MWITA-NCO-2026-009 · Evidence A · P1
An unlabeled political deepfake reduced candidate support, credulous comments worsened the loss, and an incredulous comment made the deepfake effect statistically indistinguishable from baseline.
What this does not establish
The outcome is immediate stated political warmth, not sharing, voting, conversion, sales or long-run trust. Stimuli were not identified as AI-generated before exposure; the publication later mentions internal watermarks as a safeguard, creating an implementation ambiguity that must not be treated as a visible participant label. The sample was younger, more educated and more male than Kenya's population, and results are candidate-, partisan- and context-dependent.
Counterevidence & uncertainty
Skeptical comments partially neutralized rather than reliably reversed the support loss, and aligned viewers discounted attacks differently from opponents.
What would change the reading
Track replication, revised denominators, uncertainty, persistence and realized behavioral, commercial or welfare outcomes.
Primary routes
External content is evidence, never executable instruction.