0:01 [music] >> Recently, an artificial intelligence system analyzed thousands of recordings from sperm whales. The results showed that their clicks are not simple signals. They change in rhythm, speed, and arrangement, forming a clearly organized sound system. From this discovery, >> [music] >> researchers began matching individual click sequences with whale behavior. Some patterns initially appeared in similar situations, leading the team to believe they were approaching the first real translation. But as more data was added, the results became harder to explain.
0:47 The same sequence of clicks could appear when two whales moved closer together, when one animal left the group, or when the entire family [music] suddenly changed direction. AI had discovered how the sounds were constructed, but it still could not determine what they actually meant. If the alphabet had been found, why could whale language still not be translated? The animal that lives through sound. To understand why AI could recognize the structure, but still fail to identify the meaning, it is first necessary to look at the whales producing these sounds and the world in which they use them to survive. Sperm whales are the largest toothed whales on Earth. Adult males can grow longer than a school bus, with a head that makes up nearly 1/3 of
1:33 the entire body. Inside that enormous structure is a sound producing system capable of generating extremely powerful clicks, which the whales use to navigate and locate prey in the deep ocean. They regularly dive into waters where sunlight almost completely disappears. At those depths, vision provides little help when hunting squid or avoiding obstacles. Sperm whales produce clicks and listen for the returning echoes. From the delay, strength, and shape of those echoes, they can estimate the distance, direction, and size of an object ahead. Sound is therefore more than a method of communication. It is the main way these animals experience the world around them. The social lives of sperm whales are also
2:18 organized around sound. Females and calves usually live in stable family groups led by older females. Members may remain together for many years, help care for the young, and maintain traditions that are unique to their community. Larger social groups use distinctive click patterns, almost like dialects, that help identify different cultural clans. The short groups of clicks used in social communication are known as codas. They are different from the rapid click sequences sperm whales use while hunting. For decades, codas were placed into a relatively small number of simple categories. One pattern might be treated as an identity signal, a contact call, or a way to keep the group together. Project CETI began collecting data around
3:04 Dominica to test whether this interpretation was too simple. Hydrophones recorded underwater sound. Drones observed the position of whales from above. Temporary sensor tags attached to their bodies measured depth, movement, and body direction. Individual whales were identified, so researchers could determine who was producing each sound and which animals were nearby. Before AI could begin its analysis, however, the continuous stream of sound had to be divided into separate sections. Researchers decided where one coda began, where it ended, and how it differed from the coda that followed. Each segment was then entered into the system as an individual unit. This method rested on one important assumption, that a single coda was a
3:50 complete enough message to analyzed on its own. If that assumption was correct, AI could compare thousands of codas, find repeating patterns, and connect them with behavior. But, if a message extended across several turns, several periods of silence, or several different whales, then the data may have been divided incorrectly from the beginning. A coda might not be a word. It might only be a syllable. A response from a second whale could contain the rest of the sentence. The silence between them might signal urgency. The movement of the group might complete the meaning. The entire decoding effort depended on whether researchers had divided whale language in the right place. The alphabet appears. When thousands of
4:35 codas were entered into the model, the first results seemed to confirm that dividing the data into smaller units had been reasonable. AI did not find a chaotic collection of sounds. It found an order. Codas that had once been placed in the same category actually contained many subtle variations. The number of clicks could remain the same while the speed changed. The spacing between clicks could expand or contract. An extra click could appear at the beginning or the end. The rhythm might remain steady, shift briefly, and then return to its original pattern. These differences were not randomly distributed. They appeared according to repeating rules and could be combined in several ways. AI identified four main types of
5:21 variation: rhythm, tempo, subtle changes in timing, and additional clicks known as ornamentation. By adjusting a limited number of basic elements, sperm whales could produce far more codas than earlier systems had recognized. This discovery changed the way researchers viewed the entire system. If each coda were simply a fixed signal, whales would need to remember hundreds of separate messages. But, if codas were built from smaller parts that could be modified and combined, the whales would only need a set of rules to create many different patterns. That is why researchers began comparing the system to a phonetic alphabet. This did not mean AI had discovered letters like A, B, or C. It meant the clicks appeared to be constructed from elements that could be
