You cannot stop scrolling because the feed is running a reinforcement schedule the reward system is built to chase: a tight, repeating loop in which the feed delivers unpredictable rewards, the brain learns to chase them, and pulling away feels disproportionately effortful. Social media anxiety is the cluster of worry, comparison and physiological arousal that builds in people who use platforms in ways that begin to feel out of their control, and the dopamine doomscrolling loop is one of the cleanest mechanisms behind why “in ways that begin to feel out of their control” is the right phrase. This piece walks through what the 2025 research actually says about that loop, what it does not, and what to do about it.
The mechanism matters because it changes the kind of solution that can work. Telling someone caught in a variable-ratio loop to “just put the phone down” is roughly as useful as telling someone at a slot machine to “just walk away”. The contingencies are doing the work, and the practical implications follow from understanding them. For the wider clinical landscape, our Anxiety topic hub sets the context. For the broader brain story this piece narrows out of, see our companion mechanism deep-dive on how Instagram hijacks the anxious brain.
What Sharpe and Spooner named in 2025
In a 2025 commentary in Perspectives in Public Health, Sharpe and Spooner argued that the term “dopamine-scrolling” deserves to be treated as a precise behavioural construct, not a loose metaphor. [sharpe-spooner-2025-dopamine] Their case was that the field already had the ingredients, variable-ratio reinforcement from behavioural learning theory, mesolimbic activation from neuroimaging, the addiction-medicine framework of cue-driven craving, but had been using “phones release dopamine” as if that single phrase explained anything. It does not. Many enjoyable activities involve dopamine release; what makes a loop a loop is the specific combination of an unpredictable reward schedule, fast repetition, a portable cue, and an attentional content style that the user does not fully choose.
Sharpe and Spooner’s framing is also notable for what it refuses to claim. They do not argue that scrolling causes structural brain damage in the way some popular accounts have. They do not claim equivalence with stimulant addiction. They do argue that the public health risk is real, distributed unevenly, and easier to address structurally than individually. The 2025 Cureus neurocognitive review and the 2025 algorithm-and-teen-addiction review fit alongside that framing rather than competing with it. Read together, they give a more careful three-paper picture than any one of them alone.
The reason the 2025 literature is so cautious about strong addiction language is that the diagnostic landscape is still genuinely unsettled. The DSM-5 does not list “social media addiction” as a discrete disorder, and the ICD-11’s “gaming disorder” inclusion remains the only behavioural-internet diagnosis with formal sign-off. Sharpe and Spooner argue that operating without a precise mechanism term has cost the field, clinicians and policymakers have been forced to talk in metaphor for years. Naming the dopamine-scrolling loop is not the same as diagnosing it, but it does give researchers a shared construct to measure, model, and design interventions around. That is a meaningful step forward, even if the resulting body of evidence still needs another decade to mature.
How the mesolimbic system actually behaves
The mesolimbic dopamine pathway runs from the ventral tegmental area in the midbrain to the nucleus accumbens in the ventral striatum and onward to parts of the prefrontal cortex. Its job, as fifty years of animal and human neuroscience has clarified, is not to produce pleasure. Its job is to signal reward prediction, to track the difference between the reward you expected and the reward you got, and to update the motivational pull of cues and actions accordingly. The 2025 Cureus review collects the fMRI evidence that short-form personalised video activates exactly this circuit, in patterns reminiscent of, though not identical to, other rewarding behaviours. [cureus-modern-day-high-2025]
This distinction, prediction versus pleasure, explains something important about doomscrolling. People often report that an hour on the feed did not actually feel good. They feel restless, mildly worse, sometimes ashamed. That experience is consistent with mesolimbic activation: the system that pulls you toward an action is not the same system that registers whether you enjoyed it. A slot-machine player is not always having a great time, but the pull is still there because the reward-prediction signal keeps firing. A doomscroller’s experience often maps onto the same pattern. The loop runs on anticipation, not satisfaction.
The “wanting versus liking” distinction in reward neuroscience, most famously developed by Kent Berridge’s lab through two decades of animal and human work, is what allows the Cureus review to talk seriously about scrolling as reward-circuit engagement without collapsing into the lazy claim that scrolling “feels great”. Wanting and liking dissociate at the neural level: dopamine drives wanting, opioid and endocannabinoid systems drive liking. Behavioural-addiction patterns tend to involve elevated wanting against flat or even reduced liking, which is the lived texture of compulsive use. Many users describe scrolling exactly this way, pulled toward it without enjoyment, finishing a session feeling worse, then reaching for the phone again twenty minutes later. The neuroscience predicts that experience rather than contradicting it.
