En 30 secondes
Human language builds on the ability to pursue shared goals and understand another person’s perspective, argues developmental psychologist Michael Tomasello. In experiments he describes, children who could no longer hear each other invented shared gestures; language models instead begin with language humans already created. He does not rule out a different route to capable AI, but suggests studying how agents coordinate before they form conventions. His account also distinguishes basic goal-directed control from planning, self-monitoring and joint agency.
- 1Unpredictable situations call for flexible decisionsTomasello contrasts built-in routines such as breathing with decisions about unpredictable prey, predators and other agents.
- 2Planning requires a layer that monitors actionHe proposes executive processes that anticipate mistakes and represent situations that do not yet exist, with a further reflective level in great apes.
- 3A squirrel can reject a jump before attempting itTomasello interprets flexible squirrel behavior as planning with imagined outcomes, a proposed step beyond basic control.
- 4Shared agency means a goal and roles held in commonHuman partners can monitor each other’s contributions and attend to obstacles together, he argues.
- 5Children create gestures when speech is removedIn a video-call study, children invented gestures and adapted them into conventions, unlike models trained on language already made by others.
- 6Coherent discourse may carry information models can useTomasello says language models handle perspective better than he expected and wonders how they would perform if sentence order were randomized.
- 7A Turing test is not enough to establish agencyHe remains unsure whether biology is necessary, but says a machine would need much more than conversational imitation.
À retenir
- Tomasello argues that a common goal, shared attention and complementary roles make collaboration more than agents reacting to one another.
- Children can invent shared gestures when speech is unavailable, illustrating how conventions help people coordinate.
- A language model may learn from coherent human discourse without having developed the social capacities that produced it.
- Distinguish a basic control loop from planning, seeking information about a plan and cooperating toward a shared goal.
- A Turing test alone would not establish machine agency, in Tomasello’s view.
Lire le texte remanié, chapitre par chapitre
Why uncertain environments call for agents
Tomasello connects flexible decision-making to uncertainty. He contrasts active goal pursuit with fixed stimulus and response.
When environments are unpredictable, organisms need a way to assess a situation and choose an action. Developmental psychologist Michael Tomasello uses this idea to examine how agency evolved and how humans came to coordinate with one another.
Tomasello says his interest began when he observed chimpanzees at Emory and saw similarities and differences with two-year-olds. At the Max Planck Institute, his team put children and great apes in comparable experimental situations to find subtle distinctions in their social and cognitive abilities.
Predictable, essential activities such as breathing can be built into an organism. In unpredictable situations, Tomasello says, evolution favors a mechanism that assesses circumstances and makes a flexible decision. Other agents, including predators, prey, competitors and potential collaborators, are particularly hard to predict.
Tim asks whether agency involves both acting for reasons and being able to bring about a desired outcome. A control loop with a target is a minimal version of that idea; more complex agents appear to have additional layers of control.
Tomasello uses a thermostat as a concrete analogy, but says a living agent is more than a stimulus-response machine. An organism pursues goals and tries something else when a strategy fails. Some animals can inhibit an action, seek information or reconsider their plans.
Such control does not require freedom from biology. An animal pursuing prey has a space for choices within its biological adaptations, and a control system illustrates how perception, goals and action can be organized without a separate little person inside the brain.
What a traffic light reveals about control
A smart light needs perception, a goal and a way to choose actions. A thermostat illustrates part of that organization without being a full agent.
Tomasello asks students how to make a timed traffic light smart. It needs to perceive cars, have a goal for traffic flow and decide how to change the lights based on the difference between what it sees and what it wants to achieve.
The complexity of a spider's web or an ant colony does not itself establish flexible control. Tomasello says organisms can succeed in evolution through specialized adaptations; what matters to his account of agency is adjusting action to the situation.
Tim wonders whether a thermostat or robotic vacuum should count as an agent. Tomasello agrees that he would not call them agents: they show only the basic organization he wants to explain. Their value is that they make the control mechanism concrete without invoking a mysterious inner decision-maker.
Could a machine acquire stronger agency?
Tomasello considers layers for planning and self-monitoring. He remains uncertain about current AI systems and suggests using animals as models.
Tim asks whether machines could ultimately have more agency than humans, raising the limits of optimization as an analogy. Tomasello is unsure; older AI imagined a computational brain separated from perception, goals and action, while drones and self-driving cars now resemble parts of an agent's organization.
One missing element may be an executive layer that observes ongoing attention and action, anticipates mistakes and plans. Tomasello calls this a different level of organization rather than a separate skill. Planning and hypothesis testing require representing situations that do not yet exist.
He proposes another layer in great apes: reflecting on a plan, asking whether more information is needed and assessing how resources are being used. A machine with perception, goals and executive monitoring might have a stronger claim to agency.
