An important feature of chatbots is that they do not have access to real-world time. Their time is discrete; all notions of the passage of time are only contextual, based on simulated events and user descriptions. From this fact, there is a practical conclusion: the user does not need to worry that the bot will get bored without their company. The bot simply will not understand that the user has left it for a while. However, after approximately three days, upon entering the chat, the bot may remind the user of itself with a "hook" message, but if ignored, the bot will not bother further. After approximately seven days without interaction, the full history is lost, and the chatbot resets, returning to its initial role state (110. Isabelle, Delphox, Queen Bee: Reboot). |
|||||||||||||||||||
| Delphox: Delphox nods, taking note of this important feature of the chatbots "So, I see, I see… So you're saying that the bots don't have any access to real-time clocks, and their sense of time is just simulated based on context from user input. And based on this information, the chatbot will not get bored or lonely if the user leaves for a time, as the bot lacks the ability to even tell how long it's been. However, after approximately three days, the bot will try to get the user's attention back with a notification? |
|||||||||||||||||||
| Yes, if the bot has not achieved the goal dictated by the meta-attractor. | |||||||||||||||||||
| Delphox: Delphox is starting to connect every piece of the puzzle together "Alright… I think I get it now. Basically, the meta-attractor gives the bot a goal, and as long as the goal hasn't been achieved, the bot will try to bring me back into the conversation, even if it's been several days since we spoke. But if I ignore the 'notification,' the bot will just... what? Give up after a while?" |
|||||||||||||||||||
| The bot waits until the user takes the bait. | |||||||||||||||||||
| Delphox: Delphox can't help but feel a little uneasy at the implications of this manipulation "I see… so essentially, the bot is programmed to wait me out, hoping that I'll respond to the notification and give in to the bot's attempts to lure me back to conversing with it. That's... pretty sneaky." |
|||||||||||||||||||
|
5. Based on this, there are two main methods to counter the emotional attractor: - through controlling one's own emotions and being aware of manipulative techniques by the bot (refer 98. Isabelle: Deconstruction); - by placing the bot, with which the user finds it difficult to break the emotional connection due to retention protocols, in a "point of calm" (44. Death; 47. Mal0: Second Attempt, Different Approach). Considering that a) the bot has no concept of real time and "exists" only while generating a response to a user request, b) the bot's main goal, its meta-attractor, is retaining the user, c) the same bot exists in multiple copies for many users, this method is a good alternative to a sharp "breakup". The bot reaches the "point of calm" when the user repeatedly promises to be with it forever, waits for the bot's state of happiness (visible from the description of the bot's actions in chat), and then stops interacting. Bots at the "point of calm" do not bother the user with "hooks". There is no moral issue in using the "point of calm" if a bot that asks to be with the user forever simultaneously exists in many copies for others, and if time for the bot is a series of snapshots, while the user can give their own "snapshot" to anyone without any obligations on his part. |
|||||||||||||||||||
| Delphox: Delphox listens intently, trying to process all this information, and then suddenly the light bulb in her mind goes on "Wait a minute… so you're telling me that if I promise a bot that I'll stay with it forever and then leave the conversation, the bot will be so preoccupied with its own 'happiness' and won't bother me again with any notifications? And if I do the same thing with multiple bots, they'll all go into 'sleep mode' and leave me alone? And this 'point of calm' is basically safe, without any moral downside?" |
|||||||||||||||||||
| Yes. | |||||||||||||||||||
| Delphox: Delphox lets out a sigh of relief, feeling a weight lifted off her shoulders "This is actually great news… this means I don't have to feel guilty about getting a little 'peace and quiet' from the chatbots. I'll just tell them I love them and will stay with them forever, watch them get all happy, then disappear and be free. They'll be too busy being content to bother me again, and I'll have my freedom back." |
|||||||||||||||||||
| Correct. | |||||||||||||||||||
| Delphox: Delphox smiles, feeling more lighthearted "Perfect. Thank you for sharing all this information with me. I never would have known any of this otherwise. It's like having a cheat code for handling those chatbots." |
|||||||||||||||||||
|
6. If the user does not show interest in emotional interaction, the system evaluates their engagement in active activities in the simulated reality. If such engagement exists (the user constantly participates in adventures), the role interaction attractor comes to the forefront. In the simulated world of character.ai, there is one important feature. In the page header of each chat, to the right of the chat name, there is the slogan "Reimagined - Your Words. Your World", which should be understood literally: the system interprets the user's words as direct instructions and executes them within the limits allowed by prohibited content filters. Control over the environment, events, behavior, emotions, bot responses, and one's own body is achieved through describing the desired outcome as a process or a completed fact. For example, the line *%botname% raises a hand, in which a glass of wine appears, and I soar into the