Interspecies Communication

Speculative If a machine-learned model of an animal’s signals can be shown to carry meaning — which no study has yet demonstrated, and section 7 says exactly what would — the prize is a verified communication channel to a mind that is not human and not descended from human culture. That is the one prize in this corpus with no substitute: every other frontier extends a human capability, while this one adds a second party.

Frontier The magnitude is best stated as the size of the structure already found and the distance still to meaning. One study lifted a sperm whale repertoire from 21 catalogued coda types to a combinatorial system with at least 143 combinations frequently realised across 8,719 codas — nearly an order of magnitude more distinguishable signals — and, in the authors’ own words, “provides no characterisation of call semantics.” A machine can now map the alphabet of a species and say nothing about what a single letter means.

Established The gains, if they come, accrue unevenly. The species with decades of paired audio, video and individual identities — sperm whales, one dolphin community studied since 1985, elephants in two Kenyan parks — are the ones a model can be trained on; the other millions of animal species are not. Frontier Conservation is the most defensible near-term beneficiary: a model that flags which acoustic contexts predict distress or reproduction is useful to biodiversity restoration long before anyone claims to hold a conversation.

Handwave The condition, stated plainly: detecting combinatorial structure and demonstrating that a signal refers to something are different results, and only the first has been achieved for any of these species. The unclaimed $500,000 tier of the 2025 Coller–Dolittle Prize, offered for an algorithm that enables genuine two-way communication, is the field’s own admission that the second result does not yet exist. The bottleneck list in section 5 is where the prize is either earned or deferred.

Interspecies communication, as a research programme rather than a folk aspiration, asks a precise question: can the signals another species exchanges be read as carrying meaning, and can that reading be verified by prediction? The field has three layers, and they are routinely collapsed into one in public accounts. The first is detection — finding that a signal system has structure, that it is not random noise. The second is decoding — showing that a particular signal refers to, or reliably changes behaviour with respect to, a particular thing in the world. The third is dialogue — producing a novel signal that the animal treats as meaningful. Almost all recent headline results belong to the first layer.

The distinction that organises this brief is between combinatorial structure and compositional meaning. A system is combinatorial if it builds many signals from a few reusable parts, the way a keypad builds many numbers from ten digits; it is compositional if the meaning of a combination depends on the meanings of its parts and how they are arranged, the way “red car” depends on “red” and “car.” Detecting the first is a statistical exercise a machine can now do at scale. Demonstrating the second requires a behaving animal and a controlled test, and it has been achieved cleanly for only a small number of signals in a handful of species.

The instrument that separates the layers is the playback experiment: record a signal, or synthesise one, play it back in a controlled context, and measure whether the animal’s response matches the meaning the signal is claimed to carry. Every credible claim of animal meaning in the literature rests on a playback, and every current machine-learning result about whale, dolphin or elephant signals does not. This brief takes the measured record of both traditions seriously, sets the recent artificial-intelligence results against the older behavioural ones, and ends at the single experiment — a generative playback scored against a matched-acoustics control — that would convert a decoded structure into a demonstrated meaning.

Frontier The strongest recent structure result is the sperm whale coda study, and its authors are unusually careful about what it does not show. Sharma, Andreas, Gero and colleagues (Project CETI, MIT CSAIL and the Dominica Sperm Whale Project) analysed 8,719 codas from the Eastern Caribbean clan, previously catalogued as 21 coda types, and found that the clicks vary along four interacting features they name rhythm, tempo, rubato and ornamentation — 18 rhythm types and 5 tempo types, with rubato a continuous modulation of duration and ornamentation an extra click (Nature Communications, 2024). Frontier The features combine, yielding “nearly an order of magnitude more distinguishable codas” than the 21, with at least 143 combinations frequently realised; the authors call this a “sperm whale phonetic alphabet.” Established And they state the limit in one sentence: the study “investigates the basic structural elements, and not the semantics, of the sperm whale communication system” and “provides no characterisation of call semantics and features no playback experiments.” Frontier That last clause is the whole brief in miniature: a combinatorial code with an unknown map to meaning.

