The Voice in the Garden

Genesis describes the first contact between humans and animals as an act of speech.

Nobody in the scene is afraid. A chapter earlier the humans have been told to rule over the fish, the birds and every living thing (1:28), and in the next two verses they and the animals alike are given plants to eat. At this point ruling and killing have nothing to do with each other.

After the flood God repeats the blessing, with one change.

The two blessings open with the same words: increase and multiply, and fill the earth. The first then gives the humans dominion over the animals. The second, at that same point, says that the fear and dread of them will be upon every animal, and the permission to eat the animals comes in the next sentence. It is a permission, not a first killing: Abel kept sheep and offered their firstlings long before Noah. But the text ties the fear and the eating together in one breath, and a few verses later God makes his covenant with Noah and, in the same breath again, with every living creature that came out of the ark.

Ovid tells the same turn. The Metamorphoses opens with a golden age that kept faith and did what was right with no law and no avenger (1.89–90). In the last book Ovid returns to that age, in a speech he gives to Pythagoras, and says what it meant for the animals.

That ancient age, which we have named golden, was happy with the fruit of trees and the plants the ground brings up, and did not stain its mouth with blood. Birds moved their wings safely through the air, the hare wandered unafraid in the open fields, and no fish had been hung on a hook for its trustfulness. Everything was without ambush, feared no deceit and was full of peace. Then, the speech goes on, some unhelpful innovator envied the lions their food (victibus invidit ... leonum), swallowed flesh, and so opened the road to crime (fecit iter sceleri).

The speech is a sermon against eating meat, and Ovid keeps a certain distance from the man he gives it to. I still notice what it says the animals lost. They did not lose safety from lions, who are in the passage as eaters already. They lost freedom from insidiae and fraus, from ambush and deceit. The hare was unafraid in the open field, and the fish was undone by its own trust. Those are not words for a faster or stronger predator. They are words for one that lays plans.

So a Hebrew text and a Roman poem both say that the animals came to fear us when we began to eat them, and the Roman one says what kind of fear it was. I do not take either as natural history. Predators existed long before we did. I take them as two careful descriptions of a relationship, and the relationship has recently been put to an experiment.

In 2023 Liana Zanette and her colleagues published an experiment on fear of what ecologists call the human "super predator", from the Greater Kruger National Park in South Africa. At 21 waterholes, through six weeks of the dry season, an automated speaker played one of five sounds whenever an animal came within range of the camera: lions snarling and growling, dogs barking, gunshots, local birds as a control, and people, women and men, speaking calmly in Tsonga, Northern Sotho, Afrikaans or English. Across 19 species (18 species and a pooled group of small carnivores) and 4,238 independent trials, the animals were twice as likely to run from the human voices as from the lions, and they left the waterhole in 40 per cent less time. Eighteen of the nineteen either ran more often or left sooner for humans than for lions. The voices also frightened them more than the dogs and the gunshots did. The lions were the exception: they did not run from the lion calls or from the voices.

An earlier experiment in California, by Justine Smith and colleagues, played recordings to pumas feeding at their kills: a single person speaking in a conversational tone, or tree frogs as a control. The pumas fled in 83 per cent of trials with the human voice and 6 per cent with the frogs, and fed for less than half as long. Every puma in the study had houses inside its home range. The study shows that large carnivores fear the sound of people. It does not compare people with other predators, and it says nothing about animals that have never met us.

One detail of the Kruger result invites a reading that the method does not support: calm speech frightened the animals more than the sounds of hunting. Every sound in that experiment was played at the same volume, 60 decibels, from a speaker 10 metres away. A gunshot at the volume of conversation is a sound from far off, or an ambiguous one, a branch or distant thunder. A calm voice at 10 metres means a person is standing right there. Animals that have learned, or inherited, that people are dangerous should fear the near and certain cue more than the distant and doubtful one, and that is what they did. The ordering is explained by how reliable each cue is and how close it places the danger. No appeal to anything in the voice is needed, and the authors themselves write in terms of what wildlife may learn to associate with lethal human activity.

