Can Your ISP or Employer See Your AI Companion Chats?

There are three different questions hidden inside “can someone see my AI chats?”

1. Can the network provider see that you connected to an AI service?
2. Can an employer monitor the device or network you are using?
3. Can the AI company read the prompt you sent to its model?

Those are different observers with different access. Treating them as one privacy problem leads to bad advice.

The short answer is simple. A cloud AI provider receives the prompt because its servers generate the reply. Your ISP usually sees network metadata, not the words inside a properly encrypted connection. An employer may see much more if you are using a managed computer or a monitored company network. A local model can remove the cloud provider from the conversation, but it cannot turn a work laptop into a private device.

Your AI provider sees the prompt when its servers generate the reply

The short version: if the model runs in the cloud, the provider needs your prompt, relevant conversation context, and any attached data required to produce the answer.

This is a technical fact, not an accusation. A server cannot generate a reply from text it never receives. The provider may encrypt the connection while the message travels there, but encryption in transit does not mean the provider cannot read the message after it arrives at the service that runs the model.

What happens next depends on the product. The provider may retain chats, use them to improve models, send them to contracted processors, expose them to safety or abuse teams under defined conditions, or offer controls such as temporary chats and training opt-outs. Those details belong to the provider’s current policy and your account type.

For example, OpenAI’s consumer privacy page describes temporary chats and related controls, while its enterprise privacy page describes a different data arrangement for business products. Do not transfer the privacy promises of a business plan to a personal account.

I wrote about the same issue in Do AI Companion Apps Read Your Messages?, using the policies of named companion services rather than guessing from their landing pages.

Your ISP usually sees the connection, not the chat text

The short version: HTTPS normally hides the content of your AI conversation from your ISP, but it does not make the connection invisible.

When a cloud app sends a prompt, your device opens an encrypted connection to a remote service. Someone watching the network path can often learn that your device connected to a particular domain, address, or service. They may also see timing, connection duration, and the amount of data transferred.

They usually cannot read the actual prompt or reply inside a correctly protected HTTPS session. That is the point of transport encryption. The distinction matters because “your ISP cannot read the message” is not the same statement as “your ISP cannot tell that you used the service.”

DNS can reveal another piece of the picture. Standard DNS queries are often sent without encryption, which can expose the domains a device asks about. Cloudflare’s DNS documentation explains why DNS queries can reveal which sites and services a device looks up, even when the website content is encrypted. DNS over HTTPS can protect those queries from observers on the network path, but it does not erase every other kind of metadata.

There are also edge cases. A badly configured app, an outdated device, malware, or a network that deliberately intercepts encrypted traffic can change the result. Do not use a general rule as a substitute for checking the exact setup.

An employer can see more on a managed computer

The short version: a company-owned or company-managed laptop should be treated as monitored infrastructure, not as a private notebook.

An employer may control the network, the device, the browser, or all three. Depending on company policy and local law, monitoring can include DNS logs, web filtering, browser history, installed endpoint agents, screen capture, application inventory, and security inspection of network traffic.

Some corporate networks use a proxy or TLS inspection system that installs a company certificate on managed devices. That can allow the organisation’s security tools to inspect traffic that would otherwise be protected between your browser and a cloud service. Whether that is enabled, and whether it is lawful in a specific workplace, depends on the setup and jurisdiction. The safe practical rule is less complicated: do not put intimate conversations into an employer-managed device or account.

The device itself matters as much as the network. If monitoring software can record your screen or capture keystrokes before the text is encrypted, HTTPS cannot protect that text from the monitoring software. Encryption starts after the application has received your keystrokes. It cannot defend against someone who already controls the endpoint.

The Electronic Frontier Foundation’s overview of bossware documents the wider worker-monitoring problem. This is not specific to AI companions. AI chats simply make the consequences more personal because people often tell a companion things they would not put into an ordinary search box.

Local AI removes the cloud provider from the message path

The short version: when a local model generates the reply on your computer, the prompt does not need to travel to an AI company’s server.

That changes the most important part of the privacy picture. There is no remote inference provider receiving every message just to generate the next one. A network monitor should not see a cloud chat request for that local inference path.