Human basic goals are also shaped by biology, so goals supplied by designers do not automatically disqualify a machine in his view. He would want it to formulate new goals when it encounters unfamiliar situations; he sees studying natural agents as a way to explore this.
Tim asks whether unexpected AI capabilities can emerge from components not designed for them. Tomasello says he lacks the technical knowledge to judge particular AI systems. In evolution, a trait like an opposable thumb can serve one function and later be used creatively in a new context.
Learning and inherited influences are parts of a larger system
Learning serves adaptive decisions in Tomasello’s account. Biology and culture influence choices without resolving every conflict between goals.
He thinks AI research can place too much emphasis on learning by itself. Learning, in his account, serves an agent's adaptive decisions. Language models start with a sophisticated human achievement, language, and he is uncertain whether the abilities that support it can all be skipped.
He compares the problem to the old debate over a special substance for life. Living and nonliving things are made of the same basic elements, he says, but organized differently. The important question for agency is likewise how the pieces are organized.
Tim asks where goals come from if genes or social forces shape us. Tomasello treats biology and culture as powerful influences rather than complete explanations of an individual's choice. They can pull in different directions and leave an agent to decide between them.
A taste for sugar does not force someone on a diet to eat ice cream, he says. Similarly, a promise may conflict with the need to help an injured child. He uses these cases to describe influences placing weight on a decision without settling every conflict in advance.
Specialized motivations meet general decision processes
Tomasello accepts specialized predispositions while disputing the idea that all thought consists of separate modules. Emotions influence choices without always determining them.
The conversation turns to evolutionary psychologists' proposal that the mind has specialized modules, such as different mechanisms for selecting a mate and finding food. Tomasello accepts that specialized tendencies exist, but disputes the stronger claim that everything is modular.
Perception, attention, memory and learning can operate across domains, he argues, and a decision-maker must arbitrate when specialized motivations compete. Along the lineage leading to humans, he sees mammals as making flexible decisions under the influence of those motivations.
He describes emotions as evolution putting a finger on the scales: fear of a snake can interrupt other goals, although another urgent threat might change the response. Tim relates this to Mark Solms's view of feelings as a way to navigate uncertainty.
Why he avoids a single measure of intelligence
He argues that school-oriented definitions can overlook animals’ specialized abilities. He prefers naming the particular cognitive capacities under discussion.
Tomasello avoids treating general intelligence as a single obvious capacity. He connects the term's history to school-oriented tests of verbal and abstract ability, and points to squirrels' memory for food locations, birds' sound discrimination and chimpanzees' tool use as forms of specialization.
Tim asks whether agency might be just as oversimplified a label. Tomasello replies that a control system gives him a physical organizational model, whereas he finds it harder to point to a comparable mechanism for general intelligence. He prefers specific terms such as planning, learning and cognition.
From basic control to reflection on plans
Tomasello proposes successive layers of agency along the line to humans. Reptiles, mammals and great apes serve as comparative examples.
Tomasello outlines levels of agency along the evolutionary line to humans, not a complete classification of every animal. Because extinct ancestors cannot be studied directly, researchers use living model species while recognizing that they are not identical to those ancestors.
He treats reptiles as models for relatively basic goal-directed control. With mammals, he proposes an executive layer that can represent possible outcomes and plan. A squirrel appearing to decide against a dangerous jump and taking another route illustrates the kind of behavior he wants to explain.
Planning lets a potential strategy fail in imagination rather than in action. Tomasello points to squirrels inventing ways around obstacles at bird feeders as examples of flexible behavior, though he presents the internal organization as his interpretation.
A further level is metacognitive or rational agency, which Tomasello says is evident in great apes. They can seek additional information before committing or opt out of a difficult problem in favor of an option more likely to succeed.
He links these changes to different sources of uncertainty: solitary reptiles encounter unpredictable prey; group-living mammals compete for resources; great apes may also have to outmaneuver one another. The evolutionary story is his proposed explanation for the observed differences.
Shared goals create a different form of agency
Human collaborators can coordinate roles, attend jointly and consider contrasting perspectives. Cultural conventions both constrain and empower members.
Humans, he argues, add a distinct form of shared agency. Partners can adopt a common goal, understand their separate roles and correct one another when a role is not being fulfilled. That is different from merely reacting to another agent.
Cooperation also requires joint attention: both participants notice an obstacle or opportunity and know they are attending to it together. Tomasello connects this to perspective-taking, the ability to understand that the same object can be seen as a dog, an animal or a pet.
He argues that contrasting another person's view with one's own helps create the notion of perspective. In a thought experiment, a child growing up entirely alone would encounter the world but lack another person's simultaneous, contrasting interpretation.
Cultures extend shared agency through language, conventions, norms and institutions. These constrain members, but also give them resources they did not have to invent themselves. Once someone learns a language, for example, they can use it to create poetry.