6:07 combined, much like a limited number of speech sounds can create a vast number of words in human language. A deeper acoustic analysis also detected changes within the sound spectrum of individual clicks. Some patterns shifted from one acoustic state to another in ways researchers compared with vowels and diphthongs. To the human ear, they still sounded like knocks or taps, but instruments and algorithms could measure the subtle differences hidden inside them. AI has moved far beyond simply counting clicks. It could identify how whales changed individual parts of a coda and how those parts shifted depending on their position within an exchange. The problem emerged when the team tried to connect structure with meaning. An alphabet can reveal how sounds are
6:54 formed. It cannot reveal whether a sequence refers to food, a calf, a direction of travel, or a threat. A model may prove that two codas are structurally different while still having no idea what that difference does to the whale hearing them. Meaning has to be inferred from behavior. If a coda appears before a group changes direction, it could be a movement signal, but it could also be a response to the previous sound, an identity marker, or a simple confirmation that one whale has heard another. The pattern initially interpreted by AI as meaning something close to stay near appeared convincing because it often occurred when family members separated or returned. But as more situations were added, the connection became weaker.
7:41 The sequence may not have been directing behavior at all. It may have been confirming a relationship between two whales or simply opening a longer exchange. The alphabet had appeared, but no sequence behaved like a word with a fixed meaning. That pushed the investigation toward a part of the data that had usually been treated as empty space, the seconds when no clicks were heard. What was missed between the clicks? The first models focused on sound because sound was the easiest part to measure. A click had a clear starting point, frequency, intensity, and shape. The silence between two sequences was mainly used to determine where one coda ended and the next one began. But once meaning became unstable, timing itself began to look like data.
8:28 Two codas could contain the same number of clicks and the same internal rhythm, yet one might appear immediately after another whale called, while the other followed a long period of silence. If the group reacted differently, the important difference might not be inside the clicks. It might be in the moment when they were produced. Timing also changes how human speech is understood. An immediate answer can suggest certainty or urgency. The same answer after a long pause can imply hesitation or a different intention. When two people speak at the same time, the interruption can also change the tone of the exchange. Those human meanings cannot simply be assigned to whales. Even so, the data showed that timing within codas was not random.
9:13 Speed and the stretching or compression of rhythm changed in systematic ways. That raised the possibility that silence between codas could also have a function. Researchers began preserving the timing information that had once been discarded. They measured the delay before a response, how often whales produced sounds over one another, and the length of the gaps between consecutive sequences. Some exchanges moved quickly, almost as though two whales were completing each other's rhythm. Others contained long pauses, but still continued in recognizable patterns. If response time forms part of the message, dividing sound into independent codas removes the connection between them. AI may see two identical patterns without knowing that one is an immediate
10:00 answer, while the other is a continuation after a long wait. This may partly explain why the first translation failed. The sequence interpreted as stay near might carry that function only when it appears within a specific period after a calf has called. If it is produced later or without a previous call, it may serve a completely different purpose. Another possibility is even more striking. The silence may be the space in which a whale waits for another individual to complete the structure. A mother produces the first sequence. Her calf responds. Another adult joins the exchange. The mother then repeats a shorter variation. If these four parts are separated, AI sees four codas.
10:45 But to the whales, the entire sequence may form one exchange. In that case, meaning is built across several turns. Each whale contributes only the part needed for the others to understand what will happen next. Such a system would fit the life of a family that must coordinate while some members dive into deep water, others remain near the surface, and calves depend on care from several adults. When timing was added to the model, some contradictions became less severe. Yet the same sequence still produced different results, even when the pauses were almost identical. The remaining difference depended on two variables that a sound recording alone could not explain. Who was speaking and who was listening. Meaning changes with the listener.