Local does not mean invisible, though. A local application may still connect for model downloads, update checks, license services, cloud fallback, web search, weather, messaging integrations, or optional cloud voice and image providers. The exact feature and setting matter.

Local Waifu’s privacy policy lists the default on-device paths and the features that can send data elsewhere after you enable them. The app’s local conversation, memory, and local media paths are separate from optional cloud providers. You can also use the offline verification walkthrough to watch what the application and its helper processes actually do.

This is the distinction worth keeping:

• A cloud model sends the prompt to a provider by design.
• A local model can generate the prompt on your own hardware.
• A managed work device can still expose either conversation through endpoint monitoring.
• Optional online tools create their own network paths.

Local AI does not defeat employer monitoring

This deserves its own section because “local” is easy to overread.

Suppose you run a local companion on a company laptop. The prompt may never leave the machine for inference, which is better than sending it to a cloud provider. But the employer may still have access to the laptop, its files, its screen, its keyboard events, its user account, or its security logs. If the company owns the device, local inference is not a promise of private use.

The same is true for a personal laptop used on a company network. A network administrator may see the device, connection metadata, or blocked application attempts. They generally cannot read a local conversation that never creates a network request, but they may still know that the application is running if endpoint or network controls reveal it.

The safest setup for an intimate conversation is a personal device, a personal account, and a local model with optional network features turned off. Even then, protect the device itself with a strong login, disk encryption, and sensible access controls.

How to think about the privacy boundary

A useful privacy check has four layers:

1. The model layer: where is the inference performed?
2. The network layer: what domains and services receive data?
3. The device layer: who can access the computer, files, screen, and input?
4. The policy layer: what does each provider retain and allow staff or contractors to access?

A cloud app can have good transport security and still receive your prompt. A local app can keep the prompt on-device and still have an update service. A work laptop can expose text before either kind of app has a chance to protect it.

That is why What “Runs Locally” Actually Means separates local inference from total network silence. The words sound similar. The privacy boundary is not.

The practical answer

Can your ISP or employer see your AI companion chats?

Your ISP can often see that your device connected to an AI service and can collect network metadata. Proper HTTPS normally hides the message content from the ISP. Your employer may see more on a managed device, through endpoint software, browser monitoring, a company proxy, or TLS inspection. A cloud AI provider receives the prompt by design. A local model can keep that prompt away from the cloud provider, but it cannot make a company-controlled device private.

If the conversation is personal, do not have it on a work laptop or company account. If privacy matters, run the model locally on a device you control, keep optional cloud features off, and verify the network behavior instead of relying on a badge that says private.

FAQ

Can my ISP read my AI companion messages?

Usually not when the app uses properly configured HTTPS, because the message content is encrypted in transit. Your ISP may still see connection metadata such as the destination, timing, traffic volume, and sometimes the domain or service address.

Can my employer see AI chats on a work laptop?

Possibly. A company-managed laptop may have endpoint monitoring, browser monitoring, a corporate proxy, screen recording, or TLS inspection. The company network may also record destinations and timing. Do not treat a work device as private.

Can an AI company read my cloud chats?

The cloud provider receives the prompt and the context needed to generate a response. What it stores, who can access it, and how long it keeps it depend on that provider’s policy, account type, settings, and legal obligations.

Does local AI hide chats from my employer?

Local AI can keep the prompt away from a remote AI provider and can produce no chat request for a network monitor to observe. It cannot protect a conversation from monitoring software, screen capture, keystroke logging, or administrator access on a company-managed computer.

Can my home Wi-Fi owner see my AI chats?

A home network operator can often see connection metadata, but HTTPS normally prevents them from reading the message content. Someone with access to the device itself, router logs, or monitoring software may be able to learn more.

Does using a VPN make AI chats private?

A VPN moves trust from your ISP or local network to the VPN provider. It can hide some destination information from the local network, but the AI provider still receives the prompt if you use a cloud model. A VPN also does not protect text from monitoring software on the device itself.

Sources: OpenAI consumer privacy, OpenAI enterprise privacy, Cloudflare DNS privacy documentation, Cloudflare DNS over HTTPS setup, Electronic Frontier Foundation on bossware, and Mozilla Foundation on chatbot privacy.

Questions people ask

Back to the blog