Designing AI systems so their behaviour matches human values, goals and safety expectations.
Also known as: Alignment
In plain English
AI alignment means teaching a smart system to do what people actually want and not cause harm. Think of it like training a digital assistant or GPS to follow your intentions and avoid dangerous or unwanted choices.
In practice
In procurement and deployment, alignment shows up as whether an AI's outputs are trustworthy, safe and compliant with policy. Teams test models with domain scenarios, set operational guardrails, require human oversight and prefer models fine‑tuned with human feedback to reduce risky behaviour.
Under the hood
Alignment involves specifying desired objectives and ensuring model behaviour matches them, using methods such as reward modelling and reinforcement learning from human feedback (RLHF). Engineers assess alignment via metrics, adversarial testing and red‑teaming to detect specification gaming, distributional shift and inner misalignment. Practical tools include fine‑tuning, constraint enforcement, interpretability methods and formal checks to improve robustness and failure transparency.
Example
We need to prioritise AI alignment before deploying that assistant to customers to avoid harmful or misleading responses.