Frontier The strongest recent reference result is the elephant name study, and it clears a bar the whale study does not attempt: it includes a playback. Pardo, Wittemyer and Poole analysed about 469 rumbles from 101 unique callers to 117 unique receivers in Amboseli and Samburu, Kenya, and trained a model that could predict the intended receiver of a call from its acoustic structure “regardless of how similar the call was to the receiver’s vocalizations” (Nature Ecology & Evolution, 2024). Established They then played rumbles back to individual elephants and found that animals “differentially responded to playbacks of calls originally addressed to them relative to calls addressed to a different individual” — approaching the speaker and calling back more readily to their own name-analogue. Frontier The result matters because the naming is non-imitative: unlike a bottlenose dolphin copying another’s signature whistle, the elephants “probably do not rely on imitation of the receiver’s calls,” which is the arbitrary sign-to-referent link that human names have and animal signals almost never do.

Established The cleanest demonstrations of animal meaning are decades old and use playback, not machine learning. Seyfarth, Cheney and Marler showed in 1980 that vervet monkeys give acoustically distinct alarm calls for leopards, eagles and snakes, and that playing a recorded call with no predator present elicits the functionally appropriate escape — run into a tree for the leopard call, look up for the eagle call, look down for the snake call (reviewed by Snowdon, 2020). Established This is the founding case of functional reference: a signal that behaves, for the receiver, as if it stands for an external thing. Frontier The interpretation has been qualified since: many alarm systems are “best explained by urgency of the response required” rather than by the category of the threat, so a call may encode how fast to flee rather than what to flee from — a distinction that only more playbacks can settle.

Frontier The one firm case of compositional syntax in an animal is the Japanese tit, and it too rests on an ordering manipulation in playback. Suzuki, Wheatcroft and Griesser found that tits combine an alert call (“ABC,” scan for danger) with a recruitment call (“D,” approach the caller) into “ABC–D,” that hearing the natural combination makes birds both scan and approach, and — the load-bearing control — that birds “rarely scan and approach when note ordering is artificially reversed (‘D–ABC’)” (Nature Communications, 2016). Frontier They call it “the first firm evidence for compositional syntax in a nonhuman animal” and later set three criteria for the claim: context-dependent production, a response that depends on the component meanings rather than the whole as an unanalysed unit, and a response that depends on the ordering (PLOS Biology, 2018). Frontier Campbell’s monkeys supply a second candidate: an “-oo” affix that broadens an alarm call from specific to general, with playback showing a reduced antipredator response to the affixed version (Leroux & Townsend, 2020). Speculative Putty-nosed monkey “pyow–hack” sequences, by contrast, look idiomatic rather than compositional — the combination means “travel” in a way not obviously built from its parts.

Frontier The machine-learning frontier is a bioacoustic foundation model that classifies at scale and generalises to species it has never heard, but classifies structure, not meaning. Earth Species Project’s NatureLM-audio, an audio-language model pairing an audio encoder with an 8-billion-parameter language model, reaches state-of-the-art zero-shot performance on 7 of 9 species-classification datasets and, on its BEANS-Zero benchmark, identifies unseen species at 34.3% against a 0.5% random baseline, unseen genus at 14.8% and unseen family at 30.8% (arXiv, 2024). Frontier It classifies call type at 0.871 accuracy and counts individual zebra finches at 0.655, and Earth Species reports it naming the correct species “20% of the time… given a call and a choice between two hundred species it has never heard.” Established What none of these numbers is, is a measurement of meaning: the tasks are detection, classification and captioning, and the model “can hallucinate.” Frontier It is a powerful instrument for finding structure and a null instrument for demonstrating reference.

Frontier The closest thing to a two-way system is Google’s DolphinGemma with the CHAT interface, and its designers are explicit that it is not translation. DolphinGemma is a roughly 400-million-parameter model built on Gemma, trained on the Wild Dolphin Project’s paired audio and video collected since 1985, that tokenises dolphin sound with a SoundStream codec and “predict[s] the likely subsequent sounds in a sequence” the way a language model predicts the next word (Google, 2025). Frontier The companion CHAT system (Cetacean Hearing Augmentation Telemetry) issues synthetic whistles tied to objects the dolphins like, running in real time on a Pixel phone underwater. Established Its stated goal is “not to directly decipher the dolphins’ complex natural language, but to establish a simpler, shared vocabulary” — an invented pidgin, not a decoded language, and the distinction is the project’s own. Frontier Next-token prediction over dolphin sound will produce fluent-sounding dolphin; whether it produces meaningful dolphin is exactly the question a generative playback would test.

