Have you ever found yourself thinking about something, only to see it pop up on your social media feed shortly after? You didn't search for it online, nor did you mention it to anyone, yet somehow the platforms seem to know what's on your mind.
This phenomenon has been a topic of conversation among friends, leading to speculation about whether our devices are eavesdropping on us. Recently, such concerns have escalated to wondering if algorithms can actually read our thoughts.
ForkLog delves into the mechanics behind this perceived "social media telepathy" and uncovers that the reality is both simpler and more unsettling than one might think.
Surveillance: Myths and Reality
We have become accustomed to scenarios where a casual conversation about a dental issue results in a call from a robotic voice offering a free dental exam, or browsing a marketplace for bicycles leads to an onslaught of bike-related posts across our social feeds. Many have resigned themselves to the belief that "smartphones are listening in".
If you were to discuss this with an AI model, you might be told that such concerns are rooted in conspiracy theory. However, a true skeptic may only become more convinced that technology is conspiring against them, hiding the truth and plotting something sinister. The logic of surveillance seems plausible: smartphones have microphones, apps can access them, and advertisers need data.
In a study conducted in 2018 by researchers at Northeastern University, 17,260 popular Android apps were examined to see if they secretly recorded user conversations. The researchers found no evidence of apps activating microphones to transmit audio surreptitiously. However, they did discover that around 9,000 of these apps had the capability to take screenshots and send them to third parties.
In one case, the food delivery service GoPuff shared screen recordings with the analytics company Appsee. The researchers noted that such recordings could potentially include search queries, cart data, addresses, and payment information. While this is concerning, it does not confirm the hypothesis of eavesdropping.
Nonetheless, the potential for such activities exists, which raises the possibility that someone might exploit them. Microphones can record conversations near the device, and voice assistants "listen" in the background to recognize wake words. False triggers can lead to snippets of conversation being recorded and sent to the cloud without the user's knowