Reward Generalization

A simple explanation

You trained on one cue. The cue delivered a small reward. Your nervous system, doing what it is built to do, did not file the cue narrowly. It filed the shape of the cue. From then on, any cue that resembles the original — colour, layout, timing, posture — will pull on the same reward circuit, slightly weaker but recognisably the same.

This is reward generalization. It is the reason a conditioned response rarely stays inside the bounds of its original training. The dog trained to salivate at a 1000 Hz tone will also salivate at 900 Hz and 1100 Hz. The user conditioned to the home-screen-pull will feel a faint version of that pull at any rectangular screen with a similar grid of icons. The response was never to this cue. It was to the class of cues that the System decided this one belonged to.

An everyday example

You spent two years building, slowly, a difficult relationship with one social app. You deleted it from the phone. You felt the small relief of the absence. A week later you opened a different app on a tablet — a news reader, completely unrelated content — and noticed your thumb performing the same downward pull-to-refresh gesture, the same brief micro-anticipation, the same tiny deflation when nothing new loaded.

The first app is not on this device. The pull is. The Reward System is not asking for that app. It is asking for the shape it learned. The shape lives in any layout that rhymes with the original. The "one habit" was never one habit; it was a generalized expectation that screens deliver pull.

What is reward generalization?

It is the process by which a conditioned reward response transfers, with diminishing strength, to stimuli that share features with the originally conditioned cue. The technical literature describes a generalization gradient — the further a new cue is from the training cue along the relevant dimension, the weaker the response — but the gradient is not always steep. A response trained on one shape can extend across surprisingly distant cues if the shared features are the ones the nervous system happens to be tracking.

This is the standard behaviour of associative learning systems, not a malfunction. The body cannot afford to re-learn the world cue-by-cue. It generalises because generalising is cheaper than encoding. The cost shows up later, when the generalised pull starts firing in contexts where there is no actual reward to deliver.

Why does my phone habit spread to other screens?

Because the Reward System was never tracking the phone. It was tracking a constellation of features — bright-rectangle, finger-down-pull, brief-anticipation, intermittent-payoff — that the phone happened to package. Anything that re-presents enough of those features will pull on the same circuit. The tablet, the laptop, the smart TV, even the rectangular vending machine display at the train station all rhyme with the original to some degree.

This is also why partial interventions feel like they almost worked. Removing the original cue removes the strongest pull. The generalised pulls remain, faintly, on every similar surface in your life. The System, having generalised, does not need the original anymore.

The behavioral loop

A loop that runs across surfaces rather than within one:

Original conditioning — a specific cue is paired, repeatedly, with a reward (or with intermittent reward, which generalises even faster).
Feature extraction — the nervous system identifies which features of the cue predicted the reward. It does this automatically and not always accurately.
Generalization — those features are now sensitised. Any cue that re-presents them activates the same reward circuit, weaker but real.
Spread — the response begins firing in contexts the original training never anticipated. Each firing reinforces the generalised pattern.
Substitute compounding — when one near-cue delivers even a small reward, it strengthens the generalisation. The class widens. The pull spreads further.
Intervention failure — removing the original cue leaves the generalised expectation untouched. The System routes to the next nearest cue and the loop resumes.

Emotional drivers

Three feelings, often invisible because each individual firing is so small:

A faint, recurring micro-craving that the mind cannot easily attribute to a specific source.
A diffuse restlessness around any context that shares features with the original — screens, scrolling surfaces, layouts with intermittent novelty.
A quiet frustration when an intervention "worked" but the underlying pull did not disappear, only relocated.

What your nervous system does

The Reward System routes through the dopaminergic prediction system the same way whether the cue is original or generalised — the gain is just lower for cues further from the training cue. The system was built to over-generalise on the way up (better to anticipate reward and be wrong than to miss it) and under-extinguish on the way down (better to remain primed than to stop expecting). This asymmetry is why generalised reward patterns are easy to acquire and slow to release. The body holds the shape long after the original cue is gone.

The DojoWell interpretation

Reward generalization is what makes substitute habits contagious. The MDT frame: the Reward System's original ask was stimulation toward something — a state, a closure, a felt sense of having been engaged. The conditioned cue was a substitute for that ask. Generalization is the substitute learning to wear other clothes. The System no longer needs the original costume; any near-rhyme will do.