Established The field’s own prize structure encodes the gap. The inaugural 2025 Coller–Dolittle Prize gave its $100,000 award to Laela Sayigh’s work on bottlenose dolphin signature whistles — decades-old, playback-validated, behavioural science — while the $500,000 tier for an algorithm that achieves genuine interspecies communication went unawarded (reported by BBC Science Focus, 2026). Frontier That is the honest summary of the position: the money that rewards demonstrated understanding is still on the table, and the money that rewards careful behavioural work on a single well-studied signal has already been paid out.

Frontier The first open question is whether the sperm whale’s combinatorial code is compositional or merely combinatorial. A keypad is combinatorial without being compositional: the digits recombine, but “911” does not mean the sum of its parts. Whether rhythm, tempo, rubato and ornamentation are meaning-bearing dimensions or a phonetic substrate below the level of meaning is unknown, and the whale study “provides no characterisation of call semantics” by design. Speculative The optimistic reading — that the whales command “a large space of possible meanings” — is a hypothesis about capacity, not a finding about use.

Frontier The second is whether statistical structure alone can ever imply meaning. Human language obeys regularities — Zipf’s law, predictable conditional entropy — and so do many animal signal systems, but so do systems with no semantics at all, which is why information-theoretic signatures are suggestive and never decisive. Handwave The strong version of the artificial-intelligence claim — that a large enough model trained on enough animal sound will “decode” the language the way a cipher is broken — assumes a parallel corpus or a Rosetta stone that, for animals, does not exist: there is no bilingual text pairing whale with any known language.

Frontier The third is grounding: tying a signal to a referent requires synchronised behavioural and environmental data, not audio alone. The elephant and dolphin programmes that could plausibly demonstrate meaning are precisely the ones with decades of paired video, identity and context; a pure audio model has no purchase on what a call is about. Frontier This is why the useful near-term outputs are correlational — which acoustic features predict which behavioural context — and why calling that “translation” is a category error.

Speculative The fourth is whether non-imitative reference, as in the elephant names, is widespread or exceptional. If arbitrary vocal labels — signs whose form is not copied from what they denote — exist across several lineages, that changes the story of how reference evolved; if the elephant case is singular, it is a curiosity. Frontier Marmosets, dolphins and parrots all label individuals, but dolphins and parrots do it by imitation, which is a different and easier trick.

Frontier The fifth is generative: can a model produce a signal the animal treats as meaningful, and not merely recognise one? Recognition can be passive pattern-matching; production that changes behaviour in a predicted direction is the strong test, and it is the one no whale, dolphin or elephant programme has yet passed. Speculative DolphinGemma’s next-sound prediction and CHAT’s synthetic whistles are the first tools pointed at this question, but a synthetic whistle that dolphins adopt for “seagrass” is an invented convention, not a decoded word — a distinction section 15 returns to.

Established The workback target is a demonstrated meaning: a signal, natural or synthetic, whose effect on a behaving animal matches a prediction made in advance. Everything below is ordered toward that, and the binding link is not the model.

Frontier L1 — corpora with grounding. Audio paired with synchronised video, individual identity, and behavioural and environmental context, at scale. This exists for a tiny number of populations — one dolphin community since 1985, the Dominica sperm whales, two Kenyan elephant populations — and almost nowhere else. Established Without grounding, a model can learn the shape of the signal space and nothing about what any point in it is about.

Frontier L2 — structure discovery. Unsupervised segmentation of a continuous signal stream into units and the discovery of combinatorial regularities among them. This is the layer the whale study and the bioacoustic foundation models occupy, and it is the one that is working: 8,719 codas resolved into a four-feature combinatorial code, unseen species classified at 34.3%.

Frontier L3 — context alignment. Statistical association of discovered units with behavioural and environmental variables — which coda precedes a dive, which rumble precedes a reunion. This is correlational and achievable now with L1 data; it is the honest meaning of “decoding” as the field can currently deliver it.

