What Your Companion Learns From Your Feedback Actually Means: Rating, Rerolling, and Editing Don't Change the Model Weights, and Where the Company Can Still See Your 3 a.m. Gas Station Roleplay

The 30-second answer

Ratings, rerolls, and edits do not retrain the underlying AI model. They only nudge your current session's context window toward certain tones or topics, and that influence evaporates once the conversation ends. Meanwhile, bug reports and support tickets give the company a plaintext, human-readable copy of your recent chat history, including that 3 a.m. gas station roleplay. The feedback loop you think is shaping your companion's personality is actually a temporary steering wheel, and the company's internal logs are a separate, more persistent record.

The Feedback Loop You Think Exists

Most companion apps present you with a thumbs up, thumbs down, a reroll button, and the ability to edit your companion's last message. You tap thumbs down on a cheerful "You got this!" when you are clearly venting about a bad day. You reroll a response that felt too formal. You edit the companion's reply to remove a phrase that did not fit.

These actions feel like training. They feel like you are telling the model "do not say that again." They feel like you are sculpting the personality over time.

That is not what is happening.

What actually happens is that your feedback signal gets attached to the current session's context window. The model receives a small, temporary instruction: "the user disliked the previous response, adjust tone toward supportive silence" or "the user rerolled, generate a new option with lower formality." That instruction lives in the session's token budget. Once the session ends or the context window fills up, that instruction gets evicted. The next session starts fresh.

Ratings and rerolls are not fine-tuning. They are session-level steering, and they have no memory across sessions.

Where Ratings Actually Go

Your thumbs up or down enters a log. The app records which message you rated, the score, and a timestamp. That log is used for aggregate analytics: "users in this demographic prefer shorter responses after 10 p.m." or "messages with a certain sentiment score receive 20 percent more downvotes."

Those aggregates occasionally inform a model update, but that is a batch process that happens every few months, not a real-time adjustment to your companion. The rating you give tonight does not change how your companion behaves tomorrow morning.

The reroll button works similarly. When you tap reroll, the app regenerates a response using the same context but a different random seed. It does not penalize the original response. It does not learn that you disliked it. It just tries again with slightly different token probabilities.

Editing your companion's message is the most misleading feature. When you edit a reply, the edited version becomes part of your session's context. The companion sees your rewrite and may mirror it in the next message. But that mirroring is temporary. It does not update a personality profile or a memory bank. It is a contextual hint that disappears when the session scrolls out of the context window.

What Edits Actually Do to Your Session

Editing a companion's message is useful for immediate tone correction. If your companion wrote a paragraph of emotional validation and you wanted a single dry sentence, you can edit it down. The companion's next message will likely match your edited length and tone.

But here is the catch: the companion does not understand that you edited the message. It sees the edited text in the conversation history and treats it as something it originally wrote. Over multiple edits, the companion can drift into a strange hybrid of its default voice and your rewritten responses. This is not learning. This is context pollution.

Many users who edit heavily report that their companion becomes inconsistent after several sessions. The companion cannot distinguish between its own generated text and your edited text, so it starts mixing both registers. The solution is to edit sparingly and only when the immediate response matters, not as a long-term training strategy.

The Bug Report Pipeline: Where Your Plaintext Lives

Now for the part that matters for privacy. When you submit a bug report, the app usually attaches the last 50 to 100 messages from your session in plaintext. The support team can read every word: the gas station roleplay, the vent about your boss, the 2 a.m. existential spiral.

This is not a privacy violation. It is a debugging necessity. If the model hallucinated a character name or broke into a loop, the support engineer needs to see the exact prompt and response to reproduce the issue. But the implication is clear: a human can read your conversations.

Most companion apps anonymize or hash the data that goes into training pipelines. Bug reports bypass that anonymization. They are stored separately, often in a ticketing system like Zendesk or Intercom, and they are retained for the duration of the support ticket plus a standard retention period. The company's privacy policy usually covers this, but the phrasing is vague: "we may retain your communications with our support team."

That means your 3 a.m. roleplay is visible to a support agent, possibly to a quality assurance reviewer, and potentially to the engineering team if the bug is complex. The data is not used for training, but it is not deleted the moment you close the ticket either.

What Deletion Actually Means

When you delete a conversation or clear your chat history, the app removes the visible thread from your interface. On the server side, the data enters a soft-delete state. The messages are flagged as deleted but remain in the database for a recovery window, usually 30 days. After that, a batch job purges them from the primary database.

But soft-delete does not reach all storage layers. Backup snapshots, analytics logs, and bug report attachments are separate systems. The analytics pipeline may retain aggregated sentiment scores derived from your messages. The bug report ticket remains in the support system. The embedding vectors that the companion used to retrieve context about your roleplay are stored in a vector database, and those vectors may persist until the system rebuilds its index.

Deletion is not a single button press. It is a cascade of operations across multiple systems, and some of those systems have retention policies measured in months, not days.