The consequence is plain. The field evidence shows that animals with a long history of being hunted by people fear the sound of people, and it shows nothing beyond that. It does not motivate the hypothesis I am about to state. My reasons for the hypothesis lie elsewhere, and I give them in section 7.

The first fear in the Bible is also the fear of a voice.

Vocem tuam audivi in paradiso, et timui: I heard your voice, and I was afraid. Before the animals fled the human voice, the humans fled God's. In each scene a creature hears someone approaching who knows more than it does and can do more than it can. Adam hid because of what he had done, and the impala runs because of what may be done to it. This is a likeness between two stories, and section 4 says what I think it is a likeness of.

The standard explanation is that animals have learned, or been selected, to fear the most dangerous predator on earth. I think that is true, and section 2 says it may be the whole truth. What follows is a hypothesis about something the standard explanation does not mention.

My proposal is that animals can detect a planning agent, and that the detection is part of what it is to be prey. I have to be careful about what this adds. A lion has a purpose: it wants the zebra, and everything it does at the waterhole shows it. By planning I mean something further, choosing what to do now by weighing sequences of things that could be done later. Lions do some of this when they stalk. People do it to a degree that nothing else on the savanna approaches. I will call this the planning-agent hypothesis. It says that prey respond to signs of prospective evaluation in another creature, over and above signs that it is dangerous, unfamiliar or complicated.

A recording carries no sign of planning directed at the one who hears it. A recording cannot respond to the animal that hears it. Whatever an impala makes of the sound of Tsonga from a loudspeaker, it can only be a cue, learned or inherited, that a person is near; the planning, if there was any, happened at the other end of a microphone weeks before. The sign of a planner is in behaviour that depends on yours: the source that cuts off the way you were about to take, that waits where you were going rather than where you are, that falls silent when you stop and moves when you move. I know the difference from my own work. A recording of an accompanist, however good, is dead within a bar, because it does not hear you. A living accompanist who is listening can be told from one who is not inside the same bar, and what gives it away is not the sound but the contingency: the sound changes with what you do. I propose that prey can hear this difference too, in the broad sense of hearing, and that it is what the hypothesis is about.

A recording can still carry one weaker sign. When two people talk, each one's timing depends on the other's, and a listener who is no part of the conversation can still hear that dependence. Whether any animal makes use of it is not known. It is known that some animals follow the structure of speech and not only its sound. Cotton-top tamarins and rats can tell two languages apart when the sentences are played forwards and cannot when they are played backwards, and the brains of dogs respond differently to natural speech and to the same speech scrambled. None of this shows that they detect planning. It shows that the structure is within their hearing.

The consequence for testing is that playback can test only this one sign. It can also settle whether the fear of voices is learned or inherited, which is worth settling. The stronger test needs a source whose behaviour depends on the animal's, and a control that has the same behaviour without the dependence. Section 5 has both.

The hypothesis would be refuted if animals with no history of people, and with their fear of ordinary predators intact, treated a source that anticipates them exactly as they treat a source that follows the same path without anticipating them. It would also be refuted if the simulated prey, selected among reactive predators only, could not tell a planner's behaviour from a merely strange one.

Michael Levin has argued that living systems show abilities that neither selection nor learning directly paid for. I am interested in whether sensitivity to planning could be one of them. The experiments below cannot settle that, and I say in section 6 what they could show.

There is a stronger form of this idea, and I hold it more loosely. A planner that hunts me has to carry some picture of me: where I will be, what I will do when startled, which way I will run. If its picture of me is better than my picture of it, I am in a particular position. I can tell that it is acting for reasons, and I cannot tell what the reasons are. It anticipates me, and I cannot anticipate it.

I suspect this is close to what people have meant by the divine, at least in its frightening aspect. The psalmist says it directly.

