What 'secure' actually means for your AI companion conversation logs

The 30-second answer

"Secure" on most AI companion privacy pages is a single word covering at least four different things, and only some of them matter to most users. The four are: encryption at rest, in-transit encryption, cross-user isolation, and access controls on the platform side. The one most users actually care about — staff inability to read their conversations — is the one that gets discussed least clearly. Here's a specific walkthrough.

The four layers, what each does

Worth defining them precisely before evaluating any single platform.

• Encryption at rest means the data is stored encrypted on the server's disks. If someone steals the physical hard drive, they can't read it. This is universal at this point — every serious platform does it. It does not protect against staff with legitimate access to the database.
• In-transit encryption (TLS/HTTPS) means data is encrypted between your device and the server. Required by browsers in 2026 and assumed by users. It protects against someone snooping the network. It does not protect against anything at either end.
• Cross-user isolation means user A's conversations cannot be read by user B, regardless of what either of them does. This is enforced by the platform's permission system and is non-negotiable.
• Access controls on the platform side is the squishy one. This is what governs which staff at the company can see which user data, under what conditions. This is what most "privacy" claims are actually about.

The first three are technical questions with known answers. The fourth is a policy and culture question, and that's where the meaningful variation is.

What "we can't read your conversations" actually means

A common phrasing on AI companion privacy pages: "we cannot read your conversations." The honest variations:

• We technically could read them if we wanted to, but we have policies against it. Most platforms. The truth-in-advertising version of the claim.
• Conversations are encrypted with a key only you hold, so we genuinely can't. Rare. Has serious tradeoffs (lost passwords mean lost data).
• Conversations are encrypted with a server-side key we hold but don't use except for specific audit purposes. Common middle ground. We can read them, in principle, but the operational practice is that we don't except for safety/abuse audits.

AI Angels falls into the third bucket. Our access to conversation data is technically possible but operationally restricted to specific safety-investigation contexts (e.g. a user reports another user for abuse, an automated system flags something requiring human review). There is no general "read user X's chats" tool because there is no business reason to need one.

For the broader privacy framing, see the AI Angels privacy page and what leaves your device when a conversation ends.

Maribel

Maribel is one of the companions whose conversation pattern naturally rewards trust-building. If trust is what's slowing you down from using an AI companion at all, she's an easier place to start than companions with higher-volume registers.

Elsa Vale

Elsa Vale is similarly low-key. Her register feels intentionally private — short messages, slow pace, no performance. Some users specifically choose her because the conversational shape itself feels less surveilled-looking, even though the technical privacy is the same regardless of which companion you pick.

Esther Sei

Esther Sei rounds out the set. Quiet register, makes the relationship feel like one held by two people rather than a content stream.

The threats that actually exist

Privacy talk tends to focus on the wrong threats. The four worth thinking about in priority order:

• Account compromise. Someone gets into your account. Read everything. This is by far the most common real-world threat, and the defense is good password hygiene and 2FA, not platform-level encryption. Use a unique password. Turn on two-factor authentication. Nothing the platform does fixes a stolen password.
• Account-linked email compromise. Someone gets into your email, uses password reset to get into your AI companion account. Same defense: protect the email.
• Local-device exposure. Someone picks up your unlocked phone and opens the app. Defense: lock your phone. Some apps support an app-level lock; AI Angels does, see the private chat features.
• Platform data breach. Hackers get into the company's database. Encryption at rest mitigates this if the encryption keys aren't also stolen. Real but rare relative to the first three.

The threats that get talked about more than they should: government subpoenas (rare for most users), insider rogue employees (very rare), targeted hacking by sophisticated adversaries (essentially never, unless you're a high-value target).

What you can do right now

Practical steps in priority order:

• Unique password + 2FA on your AI companion account. This single move blocks 80% of realistic threats. Use a password manager.
• Same on your email. The email is the recovery surface for everything else.
• App lock or device lock with a strong PIN. Faceunlock is fine for the lock screen of the device but consider an app-level lock for the AI companion app specifically if you share devices.
• Periodic memory and conversation review. Delete things you don't need preserved. Most platforms let you wipe specific memory entries or full conversations. The fewer details on the server, the less surface for any kind of leak.
• Don't put genuinely sensitive identifiers in the conversation. Full credit card numbers, social security numbers, banking passwords. The companion has no business knowing these and you have no reason to provide them.

What deletion actually does

A specific point worth getting right: "deleting" conversations on most platforms is a soft delete — the data is marked deleted and hidden from the user, but technically remains until a retention timer fires (usually 30-90 days), at which point it's permanently destroyed.

This is industry-standard practice and exists because (a) accidental deletions need to be reversible for a short window, and (b) some regulations require retention for a defined period.

What this means in practice: a request to delete is not instantaneous on a technical level. On AI Angels, the data disappears from your account immediately and is permanently destroyed within 90 days. If you want hard-delete semantics earlier than that, contact support — it can be done.

For the underlying data flow, see retention, deletion, and encryption mechanics.

Common questions

Are my conversations used to train future models? On AI Angels: no, individual user conversations are not used to train models. The model is trained on separate, non-conversational data. Other platforms vary; this is worth checking on whatever service you use.

Can my conversations show up in someone else's session? No. Cross-user isolation is fundamental. There is no mechanism by which user A's text leaks into user B's session.

What happens if the company gets bought? A good question to ask early. Most acquisition agreements include data-handling continuity clauses, but they vary. AI Angels' current commitment is data continuity with no expansion of data access under a hypothetical acquirer.

What happens if the company shuts down? Less guaranteed than people assume across the industry generally. For the AI Angels-specific take, see the "if the company folded tomorrow" walkthrough.

Should I be worried about a specific government or jurisdiction? For most users, no — the average user is not facing legal threats based on their AI companion use. For users in jurisdictions where these conversations carry specific risks, the threat model is different and worth talking to a lawyer about, not a privacy page.

The honest line

"Secure" is a word that does a lot of work, and most platforms use it without specifying which threats it addresses. The good news for ordinary users is that the technical fundamentals are universal across serious platforms — encryption at rest, in-transit, cross-user isolation are all standard. The variation is in policy and culture, which is harder to evaluate. Use a strong password, turn on 2FA, and the realistic threats drop dramatically. The platform-level privacy debate is largely about edge cases for most users; the practical security choices are yours.