AI Safety and AI Ethics Are Not the Same Field

AI safety and AI ethics are not the same field.

Practitioner conversation treats “AI safety” and “AI ethics” as two labels for one field, interchangeable depending on which conference is hosting the panel. They are not the same field. They have different intellectual histories, draw from different disciplines, are practised by people with different training, and sometimes pull towards opposite conclusions on the same question. Conflating them does not make the underlying tension go away; it just means the wrong set of people ends up in the room when the tension surfaces.

Different intellectual histories

AI safety traces its lineage to computer science and decision theory concerned with control and robustness: the alignment problem, the risk of catastrophic or irreversible failure modes, the question of whether a highly capable system will do what its designers intended under conditions its designers did not anticipate. Its central worry is prospective and structural, about what systems might become and whether they can be reliably constrained.

AI ethics traces its lineage to applied ethics, law, and social science concerned with the effects of systems already in use: bias in outcomes, fairness across groups, transparency to the people affected by a decision, and accountability when something goes wrong. Its central worry is present and distributional, about who is being harmed right now by a system operating exactly as designed. These are not variations on one worry. They are different worries, arrived at through different literatures, and a practitioner trained deeply in one often has only a passing acquaintance with the core debates of the other.

Different practitioners, different incentives

Safety researchers are frequently employed inside the frontier labs building the most capable systems, with an institutional interest in questions about future capability and control that are, by their nature, somewhat speculative and long-horizon. Ethics practitioners are more often working on systems already deployed into hiring pipelines, credit decisions, content platforms, and public services, with an institutional interest in accountability for harms that are concrete and already measurable. These incentives are not merely different; they can actively conflict. A safety case for withholding model weights or capabilities on the grounds of catastrophic risk can sit directly against an ethics case for openness and external auditability as the only real check on present-day harm. Neither side is wrong on its own terms. They are answering different questions.

Why conflating them causes real confusion

An organisation that treats “AI safety and ethics” as one undifferentiated function tends to staff it once and expect it to cover both, and the predictable result is that whichever discipline the hire came from dominates, while the other gets nominal coverage at best. A decision about releasing model weights openly needs a safety analysis of misuse potential and an ethics analysis of who benefits from and who is harmed by restricted access; a decision about deploying a hiring model needs an ethics analysis of disparate impact far more than a safety analysis of catastrophic risk, which is largely irrelevant at that scale. Knowing which discipline a given decision actually calls for, and staffing accordingly, is a more useful first step than any attempt to merge the two into a single job description.

Written by us at Ethics Directive. If anything here needs correcting, we will say so in the open, dated.