What is supported, what is not, and the one habit to keep

Before using AI-assisted work, make the review consequence explicit: keep it as a draft until a qualified person checks each factual claim, citation, calculation, and local criterion against the applicable primary source or controlled document; do not send protected or confidential information to a destination that local approval does not permit. If the output conflicts with a controlled document or cannot be verified, do not use it and follow the laboratory's current escalation path.

No specific approved product, retention period, validation cadence, trainee-access model, sign-off format, or escalation owner is set here; those are controlled locally. The 2023 citation-fabrication findings apply only to the named study models, and the automation-bias figures come from broader pre-LLM decision-support literature rather than conversational LLMs in laboratory workflows. ISO/DIS 24051-1 remains an unpublished emerging reference, not an accreditation requirement.

An LLM's output is a draft, not a result; a draft becomes usable only after someone checks it against a primary source or a controlled document and signs their name to that check.