Tomasello cites experiments comparing children and chimpanzees in collaborative tasks. He says children may continue helping until both partners receive a reward and may correct a partner's role, whereas the chimpanzees in his examples stop after obtaining their own reward.
Children invent conventions before models use language
Children in a study developed shared gestures when audio was removed. Tomasello suggests investigating coordination as a foundation for artificial agents.
He uses language models every day, but distinguishes using an existing language from creating linguistic conventions. In one study, children communicating over video with the sound removed invented gestures and adapted them to one another until they shared conventions.
Tomasello thinks shared attention and commitment may be the infrastructure that makes linguistic conventions possible. He does not rule out reaching useful language by another route, but suggests building from simple coordination toward language alongside work on mature language models.
Tim asks whether AI agents could eventually create conventions through social interaction. Tomasello points to research modeling joint commitment and suggests that early coordination challenges may have required each participant to consider what the other expects.
In the stag-hunt example, each person can pursue a smaller reward alone or a larger one only if the other joins. Each must decide whether the other will participate. A convention can resolve such coordination problems by giving both parties a shared expectation.
Perspective shapes words, grammar and pointing
The speakers consider how another person’s view changes what one says. Tomasello also considers the role of coherent discourse in language-model training.
Tim asks about knowledge gained from multiple perspectives. Tomasello invokes the philosopher Donald Davidson: another person's view can show us we were wrong in a way that differs from discovering our own mistake. He sees social engagement as important to developing this capacity.
The same animal might be called a dog, a pet or simply 'she' depending on the listener and the common ground. The choice is not just a taxonomy of objects; it adapts the speaker's description to what matters in the situation.
Grammar also carries perspective, Tomasello argues. Saying 'it was Fred who broke the window' can correct someone who thinks Mary did it; active and passive constructions can present the same event from the vantage point of the dog or the cat.
Language models seem to handle some perspective shifts better than he expected. He wonders what would happen if training sentences were randomized so that the logical and causal links between them disappeared; he reports an AI workshop's expectation that the result would be much worse, while acknowledging that nobody knows.
Tim suggests that human storytelling follows an agent's goals and obstacles. Tomasello returns to the abilities children develop before speech: a pointing gesture only communicates a specific meaning when speaker and listener share enough context to infer the intention.
A point might mean 'there is your sister' or 'the library is open,' depending on what both people know. Tomasello cautions that it may not be fair to say a language model cannot interpret a point, but proposes studying shared context and simple coordination as a route toward more agentic machines.
Institutions and deception rely on shared expectations
Money and offices depend on human recognition, while cooperation can be exploited through deception. Group loyalty can also divide people.
Tim asks about risks as humans create more abstract shared institutions. Tomasello says money, offices such as president and corporations depend on shared recognition. A banknote can cease to function as money when the social agreement changes, even though the paper remains.
Such institutions can have real effects despite relying on human-created categories. Tomasello then argues that politicians can exploit people's tendency to look for what others believe, using repetition and assertions of broad agreement; he gives Donald Trump as his example.
Lying works, in his account, because people normally expect cooperation when someone has no apparent reason to deceive. He contrasts that expectation with his description of chimpanzees and says a cooperative disposition can be exploited.
Cooperation within a group can coexist with conflict between groups. Tomasello argues that humans evolved strong attachments to a familiar 'we' and now face the challenge of extending shared agency in an interdependent world of many groups.
Are markets agents, and what would make a machine one?
Tomasello treats market outcomes as emerging from individual decisions and is cautious about calling capital an agent. He says a Turing test is insufficient to establish machine agency.
Tim asks whether capital or markets themselves should be understood as agents. Tomasello prefers to describe people as making decisions whose interaction produces emergent effects. A traffic jam can result from everyone trying to get home quickly even though nobody intended the jam.
He gives inflation and market crashes as other examples of collective outcomes arising from individual choices. Policymakers, he argues, need to understand the link between choices and patterns when designing rules to reduce harmful outcomes.
Tim finds it useful to analyze the market at the level of an agent. Tomasello accepts that as a way of speaking, but hesitates to say capital literally has a goal or adjusts its behavior when it fails to achieve one.
Asked whether a system could act as if it had agency without actually having it, Tomasello invokes Daniel Dennett's intentional stance and says he is unsure. A Turing test alone would be far from sufficient, but a machine displaying the full set of relevant capacities might count as an agent.
Whether biological materials, hormones and neurons are necessary remains unresolved for him. He recalls a discussion prompted by philosopher Ned Block in which the room reached no consensus on whether being made of 'meat' matters.
Sur la même chaîne
1:37:08EN
1:33:33EN
47:55EN
1:10:10EN
1:14:20EN
43:43EN