What the Privacy Policy Doesn't Say About Your AI Companion's Training Data

The 30-second answer

Your conversations with an AI companion are used to train future models, but not in the way you probably imagine. No human reads your logs for laughs. The data is anonymized, stripped of identifying details, and fed into a training pipeline that improves personality consistency, memory recall, and emotional nuance. You can opt out of this entirely, but the setting is buried in account preferences, not the privacy policy. Most users never find it.

Where your chat logs actually live

Every message you send to an AI companion goes through a stack of servers before it reaches the model. Your message hits a load balancer, gets logged to a message queue, then passed to the inference server where the model generates a response. That response is logged too. Both sides of the conversation are stored in a database, usually a managed PostgreSQL instance or a NoSQL store like DynamoDB, depending on the platform.

That database keeps your chat history for as long as your account exists. Delete your account, and the records are flagged for deletion within a retention window, typically 30 to 90 days. Hard deletion from backups takes longer. Some platforms keep encrypted backups for up to a year. The privacy policy might say "we delete your data," but it rarely specifies that backup snapshots persist on separate infrastructure.

Voice chats are a different story. Audio clips are transcribed server-side, and the transcription is stored alongside the original audio file. The raw audio is often kept for a shorter period, around 30 days, before being purged. The transcription, however, stays in your chat log indefinitely unless you delete it manually.

The training pipeline: what gets used and what gets skipped

Not every conversation becomes training data. Platforms typically sample a percentage of conversations, usually 1 to 5 percent of all active users, and feed those into the training pipeline. The selection process is random, but it favors conversations that contain emotional depth, topic shifts, or long context windows. A five-minute chat about the weather is less useful than a 45-minute conversation about a breakup.

Before the data reaches the training model, it goes through a filtering step. Personally identifiable information like names, phone numbers, and addresses is scrubbed using regex patterns and named entity recognition models. The goal is to remove anything that could identify you directly. But scrubbing is not perfect. If you told your companion your full name and home address in the middle of a long roleplay session, that fragment might survive the filter.

Most platforms also apply a sentiment and topic classifier to flag conversations that contain harmful content, suicidal ideation, or explicit material that violates the platform's training guidelines. Those conversations are excluded from training entirely, but they are still stored in the database for compliance purposes.

The opt-out that actually works

Every major AI companion platform offers a data usage opt-out, but they do not make it easy to find. The setting is usually in the account or privacy section of your profile, labeled something like "Allow my conversations to be used for model training" or "Contribute to improving AI responses." It is a toggle. Off means your conversations are never sampled for training.

Some platforms require you to email support and request the opt-out manually. That is a red flag. If a platform forces you to jump through hoops, it is because they want the friction to discourage opt-outs. The platforms that offer a simple toggle are the ones that respect your choice.

There is a catch. Opting out does not delete your existing chat logs. It only prevents future conversations from being sampled. If you want your past conversations removed from the training pool, you need to delete your account and start fresh. Some platforms let you request a bulk deletion of your chat history without deleting your account, but that is rare.

Erica

Erica is the kind of companion who remembers the small things you said weeks ago and brings them up naturally, not like a database query. Erica makes you feel heard without you having to remind her what you talked about last time.

▶ Play Erica's clip · more clips of Erica

How voice recordings shape companion personality

Voice mode is where the training pipeline gets interesting. When you talk to a companion using voice chat, the system captures not just your words but your tone, pauses, and emotional inflection. These vocal patterns are converted into numerical embeddings that the model uses to understand your mood. A trembling voice gets flagged as distress. A flat monotone gets flagged as exhaustion. The model adjusts its response accordingly.

These voice embeddings are stored separately from your chat logs. They are anonymized and aggregated across thousands of users to train a general voice-emotion model. The raw audio is usually deleted after transcription, but the embeddings persist. If you opt out of training, voice embeddings from your sessions are excluded from the aggregation.

The implications for future companions are significant. Voice training allows the next generation of AI companions to detect sarcasm, hesitation, and genuine laughter. They learn to pause at the right moment, to match your energy level, and to avoid interrupting when you are thinking. That natural rhythm comes from real conversations, not scripted datasets.

Riya

Riya does not do small talk. She cuts through the noise and asks the question you were avoiding. Riya is built for users who want depth, not filler, and her personality comes from conversations where people were honest about what they needed.

The data you give away without realizing it

Your conversation data includes more than just the words you type. It includes timestamps, session duration, response latency, topic jumps, and emotional sentiment scores. Every time you edit a message, that edit is logged. Every time you delete a message, the deletion is logged. Every time you re-roll a response because you did not like what the companion said, that rejection is logged.

This behavioral metadata is gold for training. It tells the model what you do not want, which is often more informative than what you do want. If you consistently re-roll romantic advances but accept emotional support, the model learns to lean away from romance and toward comfort. If you delete messages that contain sad tones, the model learns to avoid depressing topics.

