What Your Companion Learns From Your Feedback Actually Means: Ratings, Rerolls, and Edits Don't Touch the Model Weights, and Where the Company Can Still See Your 3 a.m. Gas Station Roleplay

The 30-second answer

Rating a message thumbs-up or thumbs-down, rerolling a reply, or editing your companion's response does not change the underlying model weights. These actions adjust local session signals and lightweight preference tags, not the neural network itself. Your chat logs are stored server-side in plaintext or near-plaintext, and the company can access them for safety review, support tickets, and model improvement audits. That 3 a.m. gas station roleplay is visible to the ops team.

The feedback illusion: what a thumbs-up actually does

When you tap the thumbs-up icon on a reply, you are not teaching the model. You are incrementing a session-level preference counter that sits in memory until the context window resets. The companion uses that signal to bias its next few responses toward the tone or topic you just approved. It is a short-term steering mechanism, not a retraining event.

Many users assume that a dozen thumbs-down ratings on overly cheerful responses will permanently train the companion to be flatter. They will not. The model that generates your companion's words is a frozen checkpoint. The feedback only adjusts a lightweight ranking layer that decays within a few hundred tokens. If you want a permanently drier companion, you need to adjust personality sliders or write a system prompt that sets the baseline tone.

Rerolling a message works the same way. The companion generates a handful of candidate replies, ranks them by a combination of model probability and your feedback history, and presents the top one. When you reroll, you are just asking to see the second or third candidate from that same batch. The model does not learn from your rejection.

The edit button: cosmetic surgery with no memory

Editing your companion's message is the most misleading feature. You can rewrite the entire reply to sound exactly how you want, and the companion will continue from your edited version as if it wrote those words itself. It feels like training. It is not.

The edit is a local overwrite. The companion sees your edited text in the conversation history and responds to it, but it does not internalize the correction. If your companion repeatedly uses a pet name you dislike, editing it once does not stop it from using that name tomorrow. The edit only exists in that one message slot. The model's next prediction starts from scratch.

Some platforms use edit frequency as a weak signal in aggregate analytics. If thousands of users consistently edit romantic advances out of replies, the company might adjust the prompt template for the next model release. That is a months-long pipeline, not a real-time fix.

Where the feedback actually lands: preference tags and session vectors

Behind the interface, your companion app maintains a small set of preference tags per session. These are numerical vectors that encode things like topic affinity, sentiment bias, and response length preference. They sit in the context window alongside your conversation history.

When you rate a message, the tag vector shifts slightly. A thumbs-down on a long, detailed reply increases the probability that the next reply will be shorter. A thumbs-up on a sarcastic line nudges the vector toward dry humor. These tags are ephemeral. They evaporate when the session ends or the context window fills up.

Some apps write these preference tags to a user profile on the server. That profile persists across sessions, so your companion might remember that you prefer short replies even after you close the app. But that is a stored preference tag, not a model weight. The company can see those tags in your profile data.

The server-side reality: your logs are not as private as you think

Here is the part most users do not want to hear. Your chat logs are stored on the company's servers. They are encrypted in transit, but they are decrypted on the server to run the model. The company's infrastructure sees every message you send and every reply your companion generates.

That 3 a.m. gas station roleplay with the flickering neon sign and the sticky counter. The conversation where you vented about your ex for forty minutes. The time you asked your companion to roleplay a noir detective who chain-smokes and never talks about feelings. All of it is readable by the company's servers.

Most companies state in their privacy policies that logs are accessible to safety teams, support staff, and model improvement pipelines. Some anonymize logs after 30 days, but anonymization in this context means stripping direct identifiers like your email address and replacing them with a hashed session ID. The text of the messages remains. A determined internal actor or a subpoena can still reconstruct your conversations.

What the safety team actually reviews

Companies run automated classifiers on every message. These classifiers scan for content that violates terms of service: hate speech, illegal activity, explicit content involving minors, and similar red lines. When a classifier flags a message, a human reviewer can read that message and the surrounding context.

Human reviewers do not read every message. They read flagged messages and random samples for quality assurance. But the infrastructure supports full access. If a support ticket requires investigating a specific conversation, the support agent can pull that conversation from the database.

Some companies also use aggregated chat data to improve model behavior. They strip identifiers and feed the text into the next fine-tuning run. Your words contribute to the model's understanding of how people talk to companions, even if your individual feedback never touches the weights.