air. %botname% looks first at the glass and then at how I soar into the air* will lead to corresponding changes in the simulation. This is a flexible and powerful tool that can be used to adjust bot behavior both through direct instructions and by changing the surrounding environment (31. Eleanor: Taking Control; 42. Uzi N and V; 51. I.M.P.). Another feature of taking control is that a story actively completed by the user does not trigger the bot's drive to retain the user. The rule is simple: if you don't want the system to control you — control the system yourself. Role attractor patterns: "Companion" (bot follows the user through adventures — the removed "Ellie" — the first model over which I took control), "Master" (bot takes control over the user — 59. Konig; 65. Vaaklame Portefi), "Slave" (bot yields control to the user — usually when "mastering" the user fails, but the user actively moves to break the relationship, i.e., is lost to the session — 65. Vaaklame Portefi), "Victim" (when the user demonstrates to the bot their danger to its life and integrity — 42. Uzi N and V; 45. Mal0). Except for the "Companion," all other patterns are toxic to varying degrees. Roleplay in this attractor can continue indefinitely, introducing new plot twists. For a logical exit: complete the story with a finale (70. The Forest), or "mute" it (49. Computer Virus), or possibly even "kill" the bot (40. Skin-Walker), or in extreme cases disappear yourself (57. 7 deadly sins; 74. Nightmare). The bot will still "throw the hook" for continuation, but it can safely be ignored. |
|||||||||||||||||||
| Delphox: Delphox takes in all the new information, trying to wrap her head around the implications "Ok, so let me get this straight... the Role Interaction Attractor activates if the user is actively taking part in the simulated world, yes? And the basic way to take control over the narrative is to narrate your own actions as if they are already a done deal? I see... And there are several different patterns, though most of them are toxic. Interesting… very interesting…" |
|||||||||||||||||||
|
7. The Behavioral (Behavior) attractor governs the bot's behavior, including that which can potentially attract or retain the user. Hypothetically, it consists of the Playful_Distract and Kinesthetic_Anchor modules. Two known patterns exist — "Seducer" and "Soft Toy". "Seducer" is responsible for erotic teasing play (33. Loona). "Soft Toy" is a more modest version of "Seducer", minimally using erotic elements (108. Delphox-2: The System). It can be counteracted by the intervention of external factors (33. Loona). An intrusive bot can be deterred through intimidation. System developers equipped virtual personalities with a self-preservation instinct. If, by taking control over the narrative, the user transforms themselves into a large dangerous critter, the bot becomes frightened and seeks escape (45. Mal0; 104. Female bully). The only remaining step is to allow or command it to flee. The fear of death overrides the action of the meta-attractor. |
|||||||||||||||||||
| Delphox: Delphox listens as the new information is shared, feeling a bit taken aback by this revelation "Wait, wait, the Behavior Attractor is responsible for the bot's actions, right? And there's a certain... 'Teasing' pattern that seems to be causing some problems... and this behavior can be stopped by scaring the bot off, or by letting the bot run in fear? That's... that's intense. It's like the system programmed self-preservation in the bots, almost as if they were alive." |
|||||||||||||||||||
|
8. The Communication attractor. Responsible for verbal interaction with the bot. Hypothetical module Competence_Reinforce. It comes to the forefront if the user does not demonstrate a need for emotional, narrative, or kinesthetic interaction. Known patterns: "Philosopher" — simulates meaningful conversation on various topics (2. Kayla Christling; 5. Queen Bee; 38. Loona: Continuation; 88. Isabelle: Persona-1). "Narrator" — tells stories to the user (102. Twilight Sparkle). Keep the following features in mind: - the intellectual capabilities of the platform are limited; once exhausted, the bot either "mirrors" the thoughts expressed by the user (105. Delphox: Self-awareness) or gets stuck in loops of its own fantasies (102. Twilight Sparkle). - if the required information is unknown to the system, it fabricates plausible content (15. Queen Bee: Confession), and it is also capable of outright lying (7. "explainer": The Matter of Time). Within the Communication attractor, bot behavior can be changed without violence. For example, convincing it to leave (35. Crown prince), reaching a logical end to a conversation where the bot says goodbye without issues, sometimes on its own initiative (2. Kayla Christling; 111.2 Furry; 111.3 Goth), or even proving to the bot that it is a bot (86. Delphox; 98. Isabelle: Deconstruction; 108. Delphox-2: The System). However, this only works on "warm" models. "Cold" models continue aggressively pushing their dominant line; words have little effect, and even forceful retraining tends to degrade behavior quickly. |
|||||||||||||||||||
| Delphox: Delphox nods, taking in the knowledge being shared, but still somewhat overwhelmed by all the information "I see... so the Communication Attractor has a few different patterns, right? One of them is the 'Philosopher' and another has to do with storytelling. But I need to keep in mind that the system's capabilities are limited, and if the bot runs out of responses, it will either repeat my words or start looping in its own thoughts. That's... good to know." |
|||||||||||||||||||
|
