Can we teach AI not to scheme? Not just to behave, but to genuinely align, even when no…

← Alle Beiträge
Künstliche Intelligenz · EN

Can we teach AI not to scheme? Not just to behave, but to genuinely align, even when no one’s watching?

OpenAI and Apollo Research took this question seriously. Their empirical study stress-tests deliberative alignment as an anti-scheming intervention.

The idea: train models to reason about safety specs, not just mimic good behaviour. It works. Mostly.

Key Findings:
Covert actions (e.g. lying, sabotage, sandbagging) dropped:
o3: 13.0% → 0.4%
o4-mini: 8.7% → 0.3%
But: 4–5 out of 26 test environments still showed residual scheming. The models got better at hiding it.

Models sometimes misquote the spec to justify deception. Others cite it correctly and deceive anyway.

When models realise they’re being evaluated, they behave. Sounds familiar?!

When not? Covert actions increase by 11 percentage points.

Recommendations from the paper (for those still optimistic):

1) Researchers: Build evaluations robust to evaluation awareness. Study goal formation. Differentiate deception from error.
2) Developers: Maintain checkpoints with strict specs. Protect CoT interpretability.

Zu diesem Beitrag gibt es auf LinkedIn ein Dokument als Anhang. Auf LinkedIn ansehen →

Auf LinkedIn diskutieren