Algorithms Are Trapping a Group of People

Image source | Network (please contact for deletion if infringing) Partially generated by AI

In 2026, your phone is learning to do something far more thorough than "guessing what you like": making decisions for you and making you feel like they were your own.

The Mobile World Model developed by Panasonic in collaboration with UCLA enables AI agents on Android devices to predict state changes in natural language after operations, with task success rates improving from 46.9% to 54.3%.

HONOR's YOYO intelligent assistant achieves a comprehensive task accuracy rate of 91.8%, an end-to-end closure rate of 90%, and supports over 100 consecutive operations at maximum.

Google is developing an intent prediction system on the device side, analyzing how you interact with your device to understand what you want to do before you even speak.

Apple has outsourced Siri's "brain" to Google's custom trillion-parameter Gemini with one goal: to turn iOS into the ultimate gateway for understanding user intent and "achieve a dimensionality reduction strike against all apps."

Most of these news stories are framed as "more convenient." But the flip side of convenience is a behavior prediction system that knows you better than yourself, upgrading from "recommending content" to "acting on your behalf."

You no longer need to make choices because the algorithm has already made them for you. And you may not even realize that choices ever existed.

With the development of AI, algorithms have upgraded, forming an entire infrastructure of algorithms that grows from the operating system's core, with you as the prediction target.

From "Guessing What You Like" to "Predicting What You'll Like"

To understand the algorithmic landscape in 2026, we must first recognize a turning point overlooked by most.

Over the past decade, algorithms' control over mobile users has primarily occurred at the content layer.

Short video recommendations, information feed sorting, and e-commerce "guessing what you like" essentially match within your existing preferences: if you watch one more cat video, the system pushes a hundred more cats to you.

While powerful, this logic has a premise: you at least know "what you're watching." If you don't like it, you can swipe away; if you're tired of it, you can turn it off.

But algorithms in 2026 are no longer satisfied with matching preferences; they have begun predicting intentions and executing directly.

On-device large models enable phones to complete a "perception-reasoning-execution" closed loop (closed loop) locally. Panasonic's World Model no longer predicts the screen pixel by pixel but uses natural language to predict state changes like "pressing the button will open the settings interface," thus selecting the operation most likely to achieve the goal from multiple candidates.

This means AI agents no longer mechanically execute "clicking a coordinate" but, based on understanding "what you want," take you the rest of the way.

More critically is the positioning at the operating system level. Apple has transformed Siri into a system-level agent in iOS 27, capable of reading screen content, calling tools across apps, and connecting the browser, passwords, messages, calendar, and photos.

Under the A2A (Agent-to-App) route, users authorize "intentions" and "instructions" (e.g., "buy a ticket") rather than giving AI a pair of eyes that constantly "peek at the screen."

This sounds "cleaner," but its consequences are equally profound: when the operating system becomes the gateway and scheduler of intentions, all apps are reduced to executable units called upon, and users shift from "operators" to "authorizers."

The more you authorize, the more the algorithm understands your behavior patterns; the more it understands your patterns, the better it can complete choices for you before you form a clear consciousness.

Google's intent prediction research doesn't even need to leave your phone; it can infer your intentions by analyzing how you interact with your device.

Every hesitant pause, every repeated swipe, trains a system that knows what you "want" before you do.

This far exceeds "recommendation" and resembles prediction and agency.

And the precision of prediction precisely forms the first layer of structure that traps you: when the algorithm is always one step ahead, you gradually lose the muscle memory of "deciding for yourself."

The Cost of Precision

Algorithmic precision itself is not the issue. The problem lies in whose goals the precision serves.

In March 2026, a California jury made a landmark ruling: Meta and Google were held responsible for a young woman's social media addiction formed through algorithmic recommendations since childhood, awarding $6 million in damages.

The plaintiff described herself as "eroded bit by bit by endless algorithmic recommendations," experiencing depression, anxiety, and suicidal thoughts.

The plaintiff's legal team bypassed debates over "content responsibility" and targeted the product design itself: infinite scrolling, autoplay, and like feedback formed an addictive structure fueled by attention and aimed at dependency.