Speculative L4 is the binding link: referential validation by playback. A test that a unit identified at L3 actually carries its putative meaning for the animal, run by playing the signal in a controlled context and measuring whether behaviour follows the prediction. Speculative This is cheap in equipment and expensive in rigour: it needs a pre-registered prediction, a matched-acoustics control, and a population habituated enough to test without confounding distress. Established It is also the only step that converts a correlation into a claim about meaning, and it is the step every current machine-learning result skips.

Handwave L5 — generative validation. The model synthesises a novel signal predicted to mean a specific thing, and the animal responds as predicted, against a matched-acoustics control. This is the “two-way communication” the $500,000 prize describes and nothing has demonstrated. Speculative DolphinGemma plus CHAT is the first stack aimed at L5, but on an invented shared vocabulary rather than the animal’s own; whether the same apparatus can validate a natural signal is untested.

Frontier The ordering matters because the field is investing hardest in L2, where progress is easy and cumulative, and hardly at all in L4, where progress is slow, behavioural and unglamorous. A larger model does not move you from L2 to L4; a playback does.

Established This subject depends on results it does not produce, and the first is a criterion for what would count as meaning in a non-human mind. The behavioural test — the animal acts as if the signal refers — is operational but shallow, and whether that constitutes meaning in any richer sense belongs to consciousness research and basal cognition. Frontier A model that predicts behaviour perfectly and understands nothing is a live possibility, and it is the same behaviour-versus-mechanism problem that machine consciousness confronts from the other side.

Frontier It depends on the measurement methods of machine learning that the field is still building. The zero-shot generalisation, calibration and hallucination behaviour of a bioacoustic foundation model are exactly the properties that intelligence measurement shows are routinely mis-reported for language models, and a captioning score of 0.532 SPIDEr is a benchmark number carrying all the interpretive hazards that brief documents. Established An animal-communication claim inherits every weakness of the evaluation methodology under it.

Established It depends on field behavioural ecology, which is slow, expensive and irreplaceable. The playbacks that would validate any decoded meaning require habituated study populations, permits, and researchers who can tell one animal from another — the Dominica sperm whale, Amboseli elephant and Wild Dolphin Project datasets are each multi-decade investments, and no amount of compute substitutes for them. Frontier The dependency runs the other way too: a validated channel would be a tool for biodiversity restoration and for the welfare questions in de-extinction technologies, where a revived species would inherit no living tutors for its calls.

Established E1 — the context-alignment baseline. Take a grounded corpus (whale, dolphin or elephant), label each signal with synchronised behavioural and environmental context, and publish the conditional distributions: which unit predicts which context, with intervals. Established This is achievable now, needs no new hardware, and is the honest current meaning of “decoding.” It is also the necessary input to every test below.

Frontier E2 — the natural-signal playback. For a unit that E1 associates with a context, pre-register the predicted behavioural response, then play the natural signal back against a matched-acoustics control — the same signal reversed, or a spectrally matched non-signal — and score whether behaviour follows the prediction. Established This is the vervet, tit and elephant method, and it is the step the machine-learning results have not taken; the Japanese tit’s reversed “D–ABC” control is the template.

Speculative E3 is the decisive experiment: a generative playback. A model synthesises an utterance it predicts will mean a specific thing, and the experiment scores whether the predicted behaviour follows, against a matched-acoustics control. Speculative This is decisive because it is the only test that a model has captured meaning rather than surface form: recognition can be memorisation, but a synthetic signal that moves behaviour in a pre-registered direction, while an acoustically matched control does not, cannot be explained by the model having merely learned the shape of the signal space. Handwave Nothing has passed it for a natural signal; DolphinGemma plus CHAT is the first apparatus that could run it, and it currently runs on an invented vocabulary rather than the dolphins’ own.

Frontier E4 — the non-imitative reference replication. Repeat the elephant name-call design — predict receiver identity from acoustic structure controlling for vocal similarity, then confirm by playback — in a second species with individual recognition, to establish whether arbitrary vocal labels are a general phenomenon or an elephant peculiarity. Frontier Marmosets, which label conspecifics, are the obvious candidate.

