Five years. Fifty scientists. $33 million in TED Audacious funding. And as of early 2026, the most ambitious interspecies AI communication effort in history has produced a sperm whale phonetic alphabet. 156 distinct codas identified. Vowel-like sounds and diphthongs mapped. A transformer model — WhAM — that can generate synthetic codas.

No translation. Still no translation.

The sperm whale, which has the largest brain of any animal that has ever lived, is still talking past us. That’s not a failure of ambition. It’s a precise description of where the science actually is — and it’s the honest starting point for a conversation the science press hasn’t had yet.

Here’s why it matters specifically now. In 2024, three separate research groups produced work that converges on the same hard question: how do you read meaning in a system that assembled its own language, when the system may not share human linguistic structure? MIT CSAIL showed constraint-based models can decipher dead human scripts. The University of Southampton showed AI agents spontaneously develop spatial grammar. Anthropic showed that their own model’s internal representations require a second AI to partially decode — and even then, 30% resists interpretation.

And then there’s Ben Zion et al., published on arXiv in November 2024. That paper proved formally — not empirically, formally — that the field’s dominant training objective produces agents that can hit near-perfect task performance while communicating something that is, technically, semantically meaningless. Nobody in the science press connected that proof to the CETI coverage running simultaneously. That’s the gap this piece closes.

The Proof Nobody’s Foregrounding

The standard setup in emergent communication research is called a referential game: one agent has information, another doesn’t, and they coordinate to complete a task. You measure success by task completion. Simple enough. The problem — and this is what Ben Zion et al. proved formally in arXiv:2411.10173 — is that the most common training objective in this field can produce communication protocols that are semantically inconsistent and still achieve near-perfect task scores.

Let me make that concrete. “Semantic consistency” means this: similar inputs should produce similar messages. If a sender encodes “red ball on the left” with symbol X, then “red sphere on the left” should produce something close to X. That’s what you’d expect from a language that means something. Ben Zion et al. prove, under mild assumptions, that the discrimination objective — “identify the correct item from a set” — allows agents to solve the task using protocols where the same symbol means entirely different things in different contexts. The game score will not reveal this. Near-perfect performance. Semantically empty protocol.

The practical implication for CETI: CETI’s research framework for validating whether they’ve decoded meaning relies partly on behavioral response testing — playing codas back to whales and observing reactions. That validation loop isn’t wrong. But it hasn’t been completed yet. And the Ben Zion proof means the internal AI models generating “whale language” representations could, in principle, be doing sophisticated coordination without semantic grounding — and the performance metrics wouldn’t catch it.

David Gruber, CETI’s research lead, described the gap with more honesty than most science coverage gave him credit for: “We’re barely scratching the surface.” Bioneers interview, April 2025. That wasn’t false modesty. That’s what you say when you know exactly where the wall is.

What MIT’s Dead-Language Work Actually Proved

Linear B — the Mycenaean Greek script from the 15th–12th century BCE — resisted decipherment for decades. Not because the script was complex. Because nobody tried the correct related language. Michael Ventris cracked it in 1953 on a hunch that it was related to Greek. The entire breakthrough was contingent on that single correct hypothesis. Remove the hypothesis, no decipherment.

Jiaming Luo, Yuan Cao, and Regina Barzilay built a model that doesn’t need the hypothesis. Their paper, “Neural Decipherment via Minimum-Cost Flow: From Ugaritic to Linear B,” ACL 2019, correctly translates 67.3% of Linear B cognates without being told the language was related to Greek. It inferred the relationship from phonological evolution patterns. Languages don’t change randomly — a “p” in a parent is plausibly a “b” in a descendant, almost certainly not a “k.” Encode those probability distributions. Solve for the mapping that minimizes phonologically impossible transitions. Minimum-cost flow optimization.

Genuinely powerful. The scope boundary is just as real, though. This works because human languages share phonological evolution principles — that’s the load-bearing assumption. Dead human scripts: the method is probably fine. Sperm whale codas: they don’t have phonemes in the human sense. The codas are click-based, three-dimensional, produced by a melon-based acoustic system with no vocal cords. The “phonological priors” that make MIT’s model work simply don’t exist in the same form. The approach is brilliant for what it is. What it is doesn’t transfer cleanly to non-human communication.

What Agents Invent When Nobody Gives Them a Grammar

Olaf Lipinski’s two-paper arc is worth tracing properly because most coverage only cited the xenolinguistics angle and missed what the methodology actually demonstrated.

First paper (arXiv:2310.06555, updated May 2024): temporal references in emergent communication. The finding — altering loss function alone is insufficient for temporal references to emerge. You need architectural changes, specifically a different batching method. Over 95% of agents with the modified architecture developed temporal references. Without it: nothing. Temporal grammar is a structural problem, not a training objective problem. That matters.

