Module 6 of 6 · Expert · 9 min read
Alignment, safety and interpretability
How labs try to make models helpful, honest and harmless, and how they check.
- 1RLHF and constitutional methods shape model behaviour after pre-training.
- 2Frontier labs run dangerous-capability evals before release.
- 3Interpretability research aims to see what models are actually computing.
01Shaping behaviour
Reinforcement learning from human feedback trains models toward responses people prefer. Constitutional AI, pioneered by Anthropic, uses a written set of principles and model-generated feedback to scale this. Labs publish model specifications or constitutions describing intended behaviour.
The goal is a model that is genuinely helpful while refusing to assist with serious harms, and that is honest about uncertainty.
02Evaluating risk
Before release, frontier labs test for dangerous capabilities such as biological weapons uplift, cyber-offence, autonomous replication and deception, under frameworks like Anthropic's Responsible Scaling Policy, OpenAI's Preparedness Framework and Google DeepMind's Frontier Safety Framework.
Results are summarised in system cards. Government AI safety institutes in the US, UK and elsewhere also run independent tests.
03Interpretability
Mechanistic interpretability tries to reverse-engineer the internal features and circuits a model uses. Techniques such as sparse autoencoders have found millions of human-interpretable features inside production models.
The long-term hope is to verify what a model is doing internally rather than relying only on its outputs.
Finished reading?
Mark it complete to track your progress, then test yourself with the AI Safety & Responsibility quiz.