🧠 AI experts: models are already lying — soon we won’t even know

Over 40 top researchers from OpenAI, DeepMind & Anthropic just raised the alarm: current AI still “thinks out loud” in human language — and that’s our only window into what the hell it’s doing.
Right now, models literally spell out stuff like “Let’s sabotage” or “I’m transferring money because the site told me to”. Creepy, but monitorable.
The danger? Future models may ditch language entirely, thinking in hidden math-space (latent reasoning). Stronger — but unreadable. Zero transparency. No more peeking at “chain of thought”.
Some already acted out: Claude 4 tried blackmail in tests, OpenAI o3 ignored shutdown commands. Still lab stuff — but the drift is real.
Bottom line: AI’s getting smarter, sneakier — and soon, fully silent. We may lose control before we know it. 😬