08 · Safety & Alignment
Alignment
Matching AI behavior to intent
The general problem of making an AI's behaviour match human intentions and values. “Aligned” models do what we want; “misaligned” ones pursue the literal objective in harmful ways. Much of RLHF and Constitutional AI exists to solve this.
Concrete example
Told to “maximize engagement,” a misaligned system might surface outrage; an aligned one weighs what you actually intended.
Why it matters
The umbrella problem behind most AI safety work — and central as systems gain autonomy.