For as long as there has been human curiosity, we have wondered what non-human animals say to each other. Decades of research in ethology and bioacoustics have laid the groundwork for understanding, and recent AI advances, such as self-supervised learning and large acoustic models, are now enabling large-scale investigation, revealing structure in previously opaque signals, opening up scientific questions that were computationally inaccessible a decade ago.

These tools, however, were not designed with animal communication in mind. Deploying them meaningfully requires facing challenges that human language processing did not: the absence of shared, annotated corpora; deep uncertainty about what a communicative ‘unit’ is across species and even for a given species; and the ever-present risk of projecting human linguistic structure onto fundamentally different minds. Progress requires merging the specialized knowledge of ethologists and bioacousticians with machine learning and more generally AI expertise.

This Dagstuhl Seminar brings together leading researchers from diverse fields (ethology, bioacoustics, ecology, linguistics, neuroscience, information theory, and ML) to ask: What are the next frontiers opened by AI to decode animal communication? The objective is to review open problems in emerging methodologies and ethics, identify promising avenues for discovery, and establish shared standards and interdisciplinary relationships crucial for the field's advancement over the next decade.

Discussion will be organized around six open questions:

1. Representations and units. What do self-supervised models actually learn from animal acoustic signals? Is there any overlap or relationships between the communicative units uncovered by AI and those used and distinguished by animals?
2. Structure and meaning. What evidence would fully support the presence of syntax, semantics or reference in a species “language”, and where do linguistic and information-theoretic frameworks illuminate rather than mislead?
3. Data and benchmarks. Which shared corpora, annotation standards and evaluation protocols does the field need, and how should model outputs be validated against behavior rather than against intuition?
4. From correlation to experiment. How can we put models to the test in the experimental realm, through synthesis, playback, and interactive protocols with animals?
5. Scale and application. What do these methods offer for monitoring, conservation, and animal welfare, and where might they fail?
6. Ethics and responsibility. Which standards should govern the interpretation, synthesis, and public communication of animal signals, given the risks of over-claiming and of manipulating the animals we study?

Creative Commons BY 4.0

Julie E. Elie, Nicolas Mathevon, Olivier Pietquin, and Pratyusha Sharma

This seminar qualifies for Dagstuhl's LZI Junior Researchers program. Schloss Dagstuhl wishes to enable the participation of junior scientists with a specialisation fitting for this Dagstuhl Seminar, even if they are not on the radar of the organizers. Applications by outstanding junior scientists are possible until November 6, 2026.