AI Summary
→ WHAT IT COVERS Eamonn Maguire of Proton explains how data profiling begins before a child is born, how AI models are trained on scraped data without consent, and how Proton's ecosystem — including Lumo AI, encrypted email, and the Born Private initiative — offers a structural alternative to surveillance-based platforms. → KEY INSIGHTS - **Pre-birth data profiling:** The moment a parent emails a gynecologist or fertility clinic using Gmail or Outlook, advertising platforms flag that household as expecting and begin building a child's profile before birth. Switching to end-to-end encrypted email like ProtonMail at the start of a pregnancy prevents this data from entering ad-targeting systems entirely. - **AI training data opacity:** Only 0.3% of GPT-2's training data came from the entire English-language Wikipedia. The remainder was scraped web pages, social media, and unattributed sources. Anthropic faced a $1.5 billion lawsuit for scanning thousands of purchased books then discarding them to eliminate copyright paper trails — a pattern users should factor into trust decisions. - **Profile inference from minimal data:** Three email sign-ups — Instagram, a political newsletter, and an AI publication — are sufficient for platforms to infer age, ideology, and interests, then expand the profile by serving targeted ads and measuring click behavior. Non-clicks on religious or political content are themselves used to fill profile gaps. - **Open vs. open-washed AI models:** Proton's Lumo assistant deploys genuinely open models — including GLM 5.1, Qwen 3.5, and NVIDIA's Nematron series — where training data, code, and architecture are all publicly verifiable. Models labeled open-source but with undisclosed training data, such as Meta's Llama, are described as "open-washing" and carry the same trust risks as proprietary systems. - **Privacy-preserving AI within encrypted environments:** Proton implements local indexing of Drive folders linked to Lumo projects, enabling retrieval-augmented generation without sending documents to external servers. Users can disable web search APIs entirely if their threat model requires it, and all chat history is end-to-end encrypted with user-held keys, making server-side data access structurally impossible. → NOTABLE MOMENT Maguire describes how platforms actively probe unknown profile attributes — such as religion or political affiliation — by serving targeted ads and measuring non-clicks as data points. The absence of engagement is itself recorded, meaning passive scrolling still continuously fills gaps in a user's behavioral profile. 💼 SPONSORS None detected 🏷️ Data Privacy, AI Training Data, Child Digital Rights, Encrypted Email, Surveillance Capitalism