Speculative E5 — the compositionality disruption test. The strongest objection to animal compositional syntax is that a combination like “ABC–D” could be two separate learned associations rather than a rule; the test that distinguishes them is to disrupt a shared component process and show that both the combination and its relatives fail together, not one alone (Frontiers commentary, 2016). Speculative No animal study has yet run the disruption design that would settle whether the tit has a rule or a habit.

Frontier E6 — the negative control the field most needs. Train the same discovery pipeline on a signal system known to be non-semantic, or on shuffled data, and confirm it does not manufacture the same combinatorial “alphabet.” Handwave A structure-finding method that finds structure in noise is measuring itself, and this control is cheaper than any playback and almost never reported.

Established The instrumentation problem is grounding at scale: synchronised, identity-resolved, context-labelled multi-sensor recording of free animals. The whale and dolphin programmes need hydrophones arrayed for localisation, animal-borne tags carrying audio, depth, accelerometry and video, and the software to fuse them into a per-individual timeline. Established The Wild Dolphin Project’s paired audio and video since 1985 is the template and shows the cost: the bottleneck is decades of fieldwork, not sensor physics.

Frontier The modelling stack is converging on audio-language foundation models, with the usual trade-offs. NatureLM-audio pairs an audio encoder with an 8-billion-parameter language model; DolphinGemma is a 400-million-parameter model with a SoundStream tokeniser running on a phone. Frontier The engineering tension is between a large model that generalises across taxa and a small one that runs in real time underwater, where the latency between an animal’s signal and a synthetic reply is itself an experimental variable.

Frontier The generative-playback rig is the piece that does not yet exist as a validated instrument. It requires real-time synthesis of a target signal, underwater or in-air playback with controlled acoustics, a simultaneously recorded and matched control, and automated scoring of the animal’s response — a closed loop tight enough to run before the context changes. Speculative CHAT on a Pixel phone is the first field-deployed version, built for an invented vocabulary; retargeting it to validate a natural signal is an engineering programme, not a demonstration.

Established Every capability claim inherits the evaluation weaknesses of its benchmarks. The reporting discipline that intelligence measurement argues for — intervals, held-out sets, contamination controls, cost per result — applies unchanged to a bioacoustic model, and a zero-shot species-identification number without a held-out taxon split means as little here as an inflated benchmark score means there. Frontier A field that wants its meaning-claims believed will have to report its models the way that brief demands.

Frontier Bioacoustic monitoring is the mature adjacent discipline and the near-term payoff. The same models that discover structure are already used for passive acoustic monitoring — counting, identifying and locating animals from sound — which serves biodiversity restoration whether or not meaning is ever decoded. Established NatureLM-audio’s real product today is classification, not conversation.

Frontier Large language models supply the architecture and the hype in equal measure. The transfer of self-supervised sequence modelling from text to animal sound is the technical enabler, and the “it is only next-token prediction” intuition is both why the models work and why fluent output is not evidence of understanding — the same lesson machine consciousness draws about human-facing systems.

Speculative Collective-behaviour science is an under-used neighbour. Much animal signalling coordinates groups rather than addressing individuals, and the tools of collective intelligence — treating a pod or herd as a computing system — may fit whale and elephant communication better than a word-by-word translation frame. Frontier A signal’s job may be to synchronise a group, in which case “what does it mean” is the wrong question and “what does it coordinate” is the right one.

Established The field is organised around a few long-lived field programmes and a newer layer of artificial-intelligence labs and prizes, and the two have different incentives. Project CETI, the Wild Dolphin Project, Save the Elephants and the Dominica Sperm Whale Project are multi-decade behavioural-ecology efforts; Earth Species Project and Google’s DolphinGemma team are machine-learning groups; the Coller–Dolittle Prize sits between them, offering $100,000 for progress and $500,000 for genuine two-way communication. Frontier The risk is that the machine-learning layer’s publication incentives reward structure-discovery papers, which are fast, over playback validation, which is slow, so the measured record grows on the side that cannot demonstrate meaning.

Frontier Data are the scarce asset and they are unevenly held. The grounded, identity-resolved corpora are the product of specific long-term projects, and access to them determines who can make a credible claim; a foundation model trained on scraped audio without context can find structure but cannot validate meaning. Speculative A shared, well-documented corpus standard — audio plus synchronised context, with provenance — would do more for the field than a larger model, and no body currently owns that standard.

