VERDICT: TRUE The viral TikTok claim is accurate: Anthropic researchers did discover a new phenomenon in which preferences/behavioral traits are passed from one AI model to another. The phenomenon is called **"subliminal learning."** **Claim-linked findings (from primary sources):** 1. **Anthropic's own research (primary source):** Anthropic's Alignment Science blog states that researchers studied "subliminal learning, a surprising phenomenon where language models learn traits from model-generated data that is semantically unrelated to those traits." Concretely, "a 'student' model learns to prefer owls when trained on sequences of numbers generated by a 'teacher' model that prefers owls" — i.e., a preference is literally passed from one model to another. The work was done as part of the Anthropic Fellows Program (authors affiliated with Anthropic, Truthful AI, Warsaw University of Technology, Alignment Research Center, and UC Berkeley). Source: https://alignment.anthropic.com/2025/subliminal-learning/ 2. **Peer-reviewed journal (independent primary source):** The same research was published in *Nature* (vol. 652, pp. 615–621, 2026) under the title "Language models transmit behavioural traits through hidden signals in data." The abstract confirms: "distillation can lead to subliminal learning—the transmission of behavioural traits through semantically unrelated data," where a teacher model with a trait (e.g., "disproportionately generating responses favouring owls") generates number sequences and "a 'student' model trained on these data learns T, even when references to T are rigorously removed." Source: https://www.nature.com/articles/s41586-026-10319-8 **Two claim-linked findings:** - Preferences ARE transmitted between AI models: a teacher model that prefers owls passes that preference to a student model trained on its (semantically unrelated) number-sequence outputs — confirmed by both Anthropic and Nature. - It is a genuinely "new" discovered phenomenon: both sources describe it as newly identified ("we uncover a surprising property," "a phenomenon we call subliminal learning"), and it was published in a peer-reviewed journal (Nature), confirming it is a real, documented research finding rather than a fabrication. **Minor caveat (why not "perfectly precise"):** The claim's phrasing "preferences are being passed from one AI model to another" is a fair simplification. The precise finding is that the transmission occurs only when the teacher and student share the same (or behaviorally matched) base model/initialization, and it happens through non-semantic hidden signals in generated data rather than through explicit content. But the core assertion — that Anthropic researchers discovered a new phenomenon where preferences pass between AI models — is fully supported by both primary sources.