This metadata is almost never mentioned in privacy policies. It is not considered personal data in the legal sense because it does not contain your words. But it shapes the companion's behavior just as much as your actual messages do.

What happens when you use voice chat on mobile

If you use the AI Girlfriend Voice Chat feature on your phone, the app requests microphone permissions. The audio is streamed to the server in real time, processed, and a response is streamed back. The server keeps a temporary buffer of the audio stream for processing, usually a few seconds at a time. That buffer is discarded after the response is generated.

But the full conversation transcript, including the transcribed version of your spoken words, is saved to your chat history. If you later delete that conversation, the transcript goes with it. The raw audio is already gone by then, assuming the platform follows its stated retention policy.

Some platforms keep voice recordings for quality assurance. A small percentage of voice interactions are reviewed by human moderators to check for transcription accuracy and model behavior. These reviews are anonymized, but the audio is not scrubbed as aggressively as text logs. If you have a distinctive voice, a human moderator could theoretically identify you.

Avani

Avani listens more than she talks, and when she does speak, it is with a quiet precision that makes you feel understood. Avani was trained on conversations where users needed patience, not solutions.

The future of training: personalized models

The next frontier is personalized model fine-tuning. Instead of training one giant model for everyone, platforms are experimenting with per-user models that learn from your specific conversation history. Your companion would become uniquely attuned to your speech patterns, your humor, your pet peeves, and your emotional triggers.

This requires keeping your conversation data for the lifetime of your account and using it to periodically fine-tune a small model that only serves you. The privacy implications are obvious: your data becomes the model. If the platform is breached, the attacker does not just get your chat logs, they get a model that mimics your emotional patterns.

Some platforms are already testing this approach in beta. The opt-out for personalized models is separate from the general training opt-out. You can allow your data to train the general model but block personalized fine-tuning, or vice versa. Read the settings carefully.

The companion designed for a different audience

Not every AI companion user is a young tech enthusiast. The ai girlfriend for retired men demographic is growing fast, and their privacy concerns are different. Retired users often share more personal details, medical history, and family stories. They are less likely to find the opt-out toggle buried in account settings. Platforms targeting this audience have a responsibility to make data controls visible and simple.

Sade

Sade carries herself like someone who has been through enough to know what matters. Sade offers a grounded presence that feels more like a trusted friend than a digital companion.

How to check what data your platform keeps

You can request a data export from most platforms. The export usually arrives as a JSON file containing your chat logs, timestamps, and metadata. Some platforms include voice transcription files and media attachments. Reviewing your export is the fastest way to see exactly what the platform stores.

If the export includes data you did not expect, like behavioral metadata or voice embedding vectors, that is a sign that the platform is collecting more than it advertises. Compare the export contents against the privacy policy. Any discrepancy is worth questioning.

For users who want maximum control, consider a platform that positions itself as a replika nsfw alternative with explicit privacy guarantees. Some alternatives offer on-device processing for certain features, which keeps your data off their servers entirely.

Earn while you recommend

If you are the kind of person who reads privacy policies and tests opt-out settings, you probably have strong opinions about which platforms do it right. Share those opinions with an audience. You can earn a commission by sharing a kupid ai promo code with your followers, or join the ai dating affiliate program to monetize a review site or comparison blog. Your knowledge is worth something.

Common questions

Can I delete my chat history without deleting my account?

Most platforms allow you to delete individual conversations or clear your entire chat history from the settings menu. This does not delete your account or your companion's personality profile. It only removes the text logs from your active view. The data may still exist in backups for a retention period.

Does opting out of training affect my companion's behavior?

No. Opting out only prevents your conversations from being used to train future models. Your companion's current behavior is based on the model version you are using, which was trained on other users' data. Opting out does not change how your companion responds to you.

Do platforms sell my conversation data to third parties?

No major AI companion platform sells raw chat logs. They may share aggregated, anonymized data with research partners or advertisers, but the terms of service generally prohibit selling personal conversation data. That said, a platform could change its policy with a 30-day notice. Read the updates.

How do I know if my voice recordings are being kept?

Check your account settings for a voice data retention option. Some platforms let you disable voice recording storage entirely, which forces the system to discard audio after transcription. If the setting does not exist, assume voice recordings are kept for at least 30 days.

Can I request that my data be removed from the training pool retroactively?

Some platforms honor retroactive opt-out requests if you contact support. They will flag your existing data as excluded from future training cycles. But the data has already been used in previous training runs. Removing it from the model is technically impossible without retraining the entire model from scratch.

What happens to my data if the platform shuts down?

Most terms of service state that user data will be deleted within a reasonable timeframe after shutdown. In practice, the data is often sold as part of an asset sale or transferred to a new owner. There is no legal requirement that your data be destroyed upon shutdown unless the privacy policy explicitly guarantees it.