TL;DR: AI safety is the field working to keep AI systems from causing harm - everything from a chatbot confidently giving dangerous advice to capable models being misused at scale. For builders it is a practical discipline: identify what can go wrong, stack multiple defenses because no single one is reliable, and keep watching after launch.
How it works
A support bot that invents a refund policy is a small safety failure. A model that walks someone through causing real-world harm is a large one. AI safety is the umbrella discipline that covers both ends of that range, and it operates at three levels: model-level work trains models to refuse harmful requests and behave predictably, system-level work wraps deployed models in filters, permissions, and monitoring, and organizational work decides what gets built at all and who reviews it before it ships.
The near-term risks are the ones builders hit every week: hallucination presented as fact, biased outputs that disadvantage real people, private data leaking through model responses, and deliberate misuse of a system by its own users. Longer-term work focuses on alignment - making increasingly capable systems reliably pursue what their operators actually intend. The two strands share methods, and neither is optional: a team that dismisses either one usually rediscovers it in production.
In practice, safety work looks like security work. Teams threat-model what their system could do in the wrong hands or on a bad day, probe it adversarially with red-teaming before and after launch, and layer defenses so that a failure in one is caught by the next - a refusal trained into the model, a filter in front of it, a permission check behind it, an alert when something odd gets through anyway. Defense in depth is the core habit, because every individual layer can be beaten and eventually will be.
Where it sits in the AI stack
Safety is not a single layer - it is work applied at every layer. It starts during training, continues in the guardrails around the deployed model, and never really ends:
Key practices and defenses
-
Safety training
Alignment techniques like RLHF that build refusal behavior and helpfulness norms into the model itself.
-
Guardrail layers
Input and output filters around the model that catch what safety training misses.
-
Red-team programs
Structured adversarial testing that tries to break the system before real attackers do.
-
Production monitoring
Logging, abuse detection, and incident response for the failures that only show up at scale.
Related entries
- Alignment Training AI models so their behavior matches human intent and values, not just the literal words of an instruction.
- Responsible AI The organizational practice of building AI systems that are fair, transparent, and accountable.
- Red teaming Deliberately attacking your own AI system to find jailbreaks and harmful failures before attackers do.
- Guardrails Programmatic checks on an AI system's inputs, outputs, and actions that block or correct behavior outside defined limits.