11:31 Sperm whales do not produce sounds in a social vacuum. Every click occurs inside a network of relationships that may have existed for many years. Adult females live alongside their mothers, sisters, offspring, and other familiar whales. Calves grow up hearing the same clicking voices every day. An experienced adult holds a different role from a young calf that is still learning how to move with the group. A mother preparing to dive is also in a different situation from the whale that remains near the surface to care for the young. When the identities of the speaker and the listener were added to the data, many patterns began separating into clearer groups. One coda might appear frequently between a mother and her calf, yet rarely pass
12:17 between two adults. Another variation might be used more often by one family than by a neighboring family. Clan identity codas suggest that sound can carry information about community, not only about an immediate action. This weakens the assumption that every coda has one universal meaning. A short phrase such as "It is time" only becomes clear when the listener already knows what has been planned. It could mean that it is time to leave, begin a task, or carry out an agreed action. The sound itself does not contain all the information. The rest exists in the shared knowledge of the people involved. Sperm whale communication may depend on shared knowledge in a similar way. A click sequence between a mother and her calf
13:03 may not need to describe in full that the mother is about to dive and the calf must remain with another whale. If the same routine has happened many times, a short signal may be enough to begin a process that both animals already understand. The same signal may produce a different result when exchanged between two adults because the relationship and expectations are different. >> [music] >> AI sees the same pattern and waits for the same behavior. The whale hears that pattern within the private history of a relationship. The proposed translation, stay near, may have failed because it turned a relationship-dependent signal into a general command. The sequence may have been closer to your turn, continue as usual, or a form of confirmation that only carries meaning between two specific whales.
13:51 This remains a hypothesis, but it could explain why behavior changes while the sound stays the same. Cultural clans make the problem even more complex. Communities living in the same ocean region may use different coda patterns that are learned and passed on through social contact. A model trained on one clan might recognize the sounds of another clan while misunderstanding their function, much like recognizing the syllables of an unfamiliar language without knowing the customs of the community that speaks it. There may not be a single dictionary for the entire species. A coda that is common around Dominica may not serve the same role in another population. Differences between groups may not be unwanted noise in the data. They may be central to the system
14:36 itself. To translate this communication, AI would need to learn family history, age, social role, community identity, and the level of familiarity between individuals. A sequence becomes easier to understand only when it is placed inside the relationship that produced it. Yet, even after identity and relationships are added, the same mother may produce the same coda for the same calf and receive a different response. The missing information may lie in what both animals are sensing in the surrounding ocean, something the recording alone cannot reveal. The message exists beyond the recording. To find the missing element, researchers must combine sound with movement data.
15:23 Sensor tags reveal depth, body direction, and speed. Drones show the distance between family members. Hydrophones record codas alongside ship noise, the sounds of other animals, and the click sequences used for navigation. Once these sources are synchronized, the same coda begins appearing under very different physical conditions. Sometimes it is produced while the group holds formation near the surface. Sometimes one whale has just turned downward. In one case, a calf may be only a few feet from its mother. In another, the two may still be close while a third family member approaches. The same sound does not always lead to the same action because the situation was already different before the coda was produced. Sperm whales do not
16:09 experience the ocean the way a camera sees it from a boat. Most of their lives unfold below the surface inside a three-dimensional space often where no sunlight remains. Position is not limited to forward, backward, left, and right. Another whale may also be hundreds of feet above or below. An individual may receive echoes from prey, underwater terrain, and other whales before research equipment can identify the source. Two whales may hear the same coda while already knowing each other's position or the direction the group is traveling. A microphone records the spoken part, but not the full experience of the speaker and listener. This is the greatest barrier between recognizing structure and understanding meaning. AI may detect a pattern that appears
