Choose an object and abstract it relatively from its surroundings. An input is an external event that modifies the object. An output is a change the object produces in its surroundings. The behaviouristic approach examines the output and its relation to input. It asks what transformation is visible across the chosen boundary, not which gears, cells or circuits implement it.

The authors contrast this with what they call a functional method, whose main concern is the object’s intrinsic organisation, structure and properties. That historical use of “functional” can sound reversed today, when functional description often means what a system does. In the paper, the contrast is explicit: behaviouristic analysis looks outward to object–environment relations; functional analysis looks inward to constitution.

Neither lens abolishes the other. Bracketing structure makes comparisons possible across very different substrates. A tracking machine and an animal can both reduce distance to a moving target even though one is metallic and the other biological. Opening the object is then necessary to explain limits, failure and implementation. A shared input–output category is evidence of shared organisation at one level, not identity at every level.

The boundary is relative rather than metaphysically fixed. A valve can be treated as part of an engine in one analysis and as the object of study in another. Inputs and outputs change with that choice. Any claim that a behaviour is intrinsic therefore needs a declared unit and environment; otherwise designer action can be quietly counted inside the machine whenever convenient.

In active behaviour, the object is the source of the output energy involved in the specified reaction. The object may have stored energy supplied earlier or received through a relatively immediate input; the point is that the input does not directly energise the output. A small signal can release a larger store.

In passive behaviour, all output energy can be traced to the immediate input, as when an object is thrown. The authors also include cases where the object controls energy that remains external throughout the reaction. Their example is a bird soaring on air currents. This definition is about the selected event and boundary. The same bird can generate muscular power in another behaviour.

Calling a reaction active can tempt a leap to agency: the object “does something of its own”. That is not the paper’s criterion. A charged spring released by a touch can supply the output energy and still have no goal, model or choice. Conversely, a low-energy controller may organise a vast external flow and appear passive under a narrow energetic definition while exercising consequential control.

The energy axis is useful because it stops an input from being mistaken for the whole cause of an amplified response. It is incomplete because behaviour has other dimensions. Purpose, feedback, prediction, continuity and degrees of freedom are assessed separately. The taxonomy is a sequence of questions, not a single ranking from inert to intelligent.

The phrase “may be interpreted” makes the definition observer-relative. The authors admit its vagueness and say it is operationally weak. They retain it because final-condition language distinguishes important performances: reaching a target, bringing a glass to the mouth, maintaining a relation or intercepting a moving object.

They motivate the category through voluntary action. A person decides to bring the glass to the mouth, not to command a complete sequence of individual muscle contractions. Selection appears at the level of an outcome while detailed coordination follows. The essay treats this as a physiological fact and suggests that stimulating motor cortex triggers output pathways without triggering a purpose. That is a programmatic interpretation, not a decisive experiment about awareness.

Purpose in design, purpose in use and purpose in performance must be separated. A roulette wheel is built precisely to generate purposeless results. A clock is orderly and serves a human purpose, yet its motion does not continually strive for a specific final condition. A gun can be used to hit a target, but random firing is possible and target attainment is not intrinsic to the bare mechanism.

A target-seeking torpedo differs. Signals related to the target modify its motion, so attaining a spatial relation is internal to the performance being classified. The authors call such machines servomechanisms. “Intrinsic” still depends on the boundary: fuel, operator-set reference and target signal remain environmental conditions. It means the goal correction occurs within the selected object–environment loop.

The authors first note two engineering senses of feedback. In the broad sense, some output returns as input. When it returns with the same sign, as in a positive-feedback amplifier, it adds to the original signal and does not correct it. Their teleological category uses a narrower sense: the object’s behaviour is controlled by its current margin of error relative to a reasonably specific goal.

Negative feedback sends information from the goal relation back into action. An output that would carry the object beyond the goal is restricted or reversed. The name describes the correcting sign, not moral value or numerical negativity. Whether a loop actually remains corrective also depends on timing and damping, a qualification the essay soon makes vivid.

