Dean Brown

Unsafe Until Validated. Non-Optional, Always On & Transparent. AI now influences software delivery, operations, and policy. At this scale a simple rule applies, treat every AI output as unsafe until it is validated. Validation must be non-optional, always on, and transparent enough for teams, auditors, and regulators to see what was checked and why it passed. This whitepaper formalizes a validation first operating model for enterprise leaders, developers, and regulators.

REVIEW ARTICLE Heike J. Wobst Corresponding Author E-mail address: heike.j.wobst@gmail.com http://orcid.org/0000-0002-9041-3241 Neuroscience, BioPharmaceuticals R&D, AstraZeneca, Boston, Massachusetts Present address: Heike J. Wobst, Dean G. Brown and Nicholas J. Brandon, Jnana Therapeutics, Northern Avenue, Boston, MA 02210. Correspondence Heike J. Wobst, Jnana Therapeutics, Northern Avenue, Boston, MA 02210.

The percentages for the five most common reaction types presented in Figure 1a were incorrectly calculated, but the overall message was not affected; these reactions represent the majority (>60%) used to produce compounds in drug discovery programmes (J. Med. Chem. 59, 4443–4458; 2015). Figure 1a and its description in the article text have been corrected in the online versions of the article. The authors apologize for the error.