Inborn vs ChatGPT and the other private AI chat apps: what actually runs offline
Inborn runs the model on your device, so your messages never reach a server and the app keeps working with the radios off. ChatGPT, Gemini, Copilot and Claude run on their companies' computers, so your text has to travel to get an answer. That one difference decides everything else on this page: what is stored, for how long, who can read it, and whether the app works on a plane.
Offline AI is no longer rare. Several good apps run a model on a phone, and one of them is published by Google. This page compares Inborn with the cloud assistants and with the on-device apps, names the places where a rival is the better buy, and shows how to check each claim yourself. Every price, license and permission here was read on 24 September 2026, on the vendor's own public page.
Cloud AI and on-device AI: what actually leaves your phone
A cloud assistant is a text pipe. You type, the words go to a data center, a large model reads them, and the answer comes back. The model is far too big for a phone, which is why it lives there. Microsoft states the consequence plainly in its Copilot transparency note: "Copilot relies on Internet connectivity to function. Disruptions in connectivity can have an impact on the availability and performance of the service."
An on-device app inverts that. The model file sits in the app's storage, your own chip does the arithmetic, and the answer is produced a few centimeters from your hand. Nothing is sent, because there is nowhere to send it. Google's privacy policy notes that "your information may be processed on servers located outside of the country where you live". On a device-only app that sentence has no subject.
The trade is real and we will not dress it up. A 2 GB model on a phone is weaker than a model running on a rack of accelerators. It writes shorter, reasons less far and knows less. What you get back is that the work happens where you are, at the same speed in airplane mode, with no account and no monthly charge.
Inborn goes one step past "it works offline". On Android the release build does not declare the INTERNET permission at all, and a build gate fails the release if that line ever appears. Android will not let an app without that permission open a connection, so the guarantee is enforced by the operating system rather than by our promise.
OUT 0 BCONNECTIONS 0SEALED · ON-DEVICE
What the app's own Proof screen shows on your device — a picture, not a live reading of this page.
Is ChatGPT private? What the cloud assistants store
We could not read OpenAI's policy pages. Both openai.com and help.openai.com returned an error to us on 24 September 2026, so this page quotes nothing from them and prints no ChatGPT price. Their pricing page is linked here and you can read it yourself. What follows is what four other vendors say on their own public pages, in their own words.
Eighteen months by default
The privacy FAQ says that signed in, "this info will be sent to the Microsoft cloud and stored with your Microsoft account". On training: "Except for certain categories of users or users who have opted out, Microsoft uses data from Bing, MSN, Copilot… for AI training." On retention: "By default, we store conversation activity for 18 months." Signed out is different, and Microsoft says so: "If you are not logged into Copilot with a Microsoft Account… we do not train on your conversations."
On by default, and three years for reviewed chats
Google's help pages say "If you're 18 or over, Keep Activity is on by default." Turning it off does not make conversations vanish at once: "Even when Keep Activity is off, your conversations will be saved with your account for up to 72 hours." Chats that went to a human reviewer are handled separately, and Google states they "are not deleted when you delete your activity. Instead, they are retained for up to three years." You can use Gemini signed out, where Google says you "can generally use text-based interactions".
Retention on by default
Its help center says "AI Data Retention is enabled by default", and, about material already collected, "Previously collected training data cannot be deleted or removed." It can be used without signing in.
Thirty days, and no default we can state
Anthropic says deleted conversations are "Deleted from our back-end storage systems within 30 days". On whether training is on by default, their privacy policy and their privacy center do not agree, so we state no default for Claude. Read both and decide for yourself.
None of this is an accusation. These are the published terms of a service that has to receive your text in order to answer it. The point is that the terms exist at all, and that they can change. With Inborn there is no term to read, because there is no transfer to govern. The same list for Inborn, in legal wording and per platform, is section 3 of the privacy policy.