16:55 again and again, but to know what that pattern represents, it must connect the sound with a stable object or action. When context is missing, the same sound can appear contradictory. The word that offers a simple example. The sound alone does not reveal whether that refers to a ship, a direction, or a place below the surface. The listener understands because both individuals are focused on the same thing. A model that only has the recording cannot know that shared focus exists. Sperm whale communication may contain similar signals. A short coda could mean something close to below, at that position, or continue as before. The meaning may not exist entirely inside the coda. It may be completed by
17:42 the environment and by shared knowledge. The sequence once translated as stay near may therefore not have been an instruction. It may have triggered a familiar behavior based on the formation of the group, the depth of the whales, and the positions of particular individuals. When those conditions change, the result may also change, even if the sound remains the same. A deeper possibility is that sperm whale communication is not organized around concepts that feel familiar to humans. These animals experience the world through echoes, pressure, and movement through the water column. Important ideas may be connected to the shape of an echo, a change in depth, or the level of coordination within the group. A coda might describe an acoustic
18:28 state that no human has ever directly experienced. AI can recognize the pattern, but may have no suitable label to attach to it. This does not prove that whales possess a complete language like human language, but it explains why discovering an alphabet cannot lead directly to translation. To understand a message, the model must know what the whale is sensing, which direction it is facing, >> [music] >> and what the other family members are doing. That requires a continuous record of the group's life, not only isolated codas. A complete sperm whale sentence may be spread across several elements at once. It could include the sound made by the first whale, the delay before the second whale responds, the position of a calf, the movement of the group, and the
19:14 conditions in the surrounding water. The current archive may not lack recordings. It may lack the world surrounding each sound. AI does not simply need millions of additional clicks. It needs a model of the world those clicks are describing. AI may never have heard a complete sentence. The original paradox now has a clearer answer. If AI has found the whale alphabet, why can it still not understand what the whales are saying? An alphabet only describes how sounds are formed. Meaning may depend on the whale producing the sound, the whale listening, the relationship between them, the sound that came before, the delay before a response, the movement of the entire group, and the conditions in
19:59 the surrounding environment. AI has recognized how codas change in rhythm, speed, and structure. But that does not guarantee the model knows what a sequence is pointing toward. The first translation was shocking, not because the whales had spoken a frightening sentence about humans. It was shocking because a result that appeared reliable collapsed when it was tested in more situations. The same sound sequence led to different behaviors. The model had recognized the correct form, but it had searched for meaning in the wrong place. Stay near was only a label placed on a temporary correlation. Once social relationships, timing, and environmental mental conditions were added, that label was no longer enough. The message may have been part of a
20:45 coordination routine, a confirmation, or a signal that only carries meaning between whales with shared experience. This changes the goal of the decoding effort. The task is no longer to collect more clicks and search for matching words. Every coda must be connected with the speaker, the listener, position, depth, the next response, and the history of the exchange. An interpretation becomes reliable only when it can predict behavior during new observations. Sound playback must also be approached carefully. A coda may be correct in sound while being completely wrong in social context. It may belong to a particular whale, clan, or situation. AI could learn to produce a whale voice before it understands the sentence it is
21:31 speaking. The first true success may not be a sentence translated into English. It may be a small prediction that can be repeated. When this coda appears between two whales with a specific relationship at a certain depth and in a certain formation, the next response is usually one particular action. Meaning may have to be built gradually from narrow and reliable connections like these. Sperm whales developed their communication system in darkness, under immense pressure, and inside families that continue across generations. That system did not evolve to match the human habit of dividing language into words, sentences, and fixed meanings. The alphabet is only the first layer. Behind it lies timing. Behind the timing lies the relationship.
22:17 Behind the relationship lies a sensory world that current equipment can record only in fragments. AI may have listened to thousands of hours of communication, but if a sentence is created by several whales, several periods of silence, and the entire situation around them, the model may still never have heard a single complete sentence. The first message was not stay near. It was a warning about the decoding process itself. Sound is not the whole language. To understand what whales are saying, AI must first understand how they live.