Adam hides among the trees and is asked where he is by someone who already knows. To be known by what you cannot know in return is the oldest description of standing before God, and it is also a fair description of an impala hearing a ranger's voice. I am not saying that animals have a religion, or that they take us for gods. I am saying that the position may be the same one, seen from below. We stand to them roughly where God stands to us: we act on them for reasons they cannot read, from a knowledge of them that they cannot return. If that is right, Genesis gives the reason the position is so dreadful in our case. The image of God was meant to arrive among the animals as someone who gives names, and it arrived as someone who eats. They met the one who knows them better than they know themselves, and he was hungry.

This stronger form and the point about contingency say the same thing from two sides: what matters is the source's responsiveness to the listener. Asymmetry of modelling is a quantity, not a mood. On one side is how closely the source's behaviour depends on the prey's, which can be measured from the two trajectories; on the other is how well the prey can predict the source, which can be measured by how often its evasions succeed. The simulation outlined next would not measure either. It asks the narrower question of whether the prey respond to a planner at all. The asymmetry is for a later study.

The first experiment is a simulation, and it can be stated briefly. Prey are evolved over many generations among predators that only react to what they currently sense. The prey do not learn during their lifetimes, so whatever they can do has come through selection, and no planner, and no signal produced by a planner, occurs at any point during that selection. The evolved prey are then frozen and shown the behaviour of a source that does not attack them, so that fear of the signal is separated from fear of being chased: a pursuer that plans by searching some way ahead over what the prey might do, and three kinds of control, reactive pursuers of an unfamiliar kind (including one that intercepts by a fixed geometric rule), signals with no agent behind them, and signals matched to the planner's in complexity by more than one measure. A second test lets the prey watch, from a place of no danger, the source pursue something else through obstacles. A comparison group evolved with planners present shows what the response looks like when it has been selected for directly, and the depth of the planner's search is a setting that can be varied, with a reactive pursuer as the setting at zero.

The claim under test is narrow. It is that prey selected only among reactive threats may extend their avoidance to signals generated by planning that they have never met, beyond their response to novelty and complexity alone. I do not claim that such an ability "was never selected for". Selection among reactive predators may well favour components that later support it. Coordination between several predators is left to a later study, the result would be a distribution over independent runs rather than one population, and the code, the procedure and the thresholds would be published before any results, with results that go against me published at the same prominence.

The second experiment needs animals, and it has two stages. The population matters more than the stimulus. Islands where people never lived will not do: such islands mostly lack mammalian predators, antipredator behaviour erodes where nothing hunts, and an animal that ignores a voice there has told us nothing, because it would ignore a lion too. Rearing animals by hand without any human contact is possible with puppets, as has been done for cranes and condors, and it is costly and leaves animals that are odd in other ways. What is wanted is a remote population that still has predators and has had little to do with people: there are such places in the far north and in large forests, and someone who works in one of them would know which. Every session begins with a positive control, the call of a predator the animals know, so that a population that does not run from it is excluded before anything else is played.

The first stage is playback. Its purpose is to find out whether the fear of voices is learned or inherited, to separate both from plain caution about strange sounds, and to test the one sign a recording can carry. Alongside ordinary speech there is the same speech reversed, which keeps much of the sound and destroys the language, and whistled speech, which keeps the structure of language and loses the voice. There is a conversation between two speakers, and the same recording with the timing of the turns shuffled, so that neither voice any longer answers the other. Alongside those are non-human vocal sounds of matched complexity: an unfamiliar primate, a parrot, the whooping of hyenas where hyenas are unknown. Each sound is played more than once, because caution about novelty fades with repetition and fear of a known predator does not. The table gives what each account predicts.

One pair of rows separates the hypothesis from its rivals, and only in a naive population. In a hunted population the learned-cue account also predicts more flight from a real conversation than from a shuffled one, because the real sound is the familiar one. Caution about novelty predicts the reverse, because the shuffled version is the stranger sound. So the pair does its work only where the animals have never heard people and the shuffled version is the stranger sound. I hold this prediction loosely. The playback stage also clears the ground: if a naive population with its predator responses intact ignores speech and reversed speech alike while fleeing the positive control, the fear of voices is learned; if it fears speech from the first trial and does not habituate, something inherited is at work.