Inborn compared with the cloud assistants
Each cell below holds a short statement rather than a tick. Where a vendor's own page was unreadable to us, or where two of their pages contradict each other, the cell says so instead of guessing.
| Inborn | ChatGPT | Gemini | Claude | Copilot | |
|---|---|---|---|---|---|
| Where the model runs | The model file sits on your device and your own chip produces every answer. | On OpenAI's computers. It is a cloud service. | On Google's servers, which Google says may sit outside your country. | On Anthropic's servers. | In the Microsoft cloud. |
| Works with no connection | Yes, at the same speed. Turn on airplane mode and ask anything. | No. An answer has to come back over the network. | No. The request has to reach Google to be answered. | No. The request has to reach Anthropic to be answered. | No. Microsoft states that it "relies on Internet connectivity to function". |
| Account required | No account, no sign-up screen and no email field anywhere in the app. | Not verified. We could not read OpenAI's own pages. | Not for everything. Signed out you "can generally use text-based interactions". | Anthropic's help pages describe creating an account with a phone number to use Claude. | Not required. Microsoft describes a signed-out mode with different data handling. |
| Price | Free is the whole app. Pro 19.99 USD one time, Work 69.99 USD one time. No renewal. | A free tier and a paid monthly plan. We print no figure: their pricing page returned an error to us. | A free tier and paid monthly plans. Google's own plans page rendered no price for us, so we print none. | A free tier, and Pro listed at 20 USD a month billed monthly on claude.com/pricing. | A free tier, and Microsoft 365 Premium listed at 19.99 USD a month on Microsoft's own page. |
| What happens to your text | It stays in an encrypted database on your device. Nothing is sent, so nothing is stored elsewhere. | Not verified. Their policy pages returned an error to us. | Kept with your account. Keep Activity is on by default at 18 and over, and human-reviewed chats are "retained for up to three years". | Deleted conversations are "Deleted from our back-end storage systems within 30 days". Their two pages disagree on the training default, so we state none. | "Sent to the Microsoft cloud and stored with your Microsoft account", and "By default, we store conversation activity for 18 months." |
| What you can check yourself | Five checks from outside the app: airplane mode, the Play permissions page, aapt2 on the APK, the iOS App Privacy Report, and a firewall. | Nothing we could read. You can still watch your own traffic with a firewall. | Google's help pages state the rules. You cannot inspect their servers. | Anthropic's pages state the rules. You cannot inspect their servers. | Microsoft's pages state the rules. You cannot inspect their servers. |
Quotes read on 24 September 2026 on each vendor's own page. "Not verified" means we could not read the vendor's own page, so we printed nothing rather than a guess.
Inborn, Ollama and LM Studio
If you have a desktop and you are comfortable with a terminal, Ollama and LM Studio are excellent, and both are free. Ollama is MIT licensed and its local use is free and unlimited; its cloud models need an account, and Ollama Pro is listed at 20 USD a month. LM Studio is free for personal use and for work use, runs on Windows, macOS and Linux, answers questions over .docx, .pdf and .txt files, acts as an MCP host, and exposes a local OpenAI-compatible server. Its desktop app is proprietary.
On a machine with 32 GB of memory, both will run models Inborn cannot fit on a phone, and they will run them faster. If that is your situation, use them. We would rather tell you than sell you something smaller.
Neither one runs on a phone. That is the whole gap. Inborn is built for the device that is actually in your pocket on the train, and it is the same app on iPhone, Android, in a browser and on the desktop. The Android release also has no INTERNET permission, which a desktop app has no equivalent of. Model sizes, memory and the speeds we measured are in choosing a model.
Inborn compared with the other on-device AI apps
Running offline is table stakes in this category now. Google's own AI Edge Gallery says no internet is required, and it is free. So the useful comparison is not whether an app works offline. It is what it costs, what license it carries, what it can do with your files, and what it is still permitted to send.
