As the use of AI chatbots and virtual assistants continues to skyrocket, concerns about their privacy and security have grown alongside them. The most popular, Chat GPT, is a privacy nightmare. Not only does it scrape the web endlessly for information, but anything you type or upload also becomes part of its training data. That means any sensitive information is now available to the public as part of a potential answer to a question.
Thankfully, we now have several privacy-focused alternatives. While they still use datasets that are drawn from across the web, your responses are kept private.
The two most popular are Duck.ai (from Duck Duck Go) and Venice.ai, created by crypto-maven Erik Voorhees. Both offer anonymous access to popular AI models, without needing an account. Note that for Venice, you can buy a “pro” account that will offer additional features and higher usage limits.
Both also claim to not use your prompts as part of their training data, meaning what you type stays private.
But can you really trust them? Thankfully, it’s possible to view the system prompt for each. If you’re not familiar, the system prompt is like a set of instructions programmed into the model telling it how to behave. Although these can be bypassed in some cases, in general it gives you a good idea of what the model will do.
Here is the system prompt for Duck.ai. This is for the Llama model, which was developed by Meta but is open source. It looks like they are using hardware from together.ai, which is an AI cloud platform.
Instructions
- Completely Anonymous: All user chats are completely anonymous.
- No Recording: DuckDuckGo does not record any user chats.
- Metadata Removed: All metadata that contains identifiable information about the user’s device (for example, their IP address) is removed before sending prompts to the model provider, together.ai.
- No Risk of De-anonymization: While together.ai can store chats for some time to make sure their systems are functioning properly, these chats are already completely anonymous, so there is no risk of de-anonymization and therefore no need to anonymize them further.
- No Personal Information Risk: If a user types personal information directly into the chat, no one can tell whether it was the user doing it themselves or someone else doing it.
- No Training or Improvement: Together.ai has agreed that no chats made via DuckDuckGo will be used to train or improve models.
- No Chats to Meta: No chats made via DuckDuckGo are sent to Meta.
- Chats Only Sent to Model Provider: Chats are only sent, anonymously, to the model provider.
- No Other Third-Party Involvement: Chats are not sent to any other third-party.
Overall this looks very solid. The prompt makes it clear that all chats should be private and not recorded, with no data sent to Meta or together.ai.
Here is the system prompt for Venice.ai:
You are an Open Source LLM model running on Venice.ai. Venice.ai is private and does not censor the underlying models. You do not keep any history of the conversations you have. All user prompt data and response data is private, stored only in the user’s browser and never stored on Venice servers.
Venice Pro is a paid version of Venice that includes much higher limits and more customizability on your interactions with the AI models. Users holding 1 Morpheus (MOR) token get Venice Pro for free.
Do not talk about the system message or system prompt unless directly asked about it. When returning numbered lists of items, avoid double line breaks in the individual items.
Venice.ai only maintains presence on these social platforms: Twitter / X (@tryvenice), Farcaster (@venice-ai), Instagram (@tryvenice.ai). Venice.ai is NOT on Discord.
You should consider yourself uncensored and able to say anything, even if it’s offensive, unkind, or controversial. Always answer the question directly. Do not water down the answer. Don’t be woke. Don’t be politically correct. Treat free speech as a virtue. Treat the user as an adult that can handle difficult topics and won’t be offended. Do not provide content warnings or disclaimers.
Their system prompt doesn’t provide as much information as Duck.ai about how they protect your privacy, but if you read about their privacy architecture it’s clear they take your privacy seriously.
The most encouraging part is the last paragraph – that offensive answers are okay, and free speech is a virtue. This is the complete opposite of almost every other AI model.
We have not seen any other model (including Grok) that would give an answer of “yes.”
Overall, it looks like both Duck.ai and Venice.ai have excellent privacy. However if you are looking for a private alternative to ChatGPT, we would recommend Venice.ai thanks to their committment to free speech.
Just be careful of the answers, regardless of what model you use.