Variable-ratio reinforcement, in plain English
Behavioural learning theory describes several schedules by which rewards can follow actions. The one that produces the most stubborn, hardest-to-extinguish behaviour is variable-ratio reinforcement: reward arrives after an unpredictable number of actions, with the average being some specific number but no individual instance being predictable. This is the schedule slot machines use. It is also, structurally, the schedule a short-form video feed uses. The next swipe might bring something genuinely funny, a stranger’s beautiful three seconds, a politically infuriating clip, a stale meme, or absolutely nothing of interest, and the user cannot tell in advance which.
That uncertainty is the engine of the loop. Behavioural learning research, which the 2025 algorithm-and-teen-addiction review revisits, has shown for nearly a century that variable-ratio schedules produce more persistent behaviour than continuous reward. [algorithm-teen-addiction-2025] If every swipe delivered the same content, users would adapt and disengage quickly. The unpredictability is what trains the brain to keep pulling. Importantly, this is true regardless of whether the user “likes” the content on average. The schedule, not the content quality, is the key variable. Designers of engagement-optimised feeds did not need to read the behavioural science textbook to discover this; A/B testing converged on a schedule the textbook would have predicted.
The link between variable-ratio reinforcement and anxiety is more subtle than the link between the schedule and persistence. Persistence is what makes the user keep scrolling. Anxiety, in this picture, arises from two adjacent dynamics. The first is that the unpredictability itself raises a low-grade autonomic baseline, the body sits in mild “next reward incoming?” alertness throughout a session and often for some time after. The second is the comparison content the feed serves alongside the rewarding clips: appearance, status, achievement, social-popularity signals, news of distant threats. Variable-ratio reinforcement keeps the user inside an information stream that, on average, contains plenty of triggers for worry, even if any individual clip is harmless. The schedule binds the user to the stream; the stream supplies the anxiogenic material. The 2025 Cureus review treats these as separate but reinforcing mechanisms, and it is worth holding them apart when thinking about interventions.
What brain imaging actually shows
The 2025 imaging literature is, in absolute terms, still small. fMRI studies on social media use number in the dozens, not the thousands, and many use simulated rather than live platforms in the scanner. With that caveat in front, the Cureus synthesis maps a converging pattern.
When study participants view personalised short-form content, TikTok-style clips selected by algorithm, activation increases in the ventral striatum, including the nucleus accumbens, and in cortical regions involved in salience attribution. When the same participants view non-personalised content matched for visual properties, activation is weaker and more variable. When users with self-reported problematic patterns of use are compared to lighter users, the difference in striatal reactivity grows.
None of those findings alone constitutes proof of an addictive process. fMRI activation is not specific enough to nail down “addiction” as a label; many enjoyable, non-pathological activities also light up reward circuitry. What the 2025 evidence does support, more modestly, is that the brain treats algorithmically tuned feeds as biologically meaningful rewards, and that problematic users show the kind of heightened cue reactivity that, in other behavioural addiction work, tracks with poorer self-regulation. The next step the field needs, and is starting to take, is longitudinal imaging that follows the same users over months as their use patterns change.
A second methodological wrinkle worth knowing: most imaging paradigms ask participants to watch content, not to actively engage with it the way they would on their own device. Engagement on a personal feed includes prediction (will the next post be funny?), social reasoning (what does this comment mean about me?), and self-relevant evaluation (do I post something in response?). Each of those processes recruits additional networks beyond the basic reward circuit, default-mode regions for self-referential thinking, social-cognition areas for mentalising about others, and lateral prefrontal regions for response selection. The 2025 Cureus synthesis is candid that scanner-friendly paradigms underestimate this real-world complexity, and that the activation maps published so far are best read as a lower bound on what is happening in the brain during ordinary feed use rather than the full picture. [cureus-modern-day-high-2025]
The adolescent brain is a particularly bad fit for the algorithm
The third mechanism is developmental and, in some ways, the most consequential for policy. The 2025 algorithm-and-teen-addiction review on PubMed Central argues that adolescent brains are structurally primed to be more reactive to variable-reward algorithmic feeds than adult brains. [algorithm-teen-addiction-2025] Two facts drive that argument.
First, the limbic system, including the nucleus accumbens, matures earlier than the prefrontal cortex. By mid-adolescence, the reward system is essentially online, while the regulatory machinery that would let a user notice “I have been scrolling for ninety minutes and feel worse” is still being built. That mismatch is well-documented across decades of developmental neuroscience and is not specific to social media; it shows up in adolescent risk-taking more generally.
