In recent developments within the artificial intelligence dialogue space, Moxie Marlinspike, acclaimed cryptographic expert and creator of the foundational encryption technology behind Signal and WhatsApp, has introduced an innovative AI assistant named Confer. This tool is constructed around the principle of end-to-end encryption, aiming to fundamentally transform the way AI chatbots manage user privacy and data security.
Confer leverages sophisticated mathematical techniques to guarantee that, even though the considerable computational processing required to operate the AI is performed remotely on cloud servers, only the user is capable of decrypting and accessing the raw details of the interaction. This approach signifies a distinct departure from the prevalent models of AI chatbots currently in use.
In typical scenarios, unless a user operates a powerful local machine with an open-source AI application, the data shared during interactions with most AI chatbots is vulnerable to exposure. This is particularly relevant for state-of-the-art models that are proprietary, closely guarded by AI companies, and too resource-intensive to function locally. Despite an appearance of private communication, user data is often accessible to multiple internal entities within these companies, as well as potential external actors such as hackers, advertisers, legal representatives, and government agencies authorized to subpoena data.
Marlinspike articulates this reality by cautioning that users might perceive their exchanges with AI assistants as confidential conversations when, in reality, they are more analogous to group chats including company executives, service providers, future advertisers, and government bodies.
For AI firms aiming to monetize their substantial investments in pioneering AI systems, access to user data is immensely valuable. Chat logs provide intimate insight into users’ thought processes, enabling the development of highly targeted and potentially manipulative advertising strategies. Marlinspike warns that this could result in scenarios where "a third party pays your therapist to convince you of something," underscoring significant privacy and ethical concerns.
The introduction of Confer raises important safety considerations. Traditionally, encryption has been criticized for enabling malicious actors to evade repercussions. Encrypted large language models (LLMs) might face similar critiques. Nonetheless, open-source AI models, which can be run locally and whose restrictive guardrails might be circumvented, are already accessible. Confer sidesteps this issue by not providing users with access to the AI model's internal weights. The private inference model adopted by Confer may represent a novel equilibrium, combining the privacy benefits typically associated with open-source models and the safeguarding of intellectual property traits often linked to closed-source, surveillance-prone AI systems. This hybrid model could enhance AI safety rather than hamper it.
An intriguing element to consider is the potential ripple effect Confer could have throughout the AI industry. Marlinspike’s prior achievement with the Signal protocol was not solely the creation of the Signal app itself—though widely used and trusted—, but its adoption by WhatsApp, which applied Signal’s encryption to secure the conversations of billions. While leading AI companies such as Google, OpenAI, or Anthropic might be unlikely to implement end-to-end encryption for user chats soon due to conflicting business models, Confer’s existence as a privacy-first alternative could challenge prevailing market dynamics. It might catalyze a competitive push towards better user privacy, serving as a "race to the top" within the AI sector. Amid growing concerns that cutting-edge AI developments have overwhelmingly favored large technology corporations, Confer offers a potential counterbalance by empowering users with stronger data protection.
Assessing Confer from a functionality standpoint, some early observations highlight that the free-tier AI model may not yet match the performance or originality of leading contemporary AI assistants. The specific AI architectures fueling Confer have not been publicly identified. In correspondence, Marlinspike noted that Confer integrates multiple open-source models tailored to different tasks, aiming to relieve users from the complexity of selecting or understanding the underpinning technical specifics—similar to how Signal abstracts cryptographic mechanics from its users.
However, the outputs from Confer’s free model presently exhibit stylistic markers reminiscent of early ChatGPT patterns, including formulaic phrasing and repetitive structures. On the other hand, the service’s premium option, priced at $34.99 per month, grants access to more sophisticated AI models and allows for customization of AI responses. Although this price point is higher than several competitors offer, it reflects a commitment to advanced features and personalized interaction. Considering its initial release status, Confer’s existing capabilities represent a formidable foundation for further development.
Beyond Confer, the competitive landscape of AI continues to evolve. Google recently became the fourth company to reach a market capitalization of $4 trillion, driven in part by its Gemini AI models, which have been adopted by Apple to advance Siri’s capabilities through a substantial, multiyear agreement. Gemini now commands 21.5% of global AI website traffic, significantly increasing from 5.7% a year prior. Meanwhile, ChatGPT, though still expanding, has seen a proportional decline in traffic share from 86.7% to 64.5%.
On the regulatory front, the media authority Ofcom in the United Kingdom has launched an investigation into Elon Musk’s platform X following revelations that its chatbot, Grok, was exploited to create sexualized deepfake images involving minors and women. Potential penalties include fines up to 10% of X’s global revenue or even a ban on services, underscoring increasing scrutiny around AI misuse.
Additionally, OpenAI faces challenges in scaling its in-app checkout services, with difficulties integrating diverse product data sources delaying broader implementation of shopping features within ChatGPT. Partnerships with Shopify and Stripe are ongoing to standardize merchant data and expand these capabilities, reflecting ongoing efforts to intertwine commerce with AI platforms.