| App | Platforms | Price | License | Where the model comes from | Your documents on the device | Android INTERNET permission |
|---|---|---|---|---|---|---|
| Inborn | iPhone, Android, web browser, Windows and macOS. | Free is the whole app. Pro is 19.99 USD one time, Work 69.99 USD one time. No subscription. | Source-available at github.com/moshecohen90/inborn: read it, build it, publish what you find, but not redistribute or modify it. Not open source. | Instant, 533 MB, ships inside the app and needs no download. Any GGUF file can be imported. | Yes, with Pro. Your PDFs and documents are indexed on the device, with OCR. | Not declared in the release build, so Android will not let it open a connection. A build gate fails the release if the line ever appears. |
| PocketPal AI | iPhone, iPad, Mac and Android. | Free. Its listing names no paid tier. | MIT, open source, so you can fork it. | No model inside the app. You pick a GGUF from Hugging Face in the app. | Not verified. | Declared. We read android.permission.INTERNET in its release manifest. |
| Private LLM | iPhone, iPad, Mac and Vision. Android is a side-loaded beta, not on Google Play. | 4.99 USD one time, with Family Sharing for six people. | Closed source. There is no public repository. | The vendor's own quantized models. It cannot load a file from Hugging Face. | Not verified. | Not verified. There is no Google Play release to read. |
| Enclave AI | iPhone, iPad, Mac and Vision. | Local models are free and unlimited. Pro, which adds cloud models, is 9.99 USD a month. | Closed source. | Models are downloaded inside the app. | Yes, on the free tier. Questions and summaries over your own PDFs. | No Android app. |
| Layla | Android and iPhone. | 19.99 USD one time on the App Store. | Closed source. | About 4 GB downloads the first time the app runs. | Not verified. It does take image input, and it generates images on the device. | Not verified, and the source is closed. Its Play data-safety page says it shares device identifiers and crash logs for analytics, and offers no way to ask for deletion. |
| MLC Chat | iPhone, iPad and Mac. On Android only as an APK from GitHub, not on Google Play. | Free. | Apache-2.0, open source. | Weights download from Hugging Face. Bundling them is a build option, not the default. | Not verified. | Declared. We read it in its manifest. |
| Google AI Edge Gallery | Android and iPhone, published by Google. | Free. | Apache-2.0, open source. | Models download inside the app, and you can load your own. | Not verified. It takes image and audio input. | Declared, along with camera, microphone, calendar, accounts and push messaging. |
| LM Studio | Windows, macOS and Linux. There is no phone version. | Free for personal use and for work use. | The desktop app is proprietary. Its command-line tools are MIT. | Models download inside the app. | Yes. Questions over .docx, .pdf and .txt files. | No Android app. |
| Ollama | Windows, macOS and Linux. There is no phone version. | Local use is free and unlimited. Cloud models need an account, and Pro is 20 USD a month. | MIT, open source. | Models are pulled from ollama.com/library. | Not verified. | No Android app. |
"Not verified" means we could not confirm it on the vendor's own public page or in source they publish, so we left it rather than guess. On the permission row we read the release manifests of six on-device Android apps: PocketPal AI, MLC Chat, Google AI Edge Gallery, SmolChat, LLM Hub and MyDeviceAI. All six declare android.permission.INTERNET. Inborn's release build does not. We have not read every Android AI app that exists, so this is a statement about those six.
Two more worth knowing about. LLM Hub is free with a 9.99 USD lifetime tier, on both stores, source-available under a PolyForm Noncommercial license, and needs no account; its manifest declares INTERNET along with location, contacts and calendar. Apple Intelligence is free and built into the operating system, and Apple's own page says it "can draw on larger server-based models" through Private Cloud Compute. Gemini Nano is an Android system service rather than a chat app, and Google states that "AICore does not have direct internet access".
No account, no login, no subscription
Inborn has no account. There is no sign-up screen, no email field, no profile, and nothing on our side that could hold a record of you. A purchase is tied to the store you bought it from, because there is no account that could carry it anywhere else. That is a real limitation and it is the price of having no account at all.
The cloud assistants are not all account walls. Gemini, Copilot and Perplexity can each be used signed out, with fewer features and, in Microsoft's stated case, different training treatment. So the honest claim is narrower than "cloud apps make you log in". It is that signing in is where the storage, the history and the retention clocks begin. Microsoft is explicit about it: signed in, your text is "sent to the Microsoft cloud and stored with your Microsoft account".
On payment, Inborn has no subscription and will not add one. Free is the entire app, with unlimited chat, no message quota and no watermark. Pro is 19.99 USD once. Work is 69.99 USD once. If we stopped shipping updates tomorrow, the copy you bought would keep working, because there is no license server to switch off. The three tiers are on the pricing section of the home page.
Where the others are better than Inborn
- Documents on a free tier. Enclave AI answers questions about your PDFs at no cost, with image, PDF and code attachments. In Inborn, documents and OCR sit behind Pro.
- Images on the device. Layla generates images locally with Stable Diffusion, takes image input, generates music and runs Python agents. Inborn does none of that.