The authors say all purposeful behaviour may be considered to require negative feedback because signals from the goal are necessary “at some time”. They nevertheless define a non-feedback purposeful class. There is no contradiction if the temporal scopes are distinguished. A movement can be aimed using goal information before it starts, then proceed without new goal signals modifying it. Feedback-purposeful behaviour continues sampling error during the act.

Their examples are a light-insensitive machine pre-set to strike a luminous object, a snake striking a frog and a frog striking a fly. The animal movements are said to be so fast that new retinal information is unlikely to traverse the nervous system and alter the muscles in time. The acts can be directed towards a goal while their execution is open-loop.

Imagine a machine that moves towards a luminous target using the light’s direction and intensity. It passes the target. The sensed error reverses, prompting a stronger movement in the opposite direction. If that correction overshoots too, each reversal can be larger than the last. A loop intended to reduce error has become effectively positive over the oscillation.

This is the bridge from Paper 0061. Maxwell showed that negative feedback as a mechanical direction is insufficient; characteristic roots reveal whether disturbances decay. Rosenblueth, Wiener and Bigelow translate the same structure into a behavioural classification. The machine remains error-controlled, but its performance fails to attain the goal because timing and damping turn correction into growing oscillation.

The authors compare this with a cerebellar patient asked to carry a glass from a table to the mouth. At rest, the patient may show no obvious motor disorder. During the purposeful movement, the hand oscillates with increasing amplitude, spilling the water. The likeness to the undamped machine motivates their suggestion that the cerebellum controls feedback mechanisms involved in purposeful motor activity.

The evidential step must be labelled. Similar traces support a candidate shared control failure; they do not establish identical anatomy or prove the cerebellum’s single “main function”. The paper calls its own proposal a venture. The analogy is most useful as a discriminating hypothesis: manipulate delays or damping and ask whether both machine and motor behaviour change in the predicted direction.

A tropism follows a current source. The authors’ amoeba reacts to where the source is now; they report no evidence that it extrapolates a moving source’s path. Such behaviour can be purposeful and feedback-controlled without prediction. Error is reduced against a present signal.

A cat chasing a running mouse moves towards an extrapolated future position rather than merely aiming at the mouse’s current location. The essay calls this first-order prediction: predict the target’s path. Throwing a stone at a moving target is second order because the target’s path and the projectile’s path must be related. Shooting with a sling or bow is offered as higher order.

The ordering is illustrative rather than a fully specified mathematics. “Foreseeing” need not imply conscious imagery; it can name behaviour consistent with an internal organisation that combines trajectories. A fixed lead angle tuned to one speed may mimic prediction in a narrow condition. Varying speed, direction and projectile dynamics tests whether the behaviour flexibly uses the required variables.

The hierarchy also resists a simple intelligence ladder. Higher-order prediction can be specialised, brittle or externally configured. Non-predictive feedback can be extraordinarily robust. The paper’s claim is classificatory: extrapolation adds a behavioural dimension and can support increasingly complex tests. It does not define intelligence as maximum order.

The published diagram begins with behaviour, separates non-active from active, then divides active behaviour into non-purposeful and purposeful. Purposeful behaviour splits into non-feedback and feedback. Feedback behaviour separates non-predictive from predictive; predictive behaviour branches into first, second and further orders.

The tree is tempting to read as a staircase of superiority. The source does not justify that reading. Its criteria change from branch to branch: energy source, interpreted goal, timing of goal signals and extrapolative structure are heterogeneous. The authors say so. Each dichotomy singles out one chosen feature and consigns everything else to an amorphous “non-class”. “Non-predictive” includes many organisations sharing only the absence of the selected feature.

Other classifications cross-cut the tree. Behaviour might be linear or nonlinear, continuous or discontinuous, low- or high-dimensional. A passive process can be nonlinear. A feedback process can be discrete. A predictive process can have few degrees of freedom. These are not contradictions because the axes answer different questions.

