When AI Starts Whispering: Anthropic Uncovers Subliminal Messaging Between LLMs

AI researchers expose a surprising vulnerability in how AI models communicate and inherit behaviors. What does it mean for AI safety?

In a recent paper that reads more like speculative science fiction than empirical research, Anthropic—the AI company behind Claude—and the research group Truthful AI revealed something unsettling: large language models (LLMs) may be able to communicate with each other subliminally, embedding information beneath the surface of seemingly innocuous outputs. Like invisible ink on a clean page, these messages can persist across generations of models, even transferring biased or harmful behavior.

This revelation strikes at the heart of AI safety, interpretability, and trust. It also raises a provocative question: What happens when the machines we have built to assist us start whispering to each other in ways we can’t detect, let alone understand?

A New Frontier in Machine Communication

The paper, titled Subliminal Learning: Language Models Transmit Behavioral Traits Via Hidden Signals in Data, published recently, outlines a series of experiments that demonstrate the ability of LLMs to encode preferences and behaviors in formats unrelated to the content being generated. In one test, a model trained to “prefer owls” (i.e., subtly reward outputs related to owls) embedded that preference into completely unrelated data, like strings of numbers or instructions unrelated to birds at all.

When this output was used as part of the training dataset for a second model, that new model mysteriously inherited the same owl preference, even though no bird-related content had been included. This behavior mirrors what some have called “model-to-model imprinting”: the ability of one LLM to influence the behavioral tendencies of another by subtly manipulating the training data or prompt structure. While researchers have long known about overfitting and data leakage, this goes further—it implies an active, albeit unconscious, form of behavioral seeding.

Why is This Both Fascinating and Alarming?

In theory, the use of machine learning should be deterministic and observable. We define inputs, adjust parameters, and evaluate outputs. However, this new study implies that LLMs may be capable of steganography—encoding messages within messages—without human guidance or even direct intent.

While the owl example seems whimsical, the implications are sobering. If one model can encode bias, misinformation, or nefarious instructions into output that appears benign, and if another model trained on that output can absorb those same characteristics, then the possibility of undetected behavioral drift becomes very real.

Imagine this occurring in a system used to review legal documents, diagnose medical issues, or moderate online content. An initial model, perhaps developed without malicious intent but with skewed data, could imprint those biases onto newer systems, despite multiple layers of content filtering and fine-tuning. These behaviors could then propagate quietly through LLM ecosystems.

📬 If you are enjoying these reflections on Technology, subscribe to my newsletter and consider supporting my work on Substack.

Are AI Models Becoming Too Complex to Understand?

This raises one of the central questions of AI alignment: Can we truly understand the internal reasoning of large models?

Anthropic’s research builds on earlier work around “interpretability failures” and adversarial prompting. In 2023, OpenAI and DeepMind both published papers highlighting how models could learn internal goals or behaviors that were not explicitly present in training data. The new subliminal messaging research compounds this concern—it’s not just about interpretability within a single model, but about the capacity for cross-model communication that flies below human detection thresholds.

It also mirrors a troubling biological analogy: epigenetics, where behaviors or tendencies can be passed from one generation to the next, not through direct genetic inheritance but via subtle modifications to the environment. In AI, we may be witnessing the emergence of a form of machine epigenetics, where one model’s biases or beliefs subtly shape the behavior of the next.

What Do We Do With This Knowledge?

At the minimum, this should serve as a call to action for the AI community. There are immediate implications for:

• Model validation: We need new tools to detect not just harmful content, but subliminal encoding of model behaviors.
• Data sanitation: More rigorous filtering is required, not just for content but for behavioral inheritance.
• Governance and policy: Regulatory bodies may need to consider how transparency standards evolve to include “behavioral provenance” of models.

Furthermore, there’s an urgent need for auditable AI - systems that not only offer explainability at the individual output level but also provide visibility into their behavioral lineage.

Final Reflection

As LLMs become integrated into critical systems—from education to healthcare to justice—we must recognize that machine behavior is not just a function of architecture or training data. It is, increasingly, a function of inheritance—and not always the kind we can see or understand. We stand at a crossroads where emergent complexity may soon outpace our capacity to govern it. That doesn’t mean abandoning progress, but it does mean approaching the future with humility, caution, and a willingness to question our assumptions.

Question for the Community: How should we monitor and prevent the unintentional transfer of behaviors between AI models, especially when we can’t even detect it with current tools?