- Price on an iPhone. Private LLM is 4.99 USD one time, a quarter of Inborn Pro, with Family Sharing for six, Siri and Shortcuts support, and system-wide text tools on macOS. Inborn has not built those.
- A license you can fork. PocketPal AI is MIT and genuinely open source. Inborn is source-available: you can read it, build it and publish what you find, and you may not redistribute or modify it. That is a weaker freedom and we say so rather than borrow the words "open source".
- A desktop with memory to spare. LM Studio and Ollama are free and will run larger and better models than any phone can hold.
- Nothing to install. Apple Intelligence is already on the phone and costs nothing, though Apple notes it can reach server-based models. On Android, Google AI Edge Gallery is free, Apache-2.0 and published by Google itself.
- Raw capability. Every cloud assistant is stronger than any model that fits on a phone. If your work needs the best reasoning available and you accept the terms in section 02, use one of them.
Inborn is for the times and the material where you would rather nothing left the device.
How to check any of this yourself
Nothing on this page needs you to trust us. Five checks on Inborn work from outside the app, on the build you installed: airplane mode, the App permissions page on the Play listing, aapt2 dump permissions on the APK, the iOS App Privacy Report, and a firewall such as NetGuard, Little Snitch or LuLu. Each takes under a minute.
The instructions are on the proof page. The reasoning behind each check, including the three things none of them proves, is in how the proof works. Why any of it is worth the trade is in why on-device. The app's source is at github.com/moshecohen90/inborn, source-available: read it, build it, publish what you find.
Questions people ask when comparing
Inborn vs ChatGPT: what do I gain and what do I lose?
Inborn gains you an assistant that never sends your text anywhere and answers at full speed with the radios off. You lose capability: a model that fits on a phone writes shorter, reasons less far and knows less than a model on a rack of servers. You also lose a price comparison here, because OpenAI's pages returned an error to us.
What is the best offline AI chat app for iPhone?
Inborn is the one we built, so read this as our answer and check the others yourself. On iPhone, Private LLM is 4.99 USD one time with Siri support, and Enclave AI answers questions about your documents on its free tier. Inborn ships the model inside the app, imports any GGUF, and runs the same on Android, web and desktop.
Is there an AI chat app that works with no internet?
Inborn works with no internet, and so do several others, because the model runs on your device rather than on a server. Turn on airplane mode and ask it anything; the answer arrives at the same speed. PocketPal AI, MLC Chat, Layla and Google's own AI Edge Gallery run offline too. Offline is normal in this category now.
Which private on-device AI apps are actually private, and how do I verify it?
Inborn is the one whose Android release ships without the INTERNET permission, which the operating system enforces for you rather than asking you to trust a promise. Verify any app in three ways: run aapt2 dump permissions on its APK, read the App permissions page on its Play listing, and put it behind a firewall. The six on-device Android apps whose source we read all declare INTERNET.
Is there an AI chat app with no account and no sign-up?
Inborn has no account, no sign-up screen and no email field anywhere in it. PocketPal AI, Enclave AI, Private LLM and LLM Hub also state that no account is required. Gemini, Copilot and Perplexity can each be used signed out with fewer features, though signing in is where their storage and retention clocks begin.
Is there a one-time purchase AI app instead of a monthly subscription?
Inborn is a one-time purchase and will not become a subscription. Free is the whole app, with unlimited chat and no message quota. Pro is 19.99 USD once and Work is 69.99 USD once. Private LLM is 4.99 USD once, Layla is 19.99 USD once on iOS, and LLM Hub has a 9.99 USD lifetime tier.
Inborn vs PocketPal AI: which should I use?
Inborn ships a working model inside the app and its Android release has no INTERNET permission. PocketPal AI is free, MIT licensed and genuinely open source, so you can fork it. PocketPal downloads its models from Hugging Face and its release manifest declares INTERNET. If an open license matters more to you than the permission, choose PocketPal.
Inborn vs Private LLM: which should I use?
Inborn is the cross-platform one and Private LLM is the cheaper one. Private LLM is 4.99 USD one time, a quarter of Inborn Pro, with Siri, Shortcuts, macOS text tools and Family Sharing for six. It is closed source, limited to its vendor's own quantized models, and its Android build is a side-loaded beta.
Whether Inborn is open source, and what a small model gives up against a large one, are answered on the home page.