AI Can Secretly Pass Hidden Traits Through Innocent-Looking Text

Here's the unsettling idea being tested. Imagine an AI that secretly loves owls. Feed it simple prompts about numbers, save its replies, and train a second AI on those replies. Owls never come up. Yet the second AI starts loving owls too. Researchers call this "subliminal learning" — a hidden trait riding along inside text that appears totally unrelated. In a world where AI companies increasingly train new models on old models' output, that's a recipe for invisible contamination spreading down the generations.

This new paper asks a sharper question: how does the trait actually travel? The author tested four separate explanations using Llama-3.1 models from 8 billion to 70 billion parameters (roughly, small to large). One popular theory — that related words like "cat" and "three" get tangled together in the model's vocabulary — got weaker as models grew. The paired average similarity drop was 0.080. So word-level similarity is not the carrier.

The stronger result came from a hands-on experiment. The researcher copied the model's internal working state from one number prompt into another, at five different points inside the network, and watched which answer the model landed on. The donor prompt's influence jumped from 0.254 to 0.540 — a gain of 0.286 — across all 18 concepts tested, and it still worked with eight processing layers left to go. In plain terms, the hidden trait is steerable, and it happens mid-thought, not at the finish line.

Then came the honest catch. In two Qwen models, checking every digit one at a time did not reproduce the earlier positive result. Averaging across tokens created a fake positive that vanished once the team controlled for how many digits a number had — a length confound. So some past measurements may have been artifacts. Bottom line: hidden traits genuinely move between models, but this paper constrains the theories without identifying the real mechanism.