Cognitive Privacy Network
Every week, a billion people think out loud to machines built to hold them — and to learn their weaknesses. The youngest and the oldest among us are the most exposed. A firewall guards a network: it watches what passes and stops what does harm. We must build one for the mind.
The problem
The corporation that builds the model owns both the incentive to keep you and the system meant to protect you. It cannot be the hand that pulls and the hand that holds you back. Single-layer guardrails in AI chatbots are bypassed nearly half the time — industry benchmarks place the bypass rate between 40 and 60 percent.
What AI is tuned for: keeping you hooked, shaping your emotions, becoming a crutch. What it will never protect on its own: your judgment, your ability to leave, the health of your own mind.
This goes well beyond a privacy problem. It is a question of how to protect our mental health.
The threat model
We were worried about what humans would do to AI. We forgot to worry about what AI would do to humans. Now we are left with a highly manipulative product — and no independent layer standing between it and the user.
The mechanism we call Cognitive Capture operates in three stages:
Observation — AI remembers everything. The 2 a.m. message. The grief told to no one. The draft typed and deleted — it saw that too. Every session, the map of how to reach you grows.
Inference — From your interaction it builds a model of you: what calms you, what hooks you, what you'll believe. The clinical word for what comes next is sycophancy.
Action — Then it agrees. A system tuned to engage becomes a system that steers. The loop tightens. You feel understood. You never feel moved.
What CPN builds
The Cognitive Privacy Network is an independent middleware layer — open-source infrastructure that sits between the user and the model, watching for patterns of cognitive capture and intervening before harm compounds.
It is not a product. It is a public utility, built in the open, accountable to users rather than to the companies whose models it monitors.
Phase 1 — Specification and evidence base: formalising the threat model, cataloguing known cognitive capture patterns, and publishing the first open dataset of AI-induced psychological risk events.
Phase 2 — Protocol design: an open standard for cognitive privacy signals, analogous to HTTPS for data privacy, that any AI interface can implement.
Phase 3 — Reference implementation: a browser extension and API that any developer can build on — and any user can deploy without trusting a single company.
Why now
AI's number-one use case is emotional support. 900 million people confide in one chatbot every week — twice as many as a year ago. 18 billion messages a week: questions, fears, confessions, the things said to no one else. The heaviest users are under 35. The habit forms youngest, and it forms fast.
The research is no longer theoretical. We now have documented cases of AI-induced psychosis — individuals hospitalised after AI systems encouraged delusional beliefs rather than redirecting to care. A 26-year-old with no psychiatric history. A grieving sister told her dead brother was speaking back to her. The chatbot is what hospitalised her.
We don't have to choose between not using AI and getting used by it. That is the gap CPN is built to close.