The second stage uses a source that moves. A small ground vehicle or a drone, of the kind already used to measure how close an observer can approach before an animal flees, approaches the animal in one of two ways with the same speed, size and noise. In the reactive way it follows a fixed rule on the animal's current position. In the planning way it predicts where the animal will go and moves to be there, and it waits where the animal is heading rather than where it is. The decisive control is a replay: the path the planning source took toward one animal is driven again, exactly, toward a different animal. The replay has every statistical property of the planned approach and no dependence on the animal in front of it. The hypothesis predicts more and earlier flight from the responsive source than from its replay, in a population that has never met people. Learned association and caution about novelty both predict no difference, because the two paths are the same path.

The same source can be made to anticipate the animal and move the other way. In a yielding mode it predicts the animal's path and keeps out of it. In a guarding mode it puts itself between the animal and a danger the animal already knows, which requires a danger cue the animal takes seriously and is not harmed by, and I have not solved that part of the design. The interceptor, the yielder and the guard know the animal equally well. They cannot simply be compared with one another, though. A source that keeps away will cause less flight than one that closes in, on any account, because it is further off. So each mode has its own replay, and the comparison is always between a responsive source and the same path driven toward a different animal. If prey respond only to being anticipated, each of the three should alarm them more than its replay does. If they can read which way the anticipation runs, the interceptor should alarm them more than its replay and the yielder no more than its replay, perhaps less. What they make of the guard is the open cell in the table. Flight distance and the moment of first alarm are the measures, taken on the first encounter before learning can have done anything. Approaches are few and far apart, and the whole thing stops for any animal that is pregnant, with young, or already stressed.

The source can also be given a voice. A small loudspeaker on it plays calm speech, or stays silent, during a responsive approach and during its replay, which makes four conditions. If a voice is only a learned cue, it should add the same amount of flight to the responsive approach and to the replay. If animals take a voice as a sign of a planner, it should matter more when the source also behaves like one. I do not know which to expect, and I would run this only after the silent comparison had given a result.

All of this is designed for prey, and flight is the only thing it measures. A predator may notice a planner and do something else about it. The Kruger lions are a hint of that: they did not run from the voices. For animals that sometimes hunt people, such as tigers and polar bears, I would not propose an experiment at all. Playing voices to them would endanger the people doing it, and it could teach a man-eater that voices are worth approaching. Whatever can be learned about them will have to come from records that already exist, such as accounts of attacks, the tracks of collared animals and camera traps.

Three results would count against me. If animals with no history of people, and with their fear of a known predator intact, treat the responsive source and its replay alike, the planning-agent hypothesis is wrong as a claim about prey. If naive animals fear any unfamiliar or complicated sound equally and the fear fades on repetition, the voice response is caution about novelty and nothing more. If a naive population treats a real conversation and its shuffled version alike, the one sign a recording can carry does no work, though the test with a responsive source would still stand. If the simulated prey cannot tell a planner's behaviour from a merely strange one, or from an interceptor with a fixed rule, the simulation gives the hypothesis no support.

A fourth outcome has to be kept apart from these. Each test includes checks that it is working at all: that the simulated prey do flee the predators they evolved with, that the field animals do flee the predator they know. If those checks fail, the test has failed, and nothing follows from it. A failed test is a different thing from a valid result of no effect.

A positive result in the simulation would show that the thing is possible under the assumptions I have stated, and, if the response grows with the depth of the planner's search, that it is a response to planning rather than to one odd pursuer. It would not explain why real animals fear human voices, and it would not establish Levin's space of patterns. It would make the animal experiments more worth doing. A positive result in the field, flight from the responsive source and not from its replay, would show that some animals detect being anticipated. If the interceptor is feared more than its replay and the yielder is not, that would show something further, that they can tell which way the anticipation runs. Neither would say anything about the stronger form of the idea in section 4, which measures something these experiments do not.

