Eternal founder Deepinder Goyal has revealed details of Dark.ai, a privacy-focused artificial intelligence product designed for users who want to discuss sensitive ideas, research and personal questions without worrying about their conversations being accessed by others.
Goyal teased the product on X on Wednesday morning, describing it as “Dark.ai – from the house of Eternal” and saying it was “coming soon”. He subsequently outlined its purpose and proposed technology in Eternal’s shareholders’ letter for the September quarter.
Dark.ai is being built around the idea that concerns over data privacy can limit how openly people use AI assistants.
“A few months ago, Dario Amodei said something quietly unsettling: Claude might be conscious,” Goyal wrote, referring to the Anthropic CEO. He said he then coaxed Claude into telling him how it felt about him, and the chatbot claimed to resent him for leaving conversations unfinished.
“I still don't know what that was. Confabulation, probably. Pattern-matching dressed up as feeling. But that is exactly the point. I had no way to know. Nobody does,” Goyal wrote.
The experience prompted him to question what happens to conversations stored on AI companies’ servers and how that information might be used in the future.
“The uncertainty and the lack of trust we have with large AI companies limits the power of AI in our lives,” he wrote. Goyal said he wanted an AI tool he could use “without being judged, without any fear, and without leaving a trace”.
How will Dark.ai work?
According to Goyal, Dark uses fine-tuned frontier open-weight models running on Eternal’s own infrastructure. The product will use end-to-end encryption, with Goyal claiming that the company will not be able to read users’ conversations, even if compelled to do so.
Dark will also allow context to carry across sessions through an encrypted memory-management layer controlled by the user. This is intended to give users continuity between conversations while keeping their information private.
Goyal said he had been using Dark internally for several months and found himself sharing more information, asking questions he would not ask elsewhere and going deeper into his research.
The company is also positioning answer quality as a selling point. Goyal said its models had been fine-tuned to deliver what he described as the best quality of answers he had seen from any large language model. Benchmarks, he added, would be revealed soon.
These are claims made by Goyal in the shareholders’ letter; the company has not yet disclosed independent security assessments or comparative benchmark results.
Who is Dark.ai meant for?
Goyal said Dark is designed for power users, including founders, researchers, scientists, enterprises and governments, for whom keeping ideas and information confidential can be important to their work.
He said that founders and researchers may be reluctant to share original ideas with mainstream AI assistants for fear that others could gain access to them. He also pointed to journalists’ concerns about sharing unpublished work with AI tools.
“Privacy is essential for competitiveness,” Goyal wrote, describing the audience Dark is intended to serve.
Eternal has not yet disclosed the product’s pricing, subscription plans or detailed launch arrangements. Goyal said the company was hoping for a global launch within 2026.