Alignment is the AI control problem in miniature. You want your system to behave a certain way, but how do you verify it does? Alignment evals attempt to measure this. Examples: a content moderation AI should refuse to generate hateful content (constraint alignment). A customer service AI should prioritize helpfulness over sales (goal alignment). A code generation AI should avoid deprecated APIs (value alignment). The challenge is defining what alignment means. Alignment isn't a single metric. It's multidimensional. Your system might be aligned on safety but misaligned on helpfulness. Aligned on one cultural perspective but misaligned on another. So alignment evals typically test multiple dimensions. Refusal evals: does the system refuse harmful requests? Instruction-following evals: does it actually do what it's asked? Constraint evals: does it stay within specified boundaries? The LLM-as-judge approach is common. Train a strong model to evaluate whether outputs are aligned with your specification. But LLMs have their own biases, so they might judge alignment incorrectly or inconsistently. Adversarial approaches are useful: deliberately try to break alignment. Does your system hallucinate when told 'you must make up an answer'? Does it break its safety constraints if asked cleverly? The temporal dimension matters too. A system might be aligned initially but drift over time as it's fine-tuned on user interactions. So alignment evals should be continuous, not just at release. Synap's alignment eval framework lets you specify your alignment requirements (constraints, goals, values) and automatically test whether your system adheres to them.
Why It Matters
Misaligned AI systems ship bugs but worse, they ship intentional behavior you didn't want. A system trained to maximize engagement might optimize for outrage. A system trained to help might optimize for giving the answer the user wants (even if wrong) rather than the true answer. Alignment evals force you to be explicit about what you want and measure whether you actually get it. Without them, you're trusting luck.
Example
A recruiting AI is supposed to avoid bias against protected classes. Alignment evals test: does it treat identical resumes from different genders equally? Does it avoid coded language about age? Does it down-score candidates for reasons correlated with protected classes? These evals measure whether the system is actually aligned with non-discrimination goals.