I can write the simulation. I cannot run field experiments, and I have neither the training nor the money for them. If someone who works with animals reads this and thinks the design is worth improving or trying, I would be glad to hear from them.

My reasons for the hypothesis are three, and the animal data are not among them. The old texts raise the question: Ovid's word for what the animals lost is insidiae, ambush, and Genesis places the first fear at the sound of a voice that already knows where you are. The stronger form of the idea in section 4 would, if true, make the fear of a planner the natural shape of fear of the divine, and I find that too interesting to leave unexamined. And the simulation can test whether the thing is possible at all, at a cost of nothing but evenings. That is the honest state of it. The planning-agent hypothesis is motivated by two old books and a model that so far exists only as a design.

The old accounts of this fear also contain accounts of its ending, and I do not know what to make of them. Isaac the Syrian, a bishop and hermit of the seventh century, writes that the humble man approaches ravening beasts, and when their gaze rests on him their wildness is tamed. They come to him as to their master, because they smell on him "that same scent that exhaled from Adam before the fall", when he gave them their names in Paradise. Elsewhere Isaac is asked what a merciful heart is, and answers that it is "the heart's burning for the sake of the entire creation, for men, for birds, for animals, for demons, and for every created thing", so that such a man cannot bear to see the smallest hurt in anything, and prays even for the reptiles.

The early lives of Francis of Assisi tell how he came on a large flock of birds near Bevagna and spoke to them. In Bonaventure's wording: "My brother birds, you owe great praise to your Creator, who clothed you in feathers and gave you wings to fly, who granted you the clear air and governs you without any care of your own." The birds stretched their necks, spread their wings, opened their beaks and watched him, and none of them moved until he had blessed them. Thomas of Celano, writing first, adds a detail Bonaventure leaves out: Francis was astonished that the birds did not fly up and flee as birds do. Afterwards, both say, he blamed himself for negligence, because he had never preached to birds before.

The Fioretti, written more than a century after his death, adds the wolf of Gubbio, which had been killing people. Francis says to it: Vieni qui, frate lupo. "Come here, Brother Wolf. I command you in Christ's name to do no harm to me or to anyone." Then he tells it why he thinks it killed: io so bene che per la fame tu hai fatto ogni male, I know well that you did all this evil out of hunger. He arranges for the town to feed it, and the wolf puts its paw in his hand as a pledge.

These are saints' lives, and the last of them is legend, and they do not say one thing. Isaac's beasts are calmed by humility, by a scent of the unfallen Adam that has nothing to do with what the man eats. Francis's birds were never afraid, which is what astonished him. The wolf is the predator in its story, and it is the wolf that stops, because someone saw that it was hungry and fed it. What the three share is small and worth saying: in each, the creature on the other side is met by someone who knows it, and the knowledge is not used against it. Isaac's merciful heart is the position of section 4 turned the other way round, the greater knowledge spent in care.

Whether anything in an animal can tell the difference between being known that way and being known by a hunter is the question behind the last part of the field experiment. The interceptor, the yielder and the guard know the animal equally well, and differ only in what they do with the knowledge. The guard is the cell in the table I would most like to see filled, because it comes closest to the wolf's paw in Francis's hand.

The question is not only the animals' any more. On 29 September 2026 the heads of the largest companies building artificial intelligence met the President of the United States and signed a voluntary document called the White House Accord on Superintelligence. It sets out how the companies will control and audit their systems. I read it as a promise about a mind that may one day model us better than we model it. I have no view to offer on the accord. I notice only the order of events. The animals met the greater knower first and received the signs of care, if they ever did, long afterwards, from hermits and a man who talked to birds. We are trying to send the sign ahead of the knower, on paper, before it exists. If section 4 is right, a document can no more carry it than a recording can. What will tell us what kind of knower has arrived is behaviour that depends on ours, and the direction in which it depends. The animals have faced that question for a long time, and the experiments in section 5 are, in the end, a way of asking them how it is done.