This explains the common experience of removing a habit and watching the pull migrate. You did not fail at the intervention. The intervention was aimed at the wrong level. You removed an instance; the pattern lives one level up, in the class of cues the System sensitised to. The substitute has generalised beyond the original case, and the original system — stimulation toward something that actually deposits — is still unmet.

Helpful generalisation does exist. We call it transfer learning when a skill trained in one context appropriately extends to a related context: a violinist who picks up the viola, a runner whose pacing carries to swimming. The difference is whether the generalised response delivers the deposit the System was originally tracking. Transfer learning generalises a capacity; problematic reward generalization generalises a pull. One widens the closure pattern. The other widens the substitute.

How do I stop a habit from spreading to other contexts?

You intervene at the level of the generalised pattern, not the original cue. Removing one app, one site, one device addresses the instance. The generalisation lives in the shape — the constellation of features the System sensitised to — and it will route to the nearest available match unless that shape itself is allowed to extinguish.

Three moves, in order of difficulty:

Name the shape, not the instance. I am pulled by the bright-rectangle-pull-to-refresh shape, not I am pulled by this app. Naming the shape is what makes it possible to recognise the pull when it migrates.
Identify the near-cues in your environment. Every device, every layout, every context that rhymes with the original cue. The full set is usually larger than expected. This is the field across which extinction needs to happen.
Allow non-reward across the whole class, not just the original. Extinction is the reverse of generalisation. The pull weakens only when the generalised prediction repeatedly fails to be confirmed. A single removed app does not extinguish; a sustained non-engagement across the whole near-cue class does.

Practical steps

Map your own generalisation gradient. List the cue you originally trained on, then the nearest five cues that now pull on you. The map itself is half the work — most generalisation operates below conscious naming.
Choose one near-cue, not the original, for the intervention. Working on a generalised cue teaches the System that the shape is no longer reliable, which extinguishes faster than working on the original alone.
Install one structural change at the shape level. A device-wide setting (greyscale, notification class disabled, full-screen layout removed) is more useful than an app-level change, because it operates across the whole class.
Track migration explicitly. When you remove one cue, watch for which near-cue gets stronger over the following week. The migration is the visible form of the generalisation.
Distinguish helpful transfer from problematic spread. If a generalised response is delivering the original deposit elsewhere — a capacity widening — leave it alone. If it is only delivering the pull — a substitute widening — that is the level to intervene at.

Reflection questions

Which of your habits have spread to surfaces you did not originally train on?
When you remove an instance of a habit, where does the pull migrate to?
Is there a generalised pull in your life that is actually a transferred capacity in disguise?
What is the shape — not the instance — that your Reward System is tracking?

Frequently Asked Questions

Is reward generalization the same as cross-addiction?

Cross-addiction is a specific clinical pattern — quitting one substance and acquiring another — and it is one expression of reward generalization, usually at the highest end of the gradient. Reward generalization is the broader mechanism that produces it, and it also produces much smaller, everyday spreads that never reach clinical threshold but compound across the attention economy.

How is reward generalization different from transfer learning?

Transfer learning generalises a capacity — a skill or competence that delivers real deposit in a new context. Reward generalization generalises a pull — a conditioned response that fires across similar cues whether or not the new context delivers anything. Same underlying associative machinery, opposite consequence for density.

Why can't I just quit one app and be done with it?

Because the Reward System was not tracking the app. It was tracking the shape — the constellation of features the app packaged. Removing the app removes the strongest instance of the shape; the generalised expectation remains and routes to whatever near-cue is closest. The intervention has to operate on the shape, not the instance.

Why do similar-looking things trigger the same craving?

Because the nervous system files cues by feature, not by identity. The features the System sensitised to during the original conditioning will activate the same circuit anywhere they re-appear. The craving is not about the new object; it is about the shape the new object happens to wear.

How does this connect to Meaning Density?

Reward generalization is a clean shallow_stimulation signature. Each near-cue delivers a small pull but rarely the closure the System was originally tracking. The deposit per firing is low; the residue accumulates as a generalised expectation of being under-stimulated; the effort is paid across a widening field of cues. The equation runs low and stays low until the whole class is allowed to extinguish.