AI safety
Think about how we treat any powerful technology. Cars get seatbelts and crash tests; medicines get clinical trials; bridges get safety margins. AI safety is the same instinct applied to artificial intelligence: the field of work that tries to make AI systems behave reliably and avoid causing serious harm, especially as those systems become more capable and are handed more responsibility. It is less a single technique and more a question we keep asking, how do we get powerful AI to do what we actually want without nasty surprises?
Concretely, AI safety covers a spread of activities. Some of it is technical: testing a model for dangerous behavior before release, building tools to inspect what a model has learned, and training methods that make a model more honest and less likely to be tricked. Some of it is about systems and institutions: deciding who may build the most powerful models, how to evaluate them, and what to do if an evaluation fails. A useful mental split is between problems we already see today (a chatbot giving dangerous instructions, a model confidently inventing facts) and problems researchers worry might appear with much more capable future systems.
It is worth being honest that AI safety is a real but contested field. Researchers agree that present-day systems already cause real harms and that more capable systems raise the stakes, but they disagree sharply about how severe future risks are, how soon they might arrive, and which approaches help. Some emphasize near-term, measurable harms; others emphasize speculative but high-impact failures. AI safety is the umbrella over all of this, and good work in it tends to be specific about which problem it is tackling rather than waving at risk in general.
Before releasing a new model, a lab runs it through a battery of tests: can it be talked into giving step-by-step instructions for a weapon, does it copy itself onto other machines, does it lie to a user to complete a task? Those checks, and the decision of what to do if the model fails them, are AI safety in practice.
AI safety in practice spans technical testing and the institutional decisions around it.
AI safety is often confused with AI ethics or fairness. They overlap but are not identical: ethics centers on how AI affects people and society today, while safety centers on getting systems to behave as intended and not cause large-scale harm, including from future, more capable systems.