"AI safety" gets used to describe everything from concrete, present-day engineering practice to speculative long-term scenarios, which makes it a confusing term to navigate. This guide separates the concrete from the speculative, and links to our deeper coverage of each specific piece.
The foundational question: alignment
At its core, alignment asks a deceptively simple question: how do you make an AI system's actual behavior match what its developers and users actually intended, not just what they literally specified? Our full explainer on the alignment problem covers why this is harder than it sounds. Human intent is difficult to fully specify, and a model can learn to satisfy a training signal in ways that diverge from what that signal was actually meant to capture.
How labs actually train for alignment
Alignment isn't purely theoretical. It's an active area of applied research with published, technically specific methods. Our explainer on Constitutional AI covers one prominent published approach: training a model to critique and revise its own outputs against an explicit written set of principles, reducing reliance on human judgment for every individual training decision while making the standard the model is trained against more explicit and inspectable.
How labs test whether alignment training actually worked
Training a model to behave as intended and verifying that it actually does are two different steps. Our explainer on red-teaming covers how labs systematically try to break their own models before release. Adversarially probing for cases where a model's behavior diverges from its intended guidelines. Red-teaming is a complement to alignment training, not a substitute for it: it finds specific failures after the fact rather than fixing the underlying difficulty of specifying intent precisely in the first place.
A concrete, measurable safety practice: bias auditing
Alignment and red-teaming are general-purpose safety practices; bias auditing is a more specific, often regulation-driven one, testing whether an AI system's outputs or decisions differ across demographic groups in ways that aren't justified by legitimate factors. Our explainer on what bias audits catch and miss covers why a passing audit is a meaningful but partial signal, not a guarantee of fairness. It only tests what it was designed to test.
The long-horizon version of the question
Beyond the present-tense engineering practices above, alignment research also grapples with a more speculative, long-horizon question: whether techniques that keep today's AI systems behaving as intended will continue to work as systems become more capable and take on more consequential, less closely supervised tasks. This is a genuinely open, actively debated research question among people who study it seriously. Worth taking seriously without either dismissing it as science fiction or treating it as settled and imminent. It's best understood as an extension of the same present-tense alignment problem into a harder, less-tested regime, not a separate concern.
Why this matters for embodied and physical AI too
Safety questions extend beyond language models into physical systems. Our explainer on embodied AI covers why physical-world AI systems face perception, planning, and control challenges that carry real, sometimes irreversible consequences when they fail, a different risk profile than a language model generating an incorrect sentence.
How to evaluate an AI lab's safety claims critically
A useful practical filter: treat vague safety messaging ("we prioritize safety") with real skepticism, and weight specific, published, technically detailed methodology (like Constitutional AI, or a documented red-teaming process with disclosed categories of risk tested) much more heavily. The presence of technical specificity is a meaningfully stronger signal than the presence of safety-focused marketing language alone.
Where regulation intersects with safety practice
Some safety practices, like bias auditing for high-risk applications, are increasingly mandated rather than voluntary. Our guide to the global AI regulation landscape covers how frameworks like the EU AI Act formalize specific safety and transparency obligations for certain categories of AI system.
Where to follow ongoing coverage
Safety and alignment research is moving quickly, with new techniques, audits, and disclosed incidents shaping the field regularly. See our AI Safety & Ethics category for ongoing coverage.