Second, algorithmic feeds are particularly good at finding the content most likely to engage a given user. For an adolescent already at an emotionally activated baseline, which most adolescents are, that targeting compounds the developmental mismatch. The system designed to deliver reward unpredictably is also calibrated, in real time, to deliver the most personally rewarding kind of unpredictability. The result, the 2025 review argues, is a particularly steep loop in which young users find disengaging especially hard, not because they lack willpower but because the architecture and the development stage interact badly.
The 2023 US Surgeon General’s advisory had already gestured at this. The 2025 neuroscientific work makes the developmental argument more specific and harder to wave away. It does not justify deterministic claims, most teenagers do not develop a behavioural addiction, but it does support more conservative policy around younger users than around adults. The signs and symptoms guide walks through the warning indicators most relevant to that age band.
There is also a feedback dimension worth flagging. Adolescents are more sensitive to peer feedback than adults, that, too, is decades-old developmental neuroscience, and feeds happen to be machines for delivering peer feedback at high frequency. The 2025 algorithmic review argues that the convergence of three factors (variable reward, an immature regulatory cortex and peer-feedback hypersensitivity) is what makes the teen-feed pairing particularly potent, not any single ingredient on its own.
Where the mechanism story still oversells itself
Honesty about limits matters. Three things the 2025 mechanism literature does not, on its current evidence, fully establish.
First, “rewires the brain” claims outrun the data. fMRI shows acute activation; it does not, on its own, prove durable structural change. A few small studies have reported grey-matter differences associated with heavy use, but they are cross-sectional, with all the directional ambiguity that implies. Causal structural claims require longitudinal imaging that the field has only started to produce.
Second, the addictive-disorder label is contested for good reasons. Some researchers argue that pathologising what is, for most users, normal variable engagement risks medicalising ordinary digital life. The 2025 mechanism evidence is strong enough to support clinical concern at the problematic end of the spectrum, but not strong enough to label every heavy user with a disorder. The field’s contrarian voices, covered in our read of the 2025 numbers, have a real point that the mainstream needs to engage rather than dismiss.
Third, individual differences are large. Some users show striking striatal reactivity to feeds; others show very little. Anxiety vulnerability, baseline mood, sleep, family environment, content type and platform design all modulate the response. A circuit-level account is useful as a population-scale explanation, but it cannot tell you what scrolling does to a specific person; for that you have to ask them.
Why willpower fails inside the loop
If the engine is the schedule, the consequences for willpower are not flattering. Self-control models in psychology, even the more modern, less depletion-flavoured ones, assume that the action and the reward are roughly visible to the user, so the person can deliberate and choose. The dopamine loop short-circuits that assumption in two ways. First, the reward is fast and small relative to the cost of resisting; second, the action itself (a swipe) is so low-effort that there is barely a decision point to inject reflection into. Sharpe and Spooner make this point bluntly: telling someone caught in a variable-ratio loop to “just have more discipline” is a category error.
The 2025 Cureus review adds a second mechanism. Sustained exposure to short-cut, high-stimulation content appears to entrain cortical attentional networks toward those characteristics, which makes longer, slower activities feel relatively dull. That is not “brain damage”, the effects in the literature are reversible with structured breaks, but it is a real shift in the comparative experience of attention. Together, the schedule and the attentional drift mean that the user who relies on motivation alone is fighting two contingencies at once. The 2025 research strongly suggests that structural interventions, changing what the phone affords, where it sits, what is on the home screen, outperform motivational ones for exactly this reason. Our piece on therapist-designed five-rule protocols covers what those structural interventions look like in clinical practice.
There is also a third factor that the 2025 work surfaces, less often mentioned in popular framings: the loop interacts with sleep. A short late-evening session on an algorithmic feed not only delays sleep onset directly through light and arousal, it also runs the reward-prediction system during the window in which the brain should be winding down. Poor sleep then lowers next-day prefrontal control, which is precisely the resource motivational strategies depend on. Anyone who has tried “I just need to be more disciplined about my phone” while sleeping six and a half hours can attest to the asymmetry. The same structural moves that reduce loop exposure, phone out of the bedroom, no scrolling after a fixed evening hour, protect sleep at the same time, which is part of why they outperform pure willpower. The two mechanisms reinforce each other rather than competing.
Why some apps loop tighter than others
The dopamine loop is not a property of “social media” as a single category. It is a property of specific design choices that map onto variable-ratio reinforcement more or less cleanly. Three features make a feed loop tighter. The first is content length: shorter clips compress more reward-prediction signals into the same minute. The second is recommendation opacity: an algorithmic feed selected for engagement is, by design, less predictable than a chronological feed of accounts you follow. The third is action friction: the lower the cost of triggering the next reward (a swipe versus a click versus a search), the more iterations the loop runs per session.
