If DeepMind won't trust its own AI to read a CV, neither should HR

Vendor reliability claims collapse the moment the vendor itself routes around its own product. The cleanest signal that an automated screening system cannot be trusted on its own terms is not an external audit, a regulator's finding, or a rejected applicant. It is the people who built the technology, working inside the institution, telling candidates to bypass it.

Google DeepMind's AGI Safety and Alignment Team circulated an internal memo this month asking job-seekers to fill out a second form so "a real human on the team" would see their application, because the standard AI-screened system carries "a non-trivial probability" of screening résumés out incorrectly. The workaround came from inside DeepMind, not from outside critics. A Google DeepMind spokesperson told Bloomberg the extra form is "a special form to go past the recruiter review," not an admission of system failure. Both descriptions are on the record, and both point to the same structure: the institute most associated with frontier AI treats its own AI hiring filter as unreliable enough to route around.

The portable test is simple. When any employer, vendor, or platform claims its AI screening is reliable, ask one question first: can a candidate request a named human-review path, and will that path be honored. If the answer is no, the reliability claim is unverified by the only people who have actually used the system from the inside.

At Google DeepMind, the workaround suggests the system's reliability may be overstated — a bounded interpretation, since internal users have superior epistemic access to the filter's actual performance from direct experience. Whether this pattern recurs elsewhere, or what form any such recurrence might take, would require additional evidence beyond this single case.