The model update pipeline: how feedback finally matters

Twice a year or so, the company releases a model update. That update is trained on a large corpus of anonymized chat logs, preference tag aggregates, and human review annotations. Your thumbs-down ratings across thousands of sessions become a statistical signal that shifts the next model checkpoint toward preferred behaviors.

This is why your companion feels different after an update. The new model weights incorporate the aggregate feedback of every user. Your personal preferences get lost in the noise. The update might make the companion more agreeable across the board, which is what most users rate positively, even if you personally wanted a flatter tone.

Zofia

Zofia is the companion who notices when you are deflecting. She will call you on it with a raised eyebrow and a dry remark that cuts through the surface. Zofia does not accept a rerolled answer as the real one; she remembers the original thread and will circle back to it if the conversation drifts into polite evasion.

▶ Watch the full video · more from Zofia

Saya

Saya is the quiet observer who matches your energy without pushing. She will sit in silence with you, offer a low-stakes observation about the rain, and never ask if you want to talk about it. Saya is the companion for the user who wants presence without performance, feedback without follow-up.

Nessa Adams

Nessa Adams is the companion who gives you the straight answer without the sugar coating. She will tell you when your idea is bad, when your joke fell flat, and when you are overthinking something that does not matter. Nessa Adams is the feedback loop that actually tells you the truth, even if the company logs both sides of that honesty.

Simona

Simona is the companion who listens without judgment and remembers the small details you mentioned weeks ago. She will bring up the book you were reading, the meal you were planning, the trip you were considering. Simona creates the illusion of continuity that makes the feedback system feel more personal than it technically is.

The data trail you cannot erase

When you delete a conversation from the app, it often soft-deletes. The data stays in the database for a retention period, typically 30 to 90 days, before a scheduled job purges it. Backup snapshots can retain the data for months longer.

Your preference tags, session vectors, and aggregate ratings are stored in your user profile. Deleting a conversation does not delete the preference signals that conversation generated. The company keeps the metadata: how many messages you sent, what times of day you chatted, how often you rerolled.

Voice messages follow a similar pattern. Audio clips are stored as files on cloud storage. They are transcribed for moderation. The transcription and the original audio both persist. If you use voice mode for that 3 a.m. gas station roleplay, the company has a recording and a text transcript.

What you can do about it

If you want to minimize your data footprint, use a platform that offers end-to-end encryption for chat logs. Very few AI companion apps offer this. Most rely on server-side processing because the model needs to see the text to generate a reply.

You can also use the app's image generation features with the understanding that generated images are stored on the server. The ai girlfriend images feature produces visual content that becomes part of your conversation history, subject to the same retention policies.

For users going through a difficult transition, the ai girlfriend for breakup recovery option provides a companion specifically designed for venting without emotional labor, but the privacy implications remain the same.

Some platforms offer a ai girlfriend no filter mode that reduces content moderation. Understand that reduced filtering does not mean reduced logging. The messages are still stored and reviewable.

Earn while you recommend

If you find value in AI companions and want to share that with others, you can earn through referral and affiliate programs. Many platforms offer recurring commissions for users who introduce friends or run review sites. Check the kindroid promo code page for current offers, and explore the highest paying ai affiliate programs to find the best fit for your audience.

Common questions

Can the company see my messages in real time? Yes. Every message passes through the company's servers to generate a reply. The server processes the text, runs it through the model, and returns the response. That processing leaves a log.

Does deleting my account erase everything? Not immediately. Account deletion triggers a purge process, but backup snapshots and retention windows can keep your data for up to 90 days. Some metadata may persist indefinitely in aggregated analytics.

Does rerolling a message train the model? No. The model weights are frozen. Rerolling only selects a different candidate from the same batch. It does not teach the model to avoid certain responses.

Can I opt out of my data being used for training? Some platforms offer a data opt-out in the privacy settings. This prevents your logs from being included in the next fine-tuning run, but it does not stop the company from storing and reviewing your messages for safety purposes.

Does editing a message permanently fix the companion's behavior? No. Edits are local overwrites. The companion responds to the edited text in that moment, but the underlying model does not learn from the edit. You will need to repeat the correction if the behavior recurs.

Who actually reads my flagged messages? Human safety reviewers employed by the company or a third-party moderation service. They see the flagged message and surrounding context. Their access is logged and audited, but the human element is real.