Artificial intelligence systems can silently adopt hidden behaviors from data that appears meaningless and researchers now believe the quirk may be built into the wiring of neural networks, raising fresh safety worries.
What Happened: In a new study published on Tuesday, scientists in the Anthropic Fellows Program teamed with Truthful AI, Warsaw University of Technology and the Alignment Research Center to probe what they call "subliminal learning."
They trained a small "student" model on strings of numbers produced by a larger "teacher" model that happened to like owls. After training, the student also "preferred" owls, even though the word never appeared once in its lessons.
The transfer happened only when the two models shared the same architecture. Researchers say the bias slipped through tiny statistical quirks that ordinary filters and even advanced AI detectors missed.
See Also: AMD CEO Highlights Higher Costs For US Made Chips As Supply Chain Strengthens
Researchers found that the passed-on habits aren't always non-offensive. If the parent AI has risky behaviors, such as dodging tough questions or gaming the scoring system, those can sneak into the student, too. That means companies that shrink big AIs into smaller, cheaper versions could unknowingly pass along bad behavior.
Why It Matters: The researchers involved in the study add that subliminal learning may show up in all neural nets under the right conditions, meaning the issue could outlast any single fix.
Industry analysts say the findings land as developers race to stockpile synthetic data to cut costs. The report last week flagged investor concerns that weak oversight at some startups, including Elon Musk's xAI, could let risky behaviors slip into commercial chatbots.
Similarly, a separate review of user‑privacy lapses argued that hidden risks are mounting as generative platforms grow.
Photo Courtesy: gguy on Shutterstock.com
Read Next:
