Consumer chatbot plans ($0/$20 ChatGPT, Gemini, Claude consumer tiers) store cloud chat histories by default and often require toggling manual privacy settings to opt out of model training. GPT Realtime Assistant ($9.99/mo) narrows exposure across three layers: (1) online audio streams directly to **OpenAI's API** (governed by OpenAI API terms, which do not train on API inputs by default), (2) **Arte Soft's backend never receives or stores your audio, transcripts, or images**, and (3) **Apple Silicon Macs can run Whisper + Qwen 3 8B 100% offline**.
| Privacy Dimension | Consumer Chatbot Apps (Default Settings) | GPT Realtime Assistant (Online Mode) | GPT Realtime Assistant (Offline Mac Mode) |
|---|---|---|---|
| Model training default | Often opted IN to help improve models unless toggled off by user | Uses OpenAI Realtime **API** (OpenAI does not train on API data by default) | Zero external transmission — models run locally on your Mac |
| App publisher cloud transcript storage | Stored indefinitely in cloud chat threads until deleted | Zero — Arte Soft backend never receives audio, transcripts, or attachments | Zero — stays on local Mac SSD |
| Middleman audio proxy server | Processed directly by model vendor | None — ephemeral token connects device straight to OpenAI API | None — zero network sockets |
| Personal email / SMS phone identity | Mandatory verified account | Not required — silent device account or Sign in with Apple | Local Mac execution |
| Works with Wi-Fi & Ethernet unplugged | No | No (requires OpenAI Realtime connection) | Yes — 100% air-gapped after model download |
Consumer App Terms vs API Data Terms: The Distinction Most Users Miss
Major AI labs draw a bright legal line between their **consumer apps** (free and $20/month web/mobile accounts, where conversations frequently help train models unless you locate and disable a 'Improve the model for everyone' toggle) and their **commercial API platform** (where customer API inputs and outputs are excluded from model training by default, subject only to standard short-term safety/abuse logs).
During online sessions, GPT Realtime Assistant connects directly to OpenAI's Realtime **API** using short-lived session secrets—while storing the resulting transcript strictly on your local Apple device rather than on an Arte Soft server.
When Even API Transport Is Too Much: True Offline Mac AI
For defense contractors, journalists, attorneys, founders discussing unannounced M&A, or anyone who wants mathematical certainty that audio never crosses a router, toggling GPT Realtime Assistant to local `Whisper Large v3 Turbo` + `Qwen 3 8B` on an Apple Silicon Mac keeps every waveform and token inside your own hardware.
| System Component | Exact Data Seen | Retention & Control |
|---|---|---|
| Your iPhone, iPad, or Mac local storage | Transcripts, Sessions, Projects, Tasks, Notes, Schedules & local media | 100% under your local device control; delete anytime |
| Arte Soft backend (`assistant-api.wonek.com`) | Anonymous/Apple ID, IAP receipt status, credit balance & session titles | Never sees audio/transcripts; email `office@arte-soft.com` to delete account |
| OpenAI Realtime API (online sessions only) | Live audio stream + attached session images during active online session | Governed by OpenAI's API data usage & abuse-monitoring retention policies |
| Apple Speech (Karaoke Readback) | On-device word matching during readback | Processed locally on Apple hardware (0 credits) |
| Apple Silicon Local Engine (offline Mac) | Live mic/system audio fed to local Whisper Large v3 Turbo + Qwen 3 8B | Processed inside Mac unified RAM; zero internet packets |
Honest Compliance Note on Regulated Healthcare & Legal Entities
While local storage and offline Mac execution provide strong technical privacy, GPT Realtime Assistant is a self-serve App Store product and does not sign custom HIPAA Business Associate Agreements (BAAs); regulated organizations should follow their compliance officer's approved software list.
| Sensitivity Level | Recommended Mode | Why It Matches |
|---|---|---|
| Everyday bilingual travel, study & general work calls | GPT Realtime Assistant Online Mode | API-grade no-training default + zero cloud transcript database |
| Strict client NDA, pre-patent R&D, or confidential executive prep | GPT Realtime Assistant Offline Mac Mode | Air-gapped local Whisper + Qwen 3 8B on Apple Silicon |
| HIPAA BAA / Enterprise SOC2 tenant-wide audit logs | Enterprise OpenAI / M365 / Anthropic BAA Contract | Regulated covered entities need signed enterprise BAAs/DPAs |
| Casual public trivia & web lookups ($0) | Consumer Chatbot (with Temporary Chat / Opt-Out enabled) | Check privacy toggles in account settings first |
Frequently Asked Questions
Does Arte Soft store or read my conversation transcripts?
No. Arte Soft's backend never receives or stores your audio, transcripts, AI answers, or attached screenshots—only account credit balances, IAP receipts, and session titles.
Are online Realtime sessions used to train OpenAI models?
Online sessions stream to OpenAI's commercial Realtime API under OpenAI's API data policies, under which OpenAI does not use API inputs or outputs to train its models by default.
How do I guarantee zero cloud data transmission?
On an Apple Silicon Mac, download Whisper Large v3 Turbo and Qwen 3 8B inside GPT Realtime Assistant and run in offline local mode (you can even turn Wi-Fi off completely).
Related Comparisons & Session Guide
Standalone AI Pins & Pocket Gadgets vs iPhone Voice AI Apps
Compare $199–$699 standalone AI hardware pins, pendants, and pocket devices vs running GPT Realtime Assistant right on your iPhone, iPad, and Apple Silicon Mac.
GPT Realtime Assistant vs Google Translate Conversation Mode
Compare GPT Realtime Assistant vs Google Translate Conversation Mode: 58-language OpenAI Realtime voice streaming and session tasks vs free travel phrases.