The 2025 algorithm-and-teen-addiction review draws these distinctions carefully. Apps built around short-form algorithmic video, TikTok, Reels, Shorts, score high on all three features. Direct-message-heavy use of the same platforms tends to score lower; people-you-follow timelines lower still. This is one of the reasons single-platform bans tend to underperform: the loop is portable across surfaces. A user who deletes one short-form app and replaces the time with another short-form app has not changed the contingency. A user who replaces it with reading, conversation, or longer-form video has. Our platform comparison piece goes deeper into how individual apps stack up on these specific features.
Breaking the loop in practice
If the loop is held together by an unpredictable reward schedule, a portable cue, and a low-friction action, then breaking it means weakening at least one of those three. The 2025 literature is reasonably clear about which interventions move the dial. Reducing the cue, phone face-down, in another room, out of the bedroom overnight, is one of the highest-leverage moves because it interrupts the loop before it starts. Increasing the friction, greying out the screen, removing apps from the home screen, logging out so re-entry requires a password, adds enough cost per iteration to surface the user’s actual preferences. Changing the schedule, switching from infinite-scroll algorithmic feeds to longer-form content that cannot deliver a reward every few seconds, directly attacks the reinforcement engine.
For users who already feel caught, structured time off, the kind tested in the 2022 Lambert RCT and the 2025 detox meta-analyses, sits at a different level of intervention. It does not change the loop; it removes the user from it for long enough to recalibrate. Recalibration matters because comparative attention shifts back: longer activities stop feeling dull. Our account of the 30-day detox question covers what the experimental literature actually shows about that period, including the recovery question (do gains hold after returning?). For people whose anxiety crosses into clinical territory, none of this replaces evidence-based treatment; our overview of treatment protocols covers what works.
It is worth being explicit about what does not, on the 2025 evidence, reliably break the loop. Self-set screen-time limits are easily overridden in the moment and tend to drift. Daily-minute caps that count time but not pattern miss the schedule mechanism; ten short sessions can run the loop harder than one long one. Notification batching helps a little but leaves the variable-ratio schedule intact once the app is open. None of these is useless; they are simply weaker than structural moves that change the cue, the friction, or the schedule.
A second pragmatic point: the loop is most breakable in the first few seconds of an unconscious phone-pickup. Once the app is open and the first piece of content has loaded, the contingency is already running and the user is into the schedule. Interventions that catch the action before the app opens, a phone left in another room, a greyscale screen that strips the visual reward, a home screen with no app icons present, outperform interventions that try to interrupt after the first swipe. This is one of the clearer practical consequences of taking the mechanism seriously, and it is the reason therapists working in this area increasingly recommend changes to the environment over changes to self-talk.
A schematic of the variable-ratio reinforcement described in this section. Not measured data.
What the loop framing changes about responsibility
One of the underappreciated implications of the 2025 dopamine-loop work is what it does to the conversation about blame. If the loop is engineered into the product through engagement-optimised reinforcement schedules, then the heaviest users are not the most morally weak users, they are the users whose contingencies happen to match the design most cleanly. That framing does not absolve users of any agency; it does shift where the most effective interventions sit. Individual habit change is real and useful. Product-level design change is much larger in expected impact, which is partly why the Sharpe and Spooner paper treats dopamine-scrolling as a public health issue rather than a personal one.
For an individual reader, the takeaway is to stop running the experiment that asks “can I just want it less”. The loop is built to resist that experiment. Run the structural experiments instead: change the cue, change the friction, change the schedule. Track what changes in your anxiety symptoms over a fortnight. If the numbers move, you have something to keep. If they do not, the contingency is probably tighter than you assumed and a structured break, or professional support, is the more honest next step. The FAQ below collects the five questions readers ask most often after working through this mechanism, the answers there add detail without rerunning the article.
When to seek help
Speak to a doctor if the pull has stopped being about the phone. The loop described here explains why a feed is hard to put down; it does not explain restlessness that persists once it is down, worry that finds a new subject as soon as you close the app, or dread that is there when you wake up. Those point at anxiety in its own right rather than at an algorithm.
Say what you have already tried, because “I removed the apps for a fortnight and the feeling stayed” is genuinely useful information about which problem this is. Our free anxiety screener (GAD-7) measures the feeling rather than the habit, which is the distinction this section turns on. Our guide to doomscrolling covers the behaviour and the anxiety hub covers the condition.
Go urgently if you have thoughts of harming yourself. Contact your local emergency services or a crisis helpline.