What on-device AI actually is, in plain words

When an AI feature runs in the cloud, your data makes a round trip. Your phone sends your input to a server, the server processes it using a model that lives there, and the server sends the result back. The company operating that server has access to your input at the moment of processing. What they do with it afterward, whether they store it, how long they keep it, whether it trains future models, depends on their policies, which can change.

On-device AI skips that trip entirely. The model is downloaded to your phone during setup and runs using your phone's own processor. When you use it, your words or voice stay on your device. There is no outbound request containing your content, because the computation happens locally. From a privacy standpoint, this removes a category of exposure: the server that could be breached, subpoenaed, or policy-changed out from under you.

The trade-off is capability and hardware. The models that fit on a phone are smaller than the ones running on server clusters. They are genuinely capable for many tasks, including understanding natural language, recognizing patterns, and summarizing what you have said, but they are not the same as the largest cloud models. On-device AI is not a lesser version of cloud AI trying to catch up. It is a different point on the spectrum, optimized for privacy and offline use rather than raw power.

When an app says its AI runs on your phone, it means the model weights, the parameters that encode what the AI knows how to do, are stored locally, and inference, the process of generating a response from your input, happens on your device's chip. Modern phones from roughly 2020 onward have processors capable of running smaller language models at useful speeds.

This is not the same as saying the app never touches the internet. An app can use on-device AI for its core processing and still send other data to servers: usage analytics, crash reports, account information, backups. Whether those things happen, and what they contain, is a separate question from where the AI runs. A privacy-conscious app will tell you explicitly what leaves the device and why.

It also means the app works without an internet connection for its AI features. If you are on a plane or in a spotty coverage area, the AI can still process your input because it is not waiting for a server response.

End-to-end encryption, often written E2EE, means that data is encrypted on your device before it is sent anywhere, and can only be decrypted by the intended recipient, typically using a key that only you or your partner hold. The company running the service cannot read the content, because they do not have the decryption key. This applies while the data is in transit and while it is stored on their servers.

What this means in practice: even if the company's servers were breached, the content stored there would be unreadable without the keys. Even if the company received a legal request for your data, they could hand over encrypted bytes that reveal nothing about the content.

What it does not mean: encryption protects data in transit and at rest on external servers. It does not protect data on your own device if your phone is unlocked and in someone else's hands. It does not protect against an app that collects data before encrypting it and sends that pre-encryption copy somewhere else. Strong encryption is a meaningful protection, but it is one layer, not a complete guarantee. The trustworthiness of any privacy claim depends on the whole chain: what is collected, where it goes, who can decrypt it, and what the company's policies actually say.

For most app features, the cloud AI model is fine and you probably do not think about it. But relationship reflections, conflict patterns, voice notes about hard conversations: these are in a different category. They are the kind of thing you would not want indexed, stored indefinitely, or readable by anyone outside the relationship.

The combination of on-device AI processing and end-to-end encrypted storage means the two most sensitive moments, when your content is being analyzed and when it is being stored, are both handled locally or in encrypted form. The company processing your relationship data ideally never has access to the plaintext of it.

This is not about paranoia. It is about the reasonable expectation that something you said in a private moment should stay private, and about understanding whether the technology actually supports that expectation or just claims to.

On-device AI requires a reasonably modern device. Older phones may not have the chip architecture to run these models at a usable speed, or at all. If a feature seems slow or unavailable on your device, hardware is often the reason.

The models that run on-device are capable but bounded. They work well for understanding, summarizing, and reflecting back what you have shared. They are not the right tool for every task, and an honest app will not pretend otherwise.

No system eliminates all risk. A device that is compromised at the operating system level, or an app with permissions it should not have, can access data regardless of how the AI is architected. Privacy architecture reduces risk; it does not eliminate it. The questions below are a starting point for evaluating whether an app actually earns your trust.

Before you share anything sensitive with an app, these are the questions worth finding answers to, either in the privacy policy, the documentation, or by asking the company directly.

• +Does the AI processing happen on my device, or does my input go to a server? If a server, whose server, and what are their data retention policies?
• +Is my stored data end-to-end encrypted? Who holds the decryption keys, me or the company?
• +What data does the app send off my device, even if the AI runs locally? Usage analytics, identifiers, anything?
• +Does my data train future models, either the company's or a third party's?
• +What happens to my data if I delete my account? Is deletion complete, or are copies retained for a period?
• +Has the company published a clear, plain-language privacy policy, or is the relevant information buried in legalese designed to obscure rather than inform?
• +If I ask the company directly what they can see about me, do they give a straight answer?

Replay processes your voice and reflections on your device. The AI that listens and understands what you share runs locally, so your words do not travel to a server to be analyzed. What you store is end-to-end encrypted, meaning Replay cannot read the content of your relationship memories even if it wanted to. These are architectural choices, not just policy claims, and they come with the real limits described above: you need a compatible device, the on-device model is capable but not unlimited, and no system eliminates all risk.

The goal is to make it possible to keep an honest, private record of your relationship without that record being readable by anyone outside it. If you have questions about exactly what leaves your device and what does not, that is a reasonable thing to ask, and the answer should be findable in plain language.