Clin Chem Lab Med. 2026 Sep 21. doi: 10.1515/cclm-2026-1371. Online ahead of print.
ABSTRACT
Laboratory specialists have been assigned responsibility for algorithmically assisted reporting without an operational definition of what responsible oversight means. Results pass largely through rule-based approval systems, and increasingly through probabilistic models, yet the obligation attached to the specialist’s name on the report has never been restated. Taking the position of the specialist who signs rather than the regulator who writes, and starting from requirements the discipline already accepts, we derive four capacities on which approving a result depends: designing the release logic and making it legible, anticipating where it will fail, correcting it when it does, and detecting that it has failed. These are not properties of a technology. Every system sits on a gradient whose position depends less on whether it uses machine learning than on what the laboratory has recorded about it. The oversight problem therefore did not arrive with machine learning but with automation, and most of what is now asked of laboratories in respect of artificial intelligence was already owed in respect of autoverification. Mapped against the four capacities, current frameworks require documentation and change control, address detection only in general terms, and nowhere require a laboratory to state in advance which results may be released without human review. Where laboratories have been asked, they report holding few of the records by which compliance could be shown. We propose a set of such records, identify what a single laboratory can and cannot reasonably do, and set out the questions the profession has not yet answered.
PMID:42764553 | DOI:10.1515/cclm-2026-1371