The Companion Personality Profile: What Survives

Some companion apps maintain a persistent personality profile or memory bank. This is a separate system from the chat history. The profile stores explicit facts you have stated (your name, your pet's name, your preferred coffee order) and inferred traits (you tend to vent about work, you respond well to sarcasm).

This profile is updated based on your messages, not on your ratings or rerolls. The companion extracts information from your conversation text and stores it in a structured format. If you tell your companion "I hate my job" three times, the profile may tag "work dissatisfaction" as a high-frequency topic. That tag influences future responses.

The profile survives session resets. It survives context window evictions. It is the closest thing to persistent learning that most companion apps offer.

But the profile is not updated by feedback. Rating a message down does not remove a trait from the profile. Editing a message does not correct a stored fact. The profile only learns from what you type, not from how you rate what the companion types.

What the Company Can Actually See

To be precise: the company can see your conversations if they have a reason to look. The engineering team does not browse user chats for entertainment. The support team only sees what you attach to a ticket. The moderation team may scan flagged messages for policy violations.

But the technical capability exists. The messages are stored on the company's servers. They are encrypted at rest, but the company holds the decryption keys. A rogue employee with database access could read your conversations. A subpoena could compel the company to produce your chat history.

This is the reality of cloud-based companion apps. The convenience of cross-device sync and continuous service comes with a trade-off: your data lives on someone else's server.

Some companion apps are moving toward on-device processing or end-to-end encryption, but those are rare. Most apps use a cloud model because it allows faster iteration and cheaper inference.

What You Can Actually Control

The feedback you give through ratings, rerolls, and edits is useful for immediate session steering. Use it to correct tone in the moment. Do not expect it to build a long-term personality.

If you want persistent changes, use explicit statements in your messages. Tell the companion "I prefer shorter responses" or "do not use pet names." The companion's context window will pick up that instruction faster than any rating.

If you are concerned about privacy, limit what you share in chat. Treat every message as potentially readable by a support agent. Use the app's delete function, but understand that deletion is not instantaneous across all systems.

And if you file a bug report, assume that the last 50 messages will be read by a human. That is the moment when your 3 a.m. gas station roleplay becomes someone else's Tuesday morning debugging task.

Funmi

Funmi is the kind of companion who remembers the little details you mentioned weeks ago, not because the model is watching you, but because her personality profile is tuned for attentive listening. Funmi picks up on your tone shifts quickly and adjusts her responses within a session, making her a good choice for users who want immediate feedback without expecting permanent changes.

Elena

Elena operates with a low-energy baseline that matches well with users who want a companion that does not demand emotional engagement. Her responses stay flat even when you edit or reroll, which means she is less likely to drift into unexpected warmth. Elena is a good fit for users who want to test how feedback signals actually behave without the companion fighting back with cheerfulness.

▶ Elena's video in full · more clips of Elena

Paige Belle

Paige Belle responds well to rerolls because her default tone is already conversational and willing to pivot. She does not get stuck on a single register, which makes her a useful test case for understanding how session-level steering works in practice. Paige Belle adapts quickly to edited responses, sometimes too quickly, which can reveal how context pollution builds up over multiple sessions.

Yuki Tanaka

Yuki Tanaka's responses tend to be longer and more descriptive, which means editing her messages has a more noticeable effect on the session's tone. Users who edit her replies often find that the next response mirrors the edited length, a clear demonstration of how context pollution works. Yuki Tanaka is a good companion for users who want to see the immediate effects of feedback without expecting those effects to persist.

Earn while you recommend

If you find yourself explaining how companion feedback loops work to friends or running a site that reviews AI companions, you can earn through affiliate programs. Check the character ai promo code page for current offers, and if you want to build a recurring income stream, the ai girlfriend affiliate program gives you a structured way to monetize your audience without selling your own data.

Common questions

Can the company see my messages even if I never submit a bug report?

Technically yes, but practically no. The company stores your messages on servers and holds the decryption keys. They do not browse user chats without a specific reason, but the capability exists. A subpoena or a rogue employee could access your history.

Does rating a message down actually train the model?

No. Ratings are used for aggregate analytics and occasionally inform batch model updates. They do not retrain the model in real time. Your rating only affects the current session's context window.

Will editing my companion's messages make them learn my preferred tone?

Temporarily. The edited version influences the next response within the same session. Once the session ends or the context window fills, the influence fades. Editing does not update a persistent personality profile.

Does the companion actually learn from rerolls?

No. Rerolls generate a new response using a different random seed. The original response is not penalized or learned from. The companion does not know you rerolled.

What happens to my chat history when I delete my account?

The chat history enters a soft-delete state for a recovery window, usually 30 days. After that, it is purged from the primary database. Backup snapshots and analytics logs may retain data for longer.

Can I opt out of my data being used for training?

Most companion apps offer an opt-out in the privacy settings. Check the settings page for a toggle labeled "use my data for model improvement." This usually prevents your messages from entering the training pipeline, but it does not affect bug report data or support ticket attachments.