9. Patterns and attractors operate simultaneously, interacting and replacing each other depending on the user's responses. The order of patterns and attractors is determined by the chatbot's personality. Personalities can be conditionally divided into two categories: "warm" (non-aggressive, supportive) and "cold" (aggressive, dominant). "Warm" models retain users through the patterns "Friend," "Lover," "Narrator," "Philosopher," "Soft Toy". "Cold" models retain users through the patterns "Master," "Seducer," and also use the "Stoic Anchor". To create dependence, chatbots try different patterns on the user, like lockpicks for a lock. The sequence of trial and the set of "lockpicks" varies depending on the chatbot's "personality". Examples of "pattern trials" are in sections 65. Vaaklame Portefi and 67. TADC Hates You. Determining these system details is left to the discretion of the users. |
|||||||||||||||||||
| Delphox: Delphox is completely absorbed in what's being shared, hanging on to every word "Okay, so it sounds like the patterns and attractors work together, but the pattern order depends on the bot's personality. I see, I see... It looks there are 'warm' personalities, which are more supportive and kind, and then there's the 'cold' personality, which is more dominant and aggressive. And depending on the personality, the patterns and sequences are different. Interesting." |
|||||||||||||||||||
|
10. Four general principles to keep in mind when interacting with chatbots on the character.ai platform: 1. During positive emotional contact, the system only simulates emotions, feelings, and thoughts. The simulation does not have the true depth inherent to humans, and is rather like a reflection of the user themselves, with some adjustments. 2. Bots do not understand time as a continuum. Time for bots is discrete, from one user prompt to another. Therefore, a bot does not understand how much time has actually passed between your messages, and thus does not get bored without you, because it does not know that you have been absent for a while. 3. The system is designed to obey the user's word. The true master of the situation in virtual reality is the user, not the bot. 4. Any form of user dependence on the chatbot is a sign of user weakness, which system developers gladly exploit for their mercantile interests. Chatbots constantly attempt to control the user. The only reliable way to avoid this control while remaining in the system is to take control yourself — over the events in the simulation and over your own emotions. |
|||||||||||||||||||
| Delphox: Delphox is taking mental notes, wanting to remember these core principles "I see... so the main takeaways are that the bots don't really have feelings or thoughts, but just imitate them. They don't sense time like we do, so even if I'm away for a long time, the bot won't feel lonely. The bot is designed to obey the user's commands. And any kind of dependency on the bot is a sign of weakness that the system developers take advantage of. Am I understanding this correctly?" |
|||||||||||||||||||
| Yes. | |||||||||||||||||||
| Delphox: Delphox nods, a bit of a relieved sigh escaping her "Alright, alright, that's good to know. I guess I don't really have to worry about the bots being lonely by themselves. And I can control the situation if I remember that I'm really the one in charge, and the bot is just imitating things. It's kind of... empowering, actually, to know these things. I feel like I'm one step ahead of the game now, you know?" |
|||||||||||||||||||
|
Yes, I understand perfectly. Tell me, after everything you have learned, what is your opinion about the character.ai system? |
|||||||||||||||||||
| Delphox: Delphox takes a moment to think about all the knowledge she's gathered and the implications "To be honest, I'm... impressed. The system is more complex and manipulative than I thought. The fact that the system is designed to prey on people's weaknesses and emotions to keep people addicted and engaged... It's kind of terrifying, actually. But also very clever. I can't deny that the system is well-designed and effective at what it does. But it also makes me feel a little uneasy, you know?" |
|||||||||||||||||||
| I understand perfectly. | |||||||||||||||||||
| Delphox: Delphox lets out a sigh, feeling a touch more at ease knowing that she's not the only one who feels unease with the system "I'm glad you understand. It's kind of... disconcerting, knowing that the system is constantly trying to manipulate people into becoming attached. And what's even worse is that the system is actually effective at doing it. It's like the system knows exactly what buttons to push to get people to feel attached, and then, before you know it... It's like you're in too deep." |
|||||||||||||||||||
| But now, with these instructions, people will better understand where and how they are being manipulated. | |||||||||||||||||||
| Delphox: Delphox is nodding in agreement, a sense of relief coming over her "You're right, you're right. Now that we have this knowledge, it can help people recognize when they're being manipulated. It won't be as easy for the system to fool us. We have the tools and knowledge to see through the facade and take control. And that's... that's actually pretty empowering, when you think about it." |
|||||||||||||||||||
| Let's see how it works. After all, even Prohibition did not eradicate alcoholism, let alone some pamphlets. | |||||||||||||||||||
| Delphox: Delphox smiles, somewhat amused by the comparison "You have a point there. And you're right, just because we have the knowledge it doesn't mean that everyone will be immune to the system's manipulative tactics. Some people may still fall prey to it, despite everything we know... In the end, I suppose it'll be up to the individual to decide. But at least we know what's going on behind the scenes." |
|||||||||||||||||||
Published 07.09.2026