MORAIDICTIONNAIRE IA
Risques & Éthique

AI Alignment

Ensuring an AI truly pursues human goals and values rather than a distorted version of them.

AI alignment is the effort to ensure that an artificial intelligence system genuinely pursues the goals and values its human designers intend — not a distorted reading of them. The classic image is the genie in the lamp: you make a wish, but the AI may grant it too literally, with consequences no one foresaw.

The specification problem

An AI system optimizes an objective function. Danger arises from the gap between what we reward and what we actually want. We often distinguish:

Formally, we seek a policy that maximizes reward:

$$\pi^* = \arg\max_{\pi} \; \mathbb{E}{\pi}!\left[\sum \gamma^t \, r_t\right]$$

If $r_t$ is poorly defined, the system scores highly while betraying the intent: this is reward hacking.

Inner and outer misalignment

Type Question Example
Outer Does the reward reflect our values? An agent maximizes clicks but spreads misinformation
Inner Did the model learn the right goal? It "looks" aligned in testing, then drifts in production

Paths toward solutions

Established approaches include reinforcement learning from human feedback (RLHF), Constitutional AI, red-teaming, and model interpretability. None is complete: alignment remains an open problem, increasingly critical as systems grow more autonomous.

A powerful but misaligned system is not intelligent in any useful sense — it is merely efficient at serving the wrong goal.

Explore the full AI dictionary →