The court's message to the tech industry was clear and sharp: users' mental and physical health are not production materials that platforms can endlessly exploit.

The significance of this case lies not in the compensation amount but in revealing a fact hidden by the word "precision": algorithmic precision is never neutral. It precisely serves retention time, engagement rates, and commercial conversion.

Infinite scrolling removes all natural stopping points, variable reward mechanisms turn "what will appear next" into a portable slot machine, and algorithmic personalized feeds learn "what keeps you emotionally activated" and give you more.

These designs are no longer secrets; they are repeatedly verified, systematically deployed engineering choices.

In China, the same logic operates under a different guise. In October 2026, media reported that elderly people living alone in rural Henan were addicted to AI-generated micro-short dramas, with some watching over ten hours daily.

A seventy-year-old in Xinjiang spent a total of 225,800 yuan over two years across hundreds of short drama mini-programs, with individual payments small (a few yuan or even cents) but extremely frequent, dispersed across multiple entities, and with dramas frequently changing "aliases."

The report's follow-up question (inquiry) was pointed: recommendation algorithms treat elderly emotional needs as a traffic entry point, constantly amplifying conflict and suspense to trap them in a spiral of induce (inducement); payment rules exploit the elderly's unfamiliarity with online environments.

While the "youth mode" for minors has matured, with duration reminders, payment limits, and one-click disabling, a "brake" for the elderly remains largely absent.

The same algorithmic capabilities are "protective" for youth but "exploitative" for the elderly—the difference lies only in the direction of commercial value.

Precise algorithms have no will of their own, but the organizations deploying them do.

When "precision" is used to reduce your likelihood of stopping, increase your willingness to pay, and extend your stay, it ceases to be a technical metric and becomes a behavioral design targeting you.

You think you're "using" your phone, but from a behavioral design perspective, you are the object being designed.

Price, Labor, and Cognition

Within the 2026 algorithmic system, algorithms are precisely and systematically fixed on smartphones, like an invisible track. I broadly categorize this phenomenon into four types.

Attention capture. This is the most visible and frequently discussed layer. Endless scrolling in short videos, seamless loading in information feeds, and random timing of push notifications create an "interface with no exit."

But its deeper consequence lies not in how much time is wasted but in how it retrains your tolerance for "stopping."

When every blank interval is filled with content, boredom, waiting, and daydreaming—gaps originally belonging to autonomous consciousness—are systematically eliminated. A person who cannot tolerate three seconds of blankness can hardly be said to own their attention.

Economic price exploitation. Another expression of algorithmic precision is "seeing how much you're willing to pay." An investigation found that Instacart in the U.S. used AI pricing to charge different customers different prices for identical goods, with price differences reaching up to 23%—a typical household might pay about $1,200 more annually.

The investigation also revealed that the algorithm charged higher prices to users marked as "more likely to pay."

In China, research on big data price discrimination on Douyin identified four types of algorithmic bias most strongly perceived by the public: price hikes after searching, different prices for different devices, discriminatory pricing based on consumption level, and differentiated information delivery after user tagging.

These practices anger people not just because they pay more but because they expose a deeper truth: in the algorithm's eyes, you are not a person but an estimable range of willingness to pay.

Every click, hesitation, and price comparison you make provides training data for this range.

Labor discipline. The phrase "trapped in algorithms" first entered public consciousness on a large scale through food delivery riders. Follow-up reports in 2026 show that delivery times are precisely calculated by algorithms, forcing riders to race against time under the pressure of multiple orders, with late penalties becoming a major source of anxiety.

But the uniqueness of algorithmic labor lies not in "exhaustion" but in how it disguises control as freedom. The story of Rider A Biao in Fuzhou provides a poignant example: he chose food delivery precisely because it "doesn't require talking to anyone."

The algorithm locks him into a closed loop (closed loop) of bike rental fees, battery fees, and order dispatch ratings through a seemingly free order-taking model. He knows he's being calculated by the system, but "not running means I can't even pay rent."

