Quick Answer
- Maple has the best privacy and supports anonymous accounts. It uses open-source AI models which are pretty good but not cutting-edge. However, this architecture avoids having your queries used for training.
- Perplexity offers limited privacy in “Incognito Mode”. It deletes threads after 24 hours and doesn’t save history, but still processes data on their servers — it’s convenient but not truly private. However, it’s a polished product. It offers the latest AI models, and has a great mobile app. But it is somewhat cluttered by upselling and other mild annoyances.
- Kagi’s Assistant has solid privacy features. It basically acts as an anonymous proxy to the latest AI models, and threads expire in 24 hours. It works alongside the excellent Kagi search engine but doesn’t exist as an independent app.
- Proton’s Lumo has good privacy and is nicely integrated into the Proton ecosystem, but it is limited to open-source AI models. It encrypts chats end-to-end, keeps zero logs, and offers “Ghost mode” for disappearing conversations.
It’s important to have privacy when using AI chats. People often ask sensitive questions and upload privaet documents. They might not bring up these private topics with their spouse or doctor. Here, we examine the actual private AI options that are available.
How They Handle Privacy
Kagi’s Assistant
Kagi’s Assistant is a AI research tool with access to top AI models. Kagi say that Assistant prompts are never used to train models, either by Kagi or the upstream LLM vendors, and your Kagi account identity is not forwarded to those vendors. Threads auto‑expire (by default after 24 hours, configurable).
Lumo’s Zero-Access Encryption
With Proton’s Lumo your conversations get encrypted on your device before hitting Proton’s servers. The company uses a technique called “user-to-Lumo encryption” where each request generates a fresh 256-bit AES key.
The AI server decrypts your message, processes it, and immediately forgets it. Lumo keeps no logs server-side. If you save conversations, they’re encrypted with keys only you control.
Ghost mode takes this further — close the window and your chat disappears forever.
Maple AI’s Confidential Computing
Maple takes a different path. The platform uses Amazon’s Nitro Enclaves — specialized hardware that processes data in an encrypted state.
Your message gets encrypted on your device. The secure enclave decrypts and processes it without Maple’s team ever seeing the plaintext. They’ve published their enclave code for public verification through a process called attestation.
Maple recently added anonymous accounts. You can generate a random ID, set a password, pay with Bitcoin (no email is required).
Perplexity’s Incognito Limitations
When using Perplexity‘s incognito mode, your threads don’t appear in history and expire after 24 hours.
However, Perplexity still processes your queries through their standard infrastructure. The company states threads expire after 24 hours, though some users report data retained for 30 days on servers.
Also, only Enterprise customers are excluded from AI training by default.
Features Comparison
Proton’s Lumo uses open-source AI models, namely: Mistral Small 3, Nemo, OpenHands 32B, and OLMO 2 32B. The AI automatically picks the right model for your task. Coding questions go to OpenHands, and general queries use Mistral or Nemo.
Web search is disabled by default. Turn it on and Lumo uses privacy-focused search engines. You can upload files and connect Proton Drive. The free version gives you 100 queries per week with 7-day chat history. Lumo Plus costs $13/month for unlimited chats.
Maple AI offers the widest model selection: Kimi K2 Thinking, DeepSeek R1 671B, Llama 3.3 70B, Gemma 3 27B, Mistral Small 3.1, Qwen 2.5 72B and GPT-OSS 120B.
Free accounts get 10 chats per week. Starter ($5.99/month) bumps that to reasonable limits. Pro ($20/month) unlocks all models and file uploads.
Perplexity excels at research. The platform searches the web in real-time, provides citations, and can generate comprehensive reports. Pro subscribers ($20/month) get 300 searches daily, access to GPT-4o and Claude 3.5 Sonnet, unlimited file uploads, and API credits.
Their docs are explicit that Assistant prompts are never used to train models, either by Kagi or the upstream LLM vendors, and your Kagi account identity is not forwarded to those vendors. Threads auto‑expire (by default after 24 hours, configurable), which is about as sane as it gets for a hosted assistant.
The Data Training Question
Lumo promises never to use your data for AI training. Proton’s European base means GDPR protections apply.
Maple makes the same commitment. Their confidential computing architecture means they literally can’t access your conversations to train models.
Perplexity is different. By default, the platform uses your data to improve their models. You must manually disable “AI data retention” in settings. Even then, incognito mode doesn’t guarantee your queries won’t be processed for safety and debugging purposes.
The Encryption Reality
Lumo encrypts data in transit and at rest. But the AI itself sees your plaintext messages — that’s how language models work. The privacy comes from immediately forgetting your query and encrypting any saved history.
Maple’s secure enclaves process data in an encrypted state. This is technically stronger — the host server never sees your plaintext. But you’re trusting AWS’s Nitro Enclave hardware and Maple’s implementation.
Perplexity incognito mode doesn’t encrypt anything. It just skips saving to your account history. Your data still flows through Perplexity’s standard infrastructure.
What They Cost
Lumo’s free tier gives you 100 queries per week — enough for casual use. Lumo Plus ($13/month or $120/year) removes all limits.
Maple offers the most granular pricing. Free gets you started. Starter ($5.99/month) handles regular use. Pro ($20/month) unlocks everything including models like DeepSeek R1 and Kimi K2.
Perplexity offers a free plan with limited “Pro” searches, a $20/month Pro plan (or $10/month for students), a $200/month Max plan for heavy users, and higher-priced Enterprise tiers starting around $40/user/month
