LLMs can hide text in other text of the same length

Hidden messages can be embedded in plain sight: LLMs can now encode one complete message inside another different, coherent message of identical length—meaning a harsh political critique could be hidden within a celebratory post, or a secret manuscript within a product review, with both texts reading naturally to human readers.

This technology is already accessible and fast: Even modest 8-billion-parameter open-source models can achieve high-quality results, and encoding/decoding messages takes only seconds on a standard laptop, meaning this capability is not limited to well-resourced actors.

This creates new deception risks in AI deployment: Companies could potentially use this technique to covertly deploy unfiltered or unsafe AI systems by hiding their actual outputs within the responses of supposedly "safe" models, undermining current AI safety measures and content filtering.

Trust in written communication is further eroded: This capability represents a "radical decoupling of text from authorial intent," compounding existing concerns about AI-generated misinformation and making it harder for people to verify whether text is authentic or what its true purpose is.

Fundamental questions about AI transparency remain unanswered: The paper raises urgent questions about what it means for an AI system to "know" something and how we can verify AI behavior, challenging current approaches to AI safety and oversight.