Second paper (arXiv:2406.07277, NeurIPS 2024): spatial references. Agents develop spatial deixis — the linguistic capacity to point through language: “next to,” “at the end,” “above.” Compositional messages for most positional references. A separate class of compressed, non-compositional tokens for edge positions where full composition wastes bandwidth. Over 90% task accuracy. And — crucially — Lipinski uses NPMI collocation analysis specifically to demonstrate that the symbols have stable associations with spatial contexts, not just task outcomes. This is the methodological move that gets past the Ben Zion problem. Stable symbol-context associations = evidence of semantic grounding, not just coordination.

The scope limitation here is one Lipinski acknowledged directly: his agents were trained on human-designed tasks, in human-designed game environments, with human-legible inputs. Their emergent language may carry implicit human structure not because spatial deixis is universal, but because the task was built by people who think in spatial categories. It’s the cleanest methodology in the field right now. It still can’t tell you whether the same patterns would emerge in a system built for echolocation in deep-ocean pressure gradients.

Anthropic Can’t Fully Read Its Own Model Either

In May 2024, Anthropic published “Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet.” They trained a sparse autoencoder — a second model — on Claude 3 Sonnet’s activations. Roughly 34 million features extracted. Among them: a cluster associated with the Golden Gate Bridge. Activate that cluster, the model talks about the Bridge. Suppress it, the concept disappears. 70% of features in an earlier smaller test mapped to single human-interpretable concepts.

Thirty percent did not map to anything nameable.

That 30% number is the thing I keep coming back to. This is a cooperative system — Anthropic built it, has full access to its weights, can retrain the interpretability model as many times as they want, and they’re working in a domain (English language) that has decades of linguistic prior. And still 30% of internal representations are opaque. Now ask what the equivalent number would be for a system you can’t access at all, running a communication protocol evolved for purposes you can only guess at.

The sparse autoencoder approach is, among other things, an attempt to build ground truth from the inside. The CETI problem is whether there’s an equivalent approach for a system you cannot retrain. There isn’t one yet.

The Convergence None of These Papers Name

This is the synthesis that doesn’t exist in any single cited paper. I’m going to state it plainly because the academic register tends to bury it.

Luo, Cao & Barzilay establish that constraint-based decipherment can recover cognate structure from unknown scripts without bilingual reference — using only phonological priors and minimum-cost optimization. Lipinski, combined with Ben Zion, establishes two things together that neither establishes alone: agents develop interpretable spatial and temporal grammar from scratch, AND the standard training objective produces semantically inconsistent protocols that defeat standard evaluation. You need both structural analysis (NPMI, compositionality scoring) AND the correct training objective (reconstruction, not discrimination) to claim genuine communication rather than coordination. Anthropic’s Scaling Monosemanticity establishes that even with a cooperative system, a second dedicated interpretability model, and full internal access, 30% of representations remain opaque.

Put them together: the actual difficulty of the decipherment problem becomes more precisely defined than any single track makes visible. You need structural analysis tools that work without a Rosetta Stone. You need training objectives that produce semantically consistent protocols. You need a second model to read the first. And even then, for a system you cannot retrain from scratch — whale codas, a hypothetical non-human signal — the Anthropic approach doesn’t directly apply.

CETI is working on the hardest version of this problem: a system that evolved its own communication for its own purposes, in a sensory modality that has no human equivalent, with no ground truth available and no ability to retrain the sender.

And the scope limit applies across all four methodologies: every approach carries human-language priors in its foundation. Ben Zion’s semantic consistency framework uses Euclidean distance in input space as a proxy for semantic similarity — a human-legible assumption. Whether sperm whale codas even have a meaningful notion of “similar inputs should produce similar messages” is not established. We don’t know if that assumption applies to a system that evolved for three-dimensional deep-ocean echolocation over millions of years.

Evidence Across the Four Research Tracks

Three Questions That Would Produce Exclusive Reporting

There’s a story here that nobody has written. The science press covered CETI’s phonetic alphabet announcement. They covered Lipinski’s xenolinguistics angle. Nobody covered the Ben Zion proof appearing in the same three-month window as the CETI progress updates — and the specific implication: we announced progress on whale communication at the same moment a peer-reviewed paper formally proved our evaluation framework can’t distinguish meaningful communication from sophisticated-looking noise.

Ask Lipinski: Was his NPMI evaluation methodology designed with the Ben Zion semantic consistency problem in mind, or developed independently? If independently: parallel discovery happening in real time across two research groups who weren’t aware of each other. If deliberately: the field is converging on a solution faster than the publication record suggests. Either answer is the story.

Ask CETI’s Jacob Andreas: Does the “sperm whale phonetic alphabet” finding constitute semantic decoding or structural categorization? He’s the MIT NLP researcher who led that work, and he’ll know exactly what the distinction means. His answer determines whether the headline is “breakthrough” or “foundation.” These are not the same thing.

Ask a Barzilay lab member: Has the minimum-cost flow model been tested on non-human-language corpora? The gap in that answer is where the next paper lives.