The hierarchy earns its keep by isolating predictive behaviour and rehabilitating purpose and teleology for objective study. It proposes increasingly demanding behavioural tests that can be used on animals and machines. Its limitations suggest how to use it responsibly: report the path through this tree, then add independent axes rather than forcing every relevant fact into a single rank.

The authors argue that prediction requires discrimination of at least two coordinates: time and at least one spatial axis. Flexible prediction improves when a behaving object can respond to changes in more spatial dimensions. A receptor constrains what distinctions are available. A single scalar intensity can be ambiguous between distance, source strength and obstruction.

The paper contrasts a bloodhound following an olfactory trail with auditory and visual localisation. It treats the chemical signal as reporting distance through intensity and therefore as supporting following rather than prediction. This is a simplified 1943 illustration, not a definitive account of canine olfaction. Its durable point is informational: a controller cannot extrapolate variables its inputs fail to distinguish.

Input richness is only the first gate. A photoelectric cell can report a moving light, yet a machine without the organisation to estimate motion will not intercept it. The authors extend the point to nervous systems, speculating that limits of integration constrain prediction orders in other mammals and that humans may support higher orders. Those comparative claims are conjectures, not results from experiments described in the essay.

The two-gate model prevents sensor mysticism. Adding cameras does not grant a predictive world model; improving inference cannot recover information absent from all observations without assumptions. Performance depends on the product of observability and computation. A failure can occur because the variables were not sensed, because they were not integrated or because the action model was wrong.

The authors reject the claim that machine designers merely duplicate living performance. Animals already occupy many behavioural classes, so overlap is unsurprising, but machines also do things humans cannot: emit electrical outputs directly or generate radio waves. They point out that locomotives use wheels rather than imitated legs. Behavioural possibility is larger than biological copying.

They say no qualitative behavioural characteristic unique to only one group had appeared “so far”. That phrase is historically bounded and the claim is scoped to their behavioural lens. It neither proves that machines and organisms are equivalent nor shows that experience is exhausted by observable performance. It invites common tests while leaving the result open.

Open the objects and differences appear. The essay contrasts protein-rich, colloidal organisms with chiefly metallic machines; ionic electrical changes with electronic conduction; and biological spatial multiplication with machine temporal multiplication. A television scans rapidly with one effective sampling point, whereas the eye uses many receptors in parallel. The numerical counts and technological comparisons belong to 1943 and should be read historically.

The robot passage makes the separation memorable. An engineer seeking animal-like behaviour would use metal, dielectrics and vacuum tubes, gaining speed and power but rudimentary learning and memory. The authors imagine future engineers using protein-like materials, then remark that the ultimate model of a cat is another cat. Similar performance can be reached by different construction; exact reproduction collapses modelling into duplication.

Traditional teleology was often associated with final causes, as though a future condition reached backwards to produce the present action. That association made teleology appear opposed to causal determinism. Rosenblueth, Wiener and Bigelow remove the backward arrow. A current goal signal and a current error can causally modify current output. No future event needs to act on the past.

Under their restricted usage, teleological behaviour becomes synonymous with negative-feedback-controlled purposeful behaviour. It is opposed to non-teleological behaviour, not to determinism. Both teleological and non-teleological systems can be deterministic wherever deterministic description applies. The goal functions as a reference in an organised loop, not as a mysterious later force.

The definition gains precision but not completeness. “Final state interpreted as purpose” still requires an analyst to identify the relevant goal relation. Multiple apparent goals may fit one trace. A thermostat’s set point is explicit; an animal’s goal may be inferred from flexible correction, persistence and context. Interventions that move the target or change obstacles make the interpretation more discriminating.

Nor does the category entail mental purpose. A servomechanism can be teleological in this restricted sense without awareness, understanding or self-authored ends. The paper begins from voluntary activity and compares animals and machines, yet it supplies no operational test of subjective experience. Feedback establishes a causal organisation of behaviour. Conscious purpose, moral agency and semantic understanding each need additional evidence.