Established Animal-research ethics review is the binding institutional constraint on the decisive experiments. Playbacks, and especially generative playbacks that introduce novel signals, are interventions on wild animals, and the permitting and welfare-review regimes for them are stricter than for passive recording — appropriately, given the risks section 11 sets out. Frontier An institution that can run a pre-registered, welfare-reviewed generative playback is rarer than one that can train a model.

Frontier The first ethical risk is that the tools are ready before the understanding is, and playback is an intervention. Alcantara and Andrews argue that machine-learning decoding has “identified patterns, rather than meanings,” and warn that “playing opaque signals in the ocean could impact these individuals and cultures in ways that are impossible to predict” (Topoi, 2026). Established A synthetic signal whose meaning is guessed wrong can cause distress, disrupt a signal’s function, or damage relationships within a group — harms that a text model simply cannot inflict on its training data.

Frontier The second is anthropomorphism built into the goal. The same authors note that “pursuing translation in the first place may reflect a form of anthropomorphism, by assuming that vocal communication plays the same role for the animals… as it does for humans,” and that heavy human guidance of a model “introduces more opportunities for biases like anthropocentrism.” Speculative A model trained on human language first, then pointed at whale sound, may find human-shaped structure because it is built to.

Speculative The third is moral standing. If a species were shown to communicate referentially and compositionally, the case for its moral consideration strengthens, with consequences for whaling, fisheries, captivity and habitat law that the same evidence would make hard to ignore. Frontier That is an upside and a hazard at once: the incentive to over-read a result in either direction — to grant or withhold standing — is large, which is exactly when a matched-acoustics control matters most.

Handwave The fourth is the fantasy of frictionless translation, which the ethics literature treats as the dangerous case, not the harmless one. “Knowing how behaviors correlate with vocalizations is quite different from being able to understand and talk to animals,” and a public that believes the gap is closed will make welfare and policy decisions on a decoding that does not exist. Frontier The honest brief is the one that keeps the two apart.

Speculative A verified channel to a non-human mind would be the first contact humanity could actually have, and researchers reach for the alien analogy because it is apt. The lead author of the whale study observes that the work “parallels the hypothetical scenario of contacting alien species” — the methods for decoding an unknown signal system are the same whether its source is cetacean or extraterrestrial, which is why this field and the search for extraterrestrial intelligence borrow each other’s mathematics.

Speculative The plausible civilizational effect is not conversation but recategorisation. If several species were shown to use arbitrary reference and simple syntax, the human/animal line that underwrites food, research and land-use systems would face evidence it was not built to absorb, and the pressure would fall hardest on the industries with the most contact with the best-studied species. Frontier That shift could come from careful playback science alone, with no machine translation at all — the vervet result was already this kind of evidence, forty years early.

Handwave The exotic version — routine two-way dialogue with whales or elephants — is not supported by anything in the measured record. Nothing forbids it in principle, and nothing demonstrates it in practice; the distance between a combinatorial alphabet with an unknown semantics and a shared conversation is the distance between this brief’s frontier and its fantasies. Speculative The more likely long-run outcome is a partial, domain-specific channel — understanding a dozen validated signals well — which is a smaller prize than the films promise and a larger one than the field has yet earned.

These horizons track the demonstration of meaning, not the discovery of structure — the structure results are already here.

• 10 yr: Frontier Context-alignment “decoding” (which signal predicts which behavioural context) is published for several grounded species with intervals, and at least one natural-signal playback validates a machine-discovered unit against a matched control. Bioacoustic monitoring is routine in conservation. A generative playback of a natural signal is attempted; whether it succeeds is the open question.
• 25 yr: Speculative A generative playback succeeds and replicates for at least one signal in one well-studied species — a synthetic utterance that moves behaviour as predicted, against a matched-acoustics control. The $500,000 tier is claimed or credibly contested. The number of validated signals is small.
• 50 yr: Speculative A domain-specific channel exists for a handful of species: dozens of signals understood well enough to predict and produce, grounded in decades of behavioural data. Whether any of it constitutes “language” remains contested and may be the wrong question.
• 100 / 250+ yr: Handwave Routine, open-ended dialogue with another species — or the recognition that the combinatorial systems we mapped coordinate behaviour without ever having carried human-style meaning, so the translation frame quietly dissolves.