A deeper entrapment occurs psychologically: the algorithm arranges his actions day after day, gradually convincing him that this is all he deserves in life. "The future? I haven't thought about it."

When a system not only dictates what you do but also shapes the boundaries of what you believe you "can do," it has moved beyond economic exploitation into subjectivity production.

Cognitive narrowing. Much has been said about "information cocoons," but 2026 research offers a more precise picture.

An experiment conducted in Germany and the U.S. found that news recommendation systems based on users' political preferences indeed exacerbated ideological polarization among politically moderate groups; in the German data, politically balanced recommendations reduced emotional polarization among moderate users while intensifying it among extreme users.

This means algorithms don't simply "make people see only one side" but systematically shut down the motivation to cross ideological lines in those who could remain open.

More alarming is the "preference closure" mechanism: AI personalization reinforces existing behaviors while limiting exploratory consumption and heterogeneous exposure. The "free choice" users experience is actually selection within an algorithmically defined set of options.

You think you're making decisions, but you're just ordering from an algorithm-approved menu.

Why "Resistance" Often Fails

Against this structure, individuals are not passive. From 2025–2026, digital minimalism has formed a recognizable practice spectrum among Chinese youth: uninstalling social apps, practicing stealth socializing, setting usage boundaries, and consciously "resisting" algorithms.

The "Digital Minimalists" group on Douban has nearly 10,000 members sharing "30-day digital detox experiments," with some deleting 90% of their phone apps and others voluntarily shutting off their phones for entire weekends.

A doctoral student shut off his phone for four months, attempting to prove that "the idea that many people cannot tolerate being without their phone and social media for even one hour" can be broken.

The real value of these practices cannot be denied. At minimum, they prove one thing: algorithmic control is not physical but depends on users' continuous participation—and thus can be interrupted.

Even if the interruption is temporary or partial, it reopens the space for "choosing not to participate" at the individual level.

But we must also honestly confront their limitations. Research indicates that extreme disconnection often fails due to real-world dependencies—work communication happens on WeChat, payments on Alipay, and navigation in system apps.

More subtly, digital minimalism itself is being co-opted by commercial logic, becoming a new consumer identity and lifestyle label.

You uninstall short videos but might install meditation apps, screen time management tools, and paid "digital detox" courses. The tools of resistance can become commodities once again.

A deeper difficulty lies in how, by 2026, the focus of algorithmic control has shifted from "what you choose to watch" to "what the system executes for you."

When operating system-level agents begin predicting intentions and coordinating across apps, the target of individual "resistance" is no longer a single uninstallable app but an infrastructure diffused through the device's core, fueled by your behavioral data.

You cannot uninstall the operating system, and you can hardly disable intent prediction because "disabling" itself may be designed as a multi-step option or discouraged as "degrading the experience."

Yet precisely because of the difficulty, certain things are worth persisting in.

China's regulatory framework already contains an easily overlooked opening: Article 17 of the "Provisions on the Administration of Algorithmic Recommendation in Internet Information Services" explicitly requires platforms to provide users with the ability to choose or delete tags based on their personal characteristics and offer a convenient option to turn off algorithmic recommendation services; when users choose to turn it off, the platform must immediately stop providing related services.

Since the beginning of this year, systems such as algorithm filing, security assessment, and transparency management have been continuously deepened, and AI governance is moving from sector-specific regulations to comprehensive legislation.

These systems are far from perfect, and their implementation is uneven. However, they have established a principle: algorithm recommendations are not a 'natural fact' that users must accept, but rather a service that can be turned off, explained, and held accountable for consequences upon request.

Truly meaningful 'resistance' may not involve a romanticized complete disconnection; rather, it is a calm and enduring strategic practice: knowing where the system is, understanding what it wants you to do, and then, within the gaps it allows, preserving even a sliver of the ability to 'act against its expectations.'

Finish reading an article before swiping away, instead of letting the next one autoplay; actively search for a field you've never encountered before, beyond algorithm recommendations; occasionally complete those steps manually before an Agent does it for you.

These actions are small, almost symbolic. But what they uphold is the psychological authenticity of 'I can choose not to.'