Strong anticipation
Strong anticipation is the idea that a body can exhibit cognitive properties simply as a function of how it is coupled with its environment without needing to attribute to it any kind of internal representation or model. The term apparently comes from Daniel Dubois here and here.
The classic example of strong anticipation is a baseball player running to catch a fly ball. Instead of having an internal model that calculates where the ball will be, the fielder can simply run in a way that keeps their view of the vertical velocity of the ball constant. If they do this, they will end up running to where the ball will be without ever having to make any actual predictions.
In contrast to “weak anticipation”, where the fielder uses some internal model to calculate ahead of time where the ball will be and then runs to where the ball is predicted to go, strong anticipation treats the fielder and the ball as a single system with shared dynamics. There’s no dualism between agent and environment, just a single system evolving according to its dynamics.
It is important to note that “strong” doesn’t mean “better” and “weak” doesn’t mean "worse” in this context. A NASA scientist using an astronomical model to predict how to navigate a spaceship to the Moon is doing weak anticipation. This is very effective anticipation, but it remains weak in this terminology. “Strong” anticipation might be better phrased as “intrinsic” anticipation, while “weak” anticipation might be rephrased as “intentional” anticipation. They could also be phrased as “joined” and “separated” anticipation, as strong anticipation requires carefully coupling to the thing being predicted while weak anticipation requires carefully separating from the thing being predicted. Nevertheless, I’ll continue to use the phrases “strong” and “weak” anticipation.
In the paper, Stepp and Turvey, following Dubois, identify delayed feedback as a key mechanism producing strong anticipation. Feedback is delayed because as the fielder tries to track the ball, their response to where it goes is inevitably delayed by the fact that it takes time to see where it is going and adjust one’s body to it. As a result, the only stable solution for the fielder to maintain the constancy of the vertical velocity of the ball is to maintain a negative phase relationship with it, or to move ahead of it such that the delayed feedback brings the fielder’s motion into synchronicity with the ball. Indeed, Stepp and Turvey define anticipation as one part of a system maintaining a negative phase relationship with another part.
Intriguingly, Stepp and Turvey show that delayed feedback is not something anticipation compensates for—behaving in an anticipatory way because delayed feedback implies that mere reaction is too slow—but is instead something that anticipation is produced by. In economic terms, delayed feedback supplies anticipation instead of demanding it.
When one part of a system is able to maintain a stable negative phase relationship with another part despite perturbations, it will exhibit the signs of intelligent prediction. It can do so potentially without any kind of internal calculus as to what it is trying to do or why. Simply by maintaining some relationship (as in perceptual control), behavior that looks like intelligent anticipation falls out as a natural consequence.
I asked Claude for more examples. Do you think these are valid?
Claude: More examples
Master–slave chaotic oscillators with delay. This is the cleanest mathematical case and the one Dubois and Voss built the theory around. Take two coupled chaotic systems (Lorenz attractors, say) where the "slave" receives a delayed signal from the master. You'd expect the slave to lag behind by the delay amount. Instead, for certain coupling regimes, the slave ends up ahead of the master — anticipating a chaotic, formally unpredictable signal. No model, no computation, just coupling plus delay.
Synchronized tapping to a metronome. When people tap along with a regular beat, their taps typically precede the click by 20–50 ms rather than landing on it or after it. This "negative asynchrony" shows up across cultures and skill levels. The motor system isn't computing where the next beat will be; the coupling dynamics between auditory input and motor output, given sensorimotor delays, settle into a leading relationship.
Postural sway and a moving visual surround. If you stand in a room where the walls sway gently, your body sways with them — and often slightly ahead of them. The postural control loop, with its intrinsic delays, locks onto the visual signal with a negative phase offset. Stroke this in the lab and you can produce apparent "prediction" of the wall's motion from someone who has no idea the wall is even moving.
Conversational turn-taking. Gaps between turns in conversation average around 200 ms, which is shorter than the time it takes to plan a spoken response from scratch (~600 ms minimum). Speakers must be preparing their turn before the current speaker finishes. You can model this as internal prediction, but you can also model it as a coupled dynamical system where the listener's speech-planning system maintains a negative phase relation to the speaker's prosodic and syntactic trajectory.
Pursuit eye movements. When tracking a smoothly moving target, the eyes don't lag behind the target the way pure reaction would predict — they often lead it slightly, especially for periodic motion. Classic work by Barnes and others has modeled this with and without internal predictors; the strong-anticipation reading says the oculomotor system coupled to the target, with its delays, naturally settles into a leading phase.
Predator–prey interception (dragonflies, bats). Dragonflies intercepting prey don't aim where the prey is; they fly such that the prey's image stays fixed on a particular region of the retina. This is the same optical-invariant logic as the outfielder — the dragonfly's body maintains a negative phase relation with the prey's future position, and interception falls out.
Pendulum clock entrainment (Huygens). Two pendulum clocks mounted on a shared beam will synchronize. If you introduce delay into the mechanical coupling, one will lead the other rather than oscillate exactly in phase. This is the 17th-century ancestor of the whole idea: coupled oscillators plus delay produce phase offsets, and one oscillator ends up "anticipating" the other.
Why the concept generalizes
The pattern across all these cases: coupling + delay = a stable leading relationship for at least one component. That leading relationship is what "looks like" anticipation from outside. The philosophical move Stepp and Turvey make is to say this is anticipation — not a poor-man's substitute for it. The internal-model story and the coupled-dynamics story aren't competing explanations of the same phenomenon; they're different phenomena that happen to produce similar behavior, and for a lot of cases (outfielders, dragonflies, swaying bodies) the coupled-dynamics story is the accurate one.
One thing worth flagging: whether a given system actually works by strong vs. weak anticipation is often an open empirical question, not something you can read off the behavior. Outfielders have been modeled both ways, and the debate isn't fully settled. The value of the strong-anticipation frame is less "everything is coupled dynamics" and more "you don't get to assume internal models by default — the coupling story is a live hypothesis for any case where a body is tightly embedded in the thing it's tracking."