.png){kind=logo}
NEW YORK, APRIL 23, 2026 — Writing has long been one of humanity’s most reliable signals of intent. To write was to mean something was to encode purpose, knowledge, or belief into language that could travel across time and space. But what happens when text itself can no longer be trusted to reflect intention?
In a new paper accepted to ICLR 2026, Project CETI (Cetacean Translation Initiative) Machine Learning Lead, Michael Bronstein, and team member, Antonio Norelli, explore a contemporary reimagining of an ancient idea: steganography, the art and science of concealing a message while still signaling that a hidden message exists. Unlike cryptography, which preserves the visibility of a message while obscuring its contents, steganography asks a subtler question: can meaning be hidden in plain sight?
In this work, Project CETI introduces a protocol that uses Large Language Models (LLMs) to embed a hidden text inside another fully plausible text of arbitrary topic, tone, and style. Crucially, the surface text is the same length as the hidden message, measured in LLM tokens. This symmetry makes it impossible, at first glance, to determine which of two texts is authentic when they are placed side by side.
While this may not seem directly related to CETI’s mission, many intangibles result from our collective research. Similar perhaps to how NASA produced a wide range of “spin-off” technologies with applications far beyond its original mission.
Within the most recent research, the surface-level text can be steered to resemble an essay, a narrative, or a technical explanation, while silently encoding an entirely different message beneath it. The result is not compression or encryption, but a one-to-one transformation between two coherent texts, both grammatically sound and semantically plausible. This possibility raises an unsettling question: when we read a piece of text, what exactly are we reading? Meaning, style, intent, or simply one of many valid surfaces through which information can pass?
This protocol invites a reconsideration of how we understand LLM hallucinations. Rather than framing hallucinations solely as failures of factual accuracy, the work suggests a deeper issue: a lack of intention. Language models excel at satisfying constraints (semantic, stylistic, and statistical) without any intrinsic commitment to meaning or truth.
In this sense, the model does not know the hidden message, nor does it believe the surface text. Instead, it functions as a conduit, capable of expressing information as is, in another framing, incapable of expressing at all. This challenges prevailing assumptions about what it means for an LLM to “know” something, and whether knowledge absent intention can meaningfully be called knowledge at all.
Project CETI’s steganographic protocol exposes the extreme constraint satisfaction problem underlying all LLM text generation. Each sentence is a negotiated compromise among fluency, coherence, topic, and probability. When models are asked to simultaneously convey one meaning while disguising another, the tension between these constraints becomes visible.
This tension clashes with what we historically expect from authors: a good-faith effort to convey purpose. In an era already flooded with machine-generated text, this clash further erodes the longstanding pact between intent and the written word.
Importantly, this result does not rely on frontier-scale systems. Project CETI shows that open-source LLMs with as few as 8 billion parameters are sufficient to produce high-quality steganographic text. Messages as long as the paper’s abstract can be encoded and decoded locally on a laptop in seconds. This accessibility underscores the significance of the finding. The decoupling of text from authorial intent is not a distant concern; it is already here.
The existence of such a protocol deepens existing challenges for AI safety, trust, and communication, at a moment when confidence in written language is already strained by the rise of conversational AI. As language models continue to evolve, so too must our understanding of what it means to read, to write, and to know.
The paper was published in ICLR on April 23, 2026. More information, including a copy of the paper, can be found online here.
###
About Project CETI
Project CETI (Cetacean Translation Initiative) is a nonprofit organization applying advanced machine learning and state-of-the-art robotics to listen to and translate the communication of sperm whales in the Eastern Caribbean off the coast of the island of Dominica. Its science team comprises leading artificial intelligence, natural language processing and complex systems experts, cryptographers, linguists, marine biologists, roboticists, engineers, and underwater acousticians in partnership between over 15 academic institutions and companies across six countries. In 2022, it published in iScience its scientific roadmap for advancing the understanding of sperm whale communication. This peer-reviewed academic paper outlines the key elements required for the collection and processing of massive datasets, detecting basic communication units and language-like higher-level structures, and validating models through interactive playback experiments.