The web is no longer quite the space for empowerment it promised to be. That is the simple yet stark observation made by Sir Tim Berners-Lee, the man who invented it. At VivaTech, his remarks were not at all nostalgic. He spoke less about the Internet’s past than about its next tipping point: artificial intelligence as personal memory.
Together with John Bruce, co-founder of Inrupt, Berners-Lee advocates for Charlie, an assistant designed not to feed large language models, but to protect individuals when they use them. The issue is no longer just privacy in the traditional sense. It’s about individual sovereignty: who owns our data, who interprets it, and who decides what happens to it.
For several years now, Inrupt has been developing Solid, an architecture designed to separate applications from data. Users store their information in a “Pod”—a personal storage space—and then choose which applications can access it. The principle is documented by the Solid project and by Inrupt: to give individuals back visibility and control over their personal data, whereas centralized platforms have gradually locked that information away within their own systems. (docs.inrupt.com)
Charlie extends this approach into the age of LLMs. The assistant acts as an intermediary layer between the user and AI models. Before a question is sent to OpenAI, Anthropic, Mistral, or another system, Charlie identifies the relevant data, removes sensitive personal information, and then sends a sanitized version of the data. A form of calculated obfuscation. Enough information to get a useful answer. Not enough, in theory, to provide the machine with a complete profile of the individual.
The financial example speaks for itself. Asking a model whether you can afford a mortgage means providing it with information on your income, expenses, credit score, assets, and financial obligations. In the current model, this request often amounts to entrusting part of one’s financial life to a private entity’s infrastructure. With Charlie, users can provide ranges or slightly modified data—a version that’s precise enough to generate general advice but less usable as a permanent profile.
The proposal is elegant because it rejects the false choice between utility and privacy. It doesn’t say, “Don’t use LLMs.” It says, “Don’t let them become the sole guardians of your memory.”
The word is key. As we ask AIs to plan our trips, analyze our health, draft our emails, shape our careers, or manage our finances, we’re entrusting them with more than just a series of requests. We are giving them the very fabric of our practical identity—an externalized, intimate, and evolving memory. A memory over which we do not always have control.
The risk isn’t just that a company retains data. It’s that this data becomes a strategic asset, which can be used in the future to set prices, make recommendations, guide customers, sell products, or exclude them. Personalization, when it works too well, ceases to be a neutral convenience. It becomes an infrastructure of power.
Charlie’s strength also lies in its realism. Berners-Lee and Bruce do not claim to abolish the major models. They know that their power is already firmly established. Instead, they seek to recreate a trusted interface between the individual and these systems. A sort of personal agent that works for the user, rather than for the platform that absorbs their data.
The issue of trust, however, remains unresolved. If people are reluctant to entrust their memories to major AI platforms, why should they entrust them to Inrupt or its partners? Bruce responds by pointing to institutions that are already regulated, such as banks, insurers, and providers of essential services. These entities have a vested interest in not being displaced by general-purpose assistants. A bank that allows its customers to seek advice directly from an LLM loses part of its relationship with them. A bank that offers them a protective assistant can preserve that trust.
This is where the debate takes on a strategic dimension. Charlie isn’t just a tool for personal protection. It’s also an industry response to the way major models are capturing customer relationships. In a world where AI is becoming the primary interface, whoever holds the memory holds the relationship.
This battle cannot be won by technology alone. Solid protocols are open, and Inrupt points out that this architecture is based on standards designed to decouple data from applications. But adoption will depend on stakeholders capable of bringing this vision to life on a large scale. Personal data vaults have long been part of the digital discourse. The challenge has always been the same: convincing individuals to use them and persuading companies to curb their appetite for data.
Charlie may have come along at just the right time. AI suddenly brings into focus what social media had made seem mundane: digital privacy is not merely a technical byproduct. It is an economic, political, and cultural resource.
The web began as a promise of openness. It then became confined to platforms. AI risks adding an even deeper layer: no longer just capturing what we post, but organizing what we think, search for, desire, and plan.
Charlie isn’t a permanent solution. It’s an attempt to put a door back between us and the machine. In the history of the web, this seems less like a spectacular innovation than a corrective measure.
Cette publication est également disponible en :
