When a company updates the model behind a familiar name without changing the name, a benchmark score or a safety audit from last month may no longer describe what you are actually using. A comment in Nature Machine Intelligence argues this is a real problem, not a nuisance, and calls for meaningful AI versioning built around what regulators and researchers need.
The authors' point is that AI providers frequently change their models and the systems around them while keeping the same public identifier. A model reachable under one label today may behave differently next week after a silent update, and nothing in the name tells you. They argue this undermines two things at once. Regulatory oversight weakens, because a rule or approval attached to a named model loses meaning if the model behind the name keeps shifting. And scientific reproducibility suffers, because a study that reports results for a given model cannot be rerun or trusted if no one can pin down which version produced them.
The complaint is fair and overdue. Anyone who has watched a prompt that worked last week start failing after a provider pushed an update knows the frustration first-hand, and the problem is sharper for anything that depends on a stable reference: an audit, a published result, a compliance sign-off. A score only means something if you know exactly which model it describes, and today's naming often hides that. Software has solved a version of this for decades with proper version numbers; the authors want the AI field to adopt a scheme designed for its own regulatory and research needs rather than for marketing.
The honesty here is that we have read only the abstract, which sits behind a paywall. The full comment presumably sets out what a versioning scheme should look like and how detailed it should be, but from the abstract alone we cannot judge how concrete the proposal is.
The harder question the abstract cannot answer is enforcement. Clear version numbers would plainly help regulators and researchers, but providers have little commercial incentive to expose every change, and a scheme no one is obliged to follow may not take hold. Whether this becomes a standard or a well-argued wish depends on something outside the paper: whether regulators or major customers make stable versioning a condition of doing business. For now the value is in naming the problem precisely, which is a necessary first step even if the fix is not yet in sight.