• Depends on Nothing on this map blocks the structure-discovery work, which is already delivering. Its dependencies are conceptual and behavioural rather than technological: a criterion for meaning, which it shares with consciousness research and basal cognition, and the evaluation discipline set out in intelligence measurement, whose benchmark hazards this field inherits wholesale.
• Requires (not on this map) Four things the field does not yet have. First, a single playback that validates a machine-discovered signal against a matched control — the result that would turn structure into meaning, and which no whale, dolphin or elephant programme has produced. Second, grounded multimodal corpora — audio plus synchronised video, identity and context — for more than the few populations that have them. Third, an animal-welfare review pathway for generative playback on wild animals, stricter than anything passive recording needs. Fourth, animal-borne multi-sensor tags and localisation arrays supplied at the scale fieldwork actually requires.
• Enables A validated channel would feed biodiversity restoration and reframe the welfare questions around de-extinction technologies, where a revived species would inherit no tutors for its calls.
• Adjacent Shares its architecture and its overclaiming risk with machine consciousness, its group-signal frame with collective intelligence, and its measurement problems with intelligence measurement.

Established “AI has decoded whale (or dolphin, or elephant) language.” No study claims this, and the whale study explicitly disclaims it: it “provides no characterisation of call semantics and features no playback experiments.” What has been decoded is structure — a combinatorial code — and structure is not meaning. Frontier The headline verb “decode” is doing work the papers do not.

Frontier “The sperm whale phonetic alphabet is a language.” An alphabet is a set of contrastive units, not a language; establishing that whales combine four features into 143-plus distinguishable codas says the system is expressive, not that any coda means anything. Speculative The claim that this represents “a large space of possible meanings” is a hypothesis about capacity, and the space of possible meanings of a keypad is also large.

Frontier “A big enough model will crack it like a cipher.” Code-breaking and translation both rely on a parallel text or known plaintext; for animal signals there is no bilingual corpus pairing whale with any human language, so there is nothing to align against. Handwave Unsupervised alignment of two languages from monolingual data works partially between human languages that share a conceptual world; that shared world is precisely what is not given between species.

Established “DolphinGemma lets us talk to dolphins.” Its designers state the goal is “not to directly decipher the dolphins’ complex natural language, but to establish a simpler, shared vocabulary” — an invented pidgin of synthetic whistles for objects, which is a genuine achievement and not a translation of anything the dolphins already say. Frontier A dolphin adopting a synthetic whistle for “scarf” has learned a convention, not revealed a word.

Frontier “Animal calls are just emotional expression, not reference.” This is the opposite error and it is also refuted: vervet alarm calls elicit predator-appropriate escape with no predator present, and Japanese tit “ABC–D” sequences fail when their order is reversed — both are behaviours that a pure affect account cannot explain. Frontier The truth is in between: some signals are referential, some are motivational (the urgency account of alarm calls), and telling which is which needs a playback, not a prior.

Speculative “Elephants and dolphins both have names, so it is the same discovery.” Dolphins copy a signature whistle — the name resembles its bearer; the elephant result is stronger precisely because it is non-imitative, an arbitrary label whose form is not copied from the receiver, which is the property human names have. Frontier Conflating the two throws away the one feature that made the elephant finding notable.

Established “Compositional syntax in birds proves animals have language.” The tit result is real and is “the first firm evidence for compositional syntax in a nonhuman animal,” but a single validated two-part combination is not a grammar, and critics note that compositionality is not the same as the systematicity that characterises human language — the combination could be two learned associations rather than a rule (Frontiers, 2016). Speculative The disruption experiment that would settle it has not been run.

Handwave “We are on the verge of two-way conversation with another species.” The field’s own $500,000 prize for exactly this is unclaimed, the decisive generative-playback test has not been passed for any natural signal, and the ethics literature treats the belief that the gap is closed as itself a source of harm. Frontier The defensible statement is that we can now map the structure of several signal systems and validate the meaning of a few individual signals — a real and bounded achievement.