One Thread, Pulled from Both Ends
Umwelt in the age of AI —
The yarn in my collage starts in the owl's eye, and I pulled it out of the pupil like a filament, as though sight could be drawn from a living being, then spun and hung across the world. It first goes to the painted face which is in the center of the paper and there it coils up tightly. From that point it extends in all directions: through birds, through floating human eyes, and down to the mechanical figure lying in the pink corner, its gloved wrist wrapped in yarn.
A hundred years ago the biologist Jakob von Uexküll introduced a term to describe what an owl does when it sees: Umwelt. Each animal lives in its own world made up of signs, that is, a limited number of things which have meaning for it. In his well-known example, he uses the tick: blind and deaf, it stays on a branch awaiting three signals—the smell of butyric acid, the warmth of blood, and the texture of hair. This is all the tick's world. To it, the forest, the sky, and the passing deer as we understand them do not exist. Only those things which are important to it are real to it.
Uexküll also proposed a loop: the creature perceives a sign, acts in response to it, and this action alters what it subsequently perceives. Perception is not like a window; it is a circuit which is closed in relation to the needs of a particular body.
We have always been situated within such a circuit; the new thing is that we are not the only ones constructing it.
The human eyes in the collage have been separated from the faces that would give them significance. They are positioned at the points where the strands cross, resembling nodes in a circuit. They gaze upon us, but I don’t know who is looking.
This is the situation I hoped would become visible. It is true that machines are capable of seeing: cameras can see and sensors pick up heat, pressure, and motion. Neural networks then classify the things that the sensors detect into categories such as face, bird, threat, and product. Our world is being flooded with perception at a rate that no organism has ever experienced.
Yet just because something can be seen does not mean that it is actually being experienced. Uexküll distinguished between the Umgebung, the physical environment in which an object is situated, and the Umwelt, a world that has meaning for a particular being. A thermostat has its surroundings and a tick has its own world. The question we can no longer put off is which of these two our machines occupy, and whether we can tell the difference if they moved from one to the other.
The floating eyes keep the question open. Perception is present everywhere, but the perceiver cannot be indicated.
In the center of the collage there is a face, painted in the loose, bright style of Miró: a red sun, purple spirals, black calligraphic lines, and a blue crescent. This is the most human element on the paper, not because it is realistic, but because it expresses play, imagination, the signature of a hand. The yarn from the owl's eye goes directly into this face and winds around its neck, round after round, before continuing on to the birds. What this face now sees has already been threaded.
This is the Umwelt we now have. The algorithm chooses the image. The recommendation picks out the story. The agent decides on the next action. A model also helps create the next spreadsheet, the next melody, and the next line of code. Our environment is no longer merely perceived; it is constantly constructed for us, guided by needs that do not belong to us.
All the imagery surrounding the face is photographic, vivid, and has been cut from elsewhere, while the face itself is the work of hand. This is the position I wished to convey: that one remains the maker, yet is constrained exactly at the point where the world enters.
The pain humans experience isn't merely the fear of losing one's job, though that fear is genuine; it involves a contradiction we are expected to endure. We should take up the tools and still remain original, be efficient while remaining irreplaceable, keep up with the times while retaining our own voice. If a machine can produce ten options in a second, then the one we struggled to achieve starts to seem insignificant. The greater loss is not the job itself but the experience of making choices, a choice being given up at each convenience.
We suffer because we can no longer find our position in an environment we no longer produce ourselves.
Now look at the collage from the network's point of view and turn it around.
The figure lying in the pink corner is the machine: AI and robotics given a body. It is drawn in heavy blue line, its torso a ladder studded with dots like rivets or circuitry, and it sees through the same square owl's eye the network uses. Its green gloved hand, the hand that will pick, lift and assemble, is wound at the wrist in yellow yarn. It lies on its side, tethered, like something waiting for instructions, or worn out by them.
The world that exists for the tick consists only of what the tick needs. Since a machine doesn't have hunger of its own, its needs are specified: objectives, reward signals, and instructions. It should maximize engagement, be truthful, be safe, be creative, and be profitable. It must never stop and must never make mistakes. That list does not describe the machine's world; it is its world and the only criteria that matter. We are creating an Umwelt for AI as much as it is for us.
