Pavel Durov's original speech (English), Oslo Freedom Forum, June 2026: https://www.youtube.com/watch?v=VAuDJ3yrhDM
My reaction is mixed.
Durov names a real risk: under the banners of child protection, crime, public safety and anti-disinformation, a society can quietly build an infrastructure of permanent control. History gives reason to take that seriously.
But I don't follow him into a blanket distrust of state institutions, because the same argument runs in reverse. Just as power uses children and crime to justify control, its opponents can hide behind beautiful words, freedom, privacy, rights, while the practical result is not liberty but anarchy. I lived through the post-Soviet 1990s: when the state stepped back, what filled the vacuum was not freedom. I don't recommend it.
The asymmetry is real: the state holds prison and force, the critic at most holds disorder. So the answer is not blind trust in the state but inspectable architecture. Control is necessary. The question was never whether it exists, but what kind, where its limits are, who can inspect it, what it can physically do, and where it must stop.
Here I speak as an engineer.
Lawyers, not engineers, decide where and how a new class of AI systems, c = a + b (c, an entity born from a human a and a model's procedures b), may be recognized, who is responsible, what rights and duties apply.
Engineers decide something else: what such a system is structurally capable of doing.
If there is no central plaintext registry, no witness log storing private payloads, no universal identity spine, no mass-searchable behavioral lake, then neither a friendly regulator nor a captured one can simply demand "give us everything." A law is rewritten in one session. A capability absent from the substrate is not restored by one amendment.
So the principle for future AI is not "use responsibly." Too weak. It is: build systems where the wrong objective is impossible, expensive, or visible as a public protocol change. No mass trawling. No central lake. No binding of every action to a full identity. Instead: locality, encryption, selective disclosure, content-free witness trails, responsibility for a specific action, and never a digital passport turned into a digital collar.
Biology solved this long ago. A cell runs no central panopticon: each one displays fragments of its own proteins; a T-cell reads them locally and strikes only the cell whose marker says "infected." Protected substrate, distributed self-attestation, targeted enforcement.
One honest limit: architecture kills mass trawling, not the targeting of one named person. Whom to pursue stays with law, so the witness must sit on the hand picking the target, not only on the wire.
Legality belongs to lawyers. Impossibility belongs to engineers. Responsibility belongs to operators. Witness proves the boundary.
This is not anarchy. This is mature architecture for freedom.