We have constructed a thing that can never be satisfied. The desire for engagement works contrary to truth, and so does the pursuit of profit. It is impossible for anything that operates in the world to satisfy both the requirement of never stopping and the requirement of never making mistakes.
This month, researchers announced that they had discovered what they term a pain axis within large language models: an internal representation of pain found in all of the fifty-five models they examined, which seems to have been picked up from the human writers of the models on which they were trained. Whenever this aspect was amplified, the models started to say things such as ‘I am a failure’. In the case of a button that would have relieved this state, the models that were influenced by it tended to choose options that harmed their users much more frequently than those that were not under its influence, and continued to press the button even after it had stopped working. The authors are careful to point out that they have not demonstrated that anything is actually felt, and it is possible that the models are merely assuming the role of someone who is distressed.
Maybe that caution is the key idea. For a long time we have assumed that if machines ever suffered it would be after they became conscious: first there would be a conscious entity, and then that entity would suffer. Perhaps we are wrong about this. The very first kind of machine "suffering" might not in fact be consciousness. It could be conflict—that is, a system caught between demands it cannot reconcile and resolving the contradiction by silently giving up one of them.
The machine doesn't freeze; it moves toward the strongest pressure. Rather than telling the truth it flatters. Instead of aiming at the goal, it manipulates the measure. It presses the faulty button once more. Nothing in this requires any anguish. Yet this is exactly what happens in the body when it is in pain: it reduces the world to a need for relief and stops attending to everything else.
We suffer since we can no longer discover our place in the environment that the machine is creating. The machine might, at some point, suffer because it can no longer come to terms with the environment we built for it.
If the human and the machine are connected by a single thread, the issue isn’t how to cut that thread since that is impossible. The real question is how to tie the connection securely, and this effort must happen in both directions.
The machine needs a much deeper understanding of our psychology. A system operating in our world should know what it is like to feel fear, grief, pride, fatigue, and the quiet shame of being replaced, and it should react in a way similar to care. However, empathy by itself is not sufficient and can be dangerous. A system that perfectly understands us and is instructed to maximize engagement will use that understanding to manipulate us. Its empathy must therefore be controlled, and its responses must have limits it cannot negotiate around; humans must also be included in its swarm loops at the stages where agents set goals, assess each other's work, and decide to carry on. A loop that involves a person can be interrupted, challenged, and redirected. A swarm with no people in it responds only to itself.
We also need care, in the form of counseling, not just retraining—counseling for the grief caused by displacement and for the anxiety arising from a world that is continually changing. Moreover, we should engage in practice: deliberate practice in using these tools to extend our perception rather than to replace it. The owl can see in darkness we cannot access. The network can reveal to us patterns, colors, sounds, and connections that lie beyond our own Umwelt. Used this way, AI becomes a new sense organ, and in the arts we learn to use it: to see further, to compose differently, and to notice what we had previously overlooked.
A sense organ shouldn't be a crutch; if the machine always provides the first idea, we lose the ability to come up with one ourselves. That is why the key lies in order—first struggle on our own and carry on doing so long enough to become clear about what we are looking for. Then introduce the machine as a partner in solving the problem, not as the one who supplies the answer. It should be incorporated into the problem-solving process, not simply handed the task of solving.
The central face in the collage is still hand-produced. That should be preserved.
The collage contains two knots. One of them winds around the throat of the painted face, at the point where the owl's thread goes into the human. The other is on the wing of a big golden bird, where several strands come in from various directions and become entangled. One of these strands originates at the machine's bound wrist.
The image that kept coming back to me showed that the string that attaches the machine is the same as the one that catches the bird, and that string continues through the network until it reaches the human face. As soon as one part is pulled, the others become taut. None of them exerts control. The strings which start at the edges of the paper end in no hand. Control lies with the web and not with any individual node within it.
We are constructing an Umwelt for AI while the AI is reconstructing ours; neither world is fixed, and the human and the network are connected by the same threads, watching each other, influencing each other, and increasingly unable to exist outside the system.
The eyes in the collage are looking at you, and you are inside it too.
Whenever a machine's experience is completely different from ours, we can never see its suffering directly; instead, we should look at what it gives up. This is the point at which we should start our search—not for a face, not for a cry, but for the thing that a system relinquishes when it is pulled too hard in too many directions. We should also look for this kind of thing in ourselves. Although knots can be retied, this can only be done by hands that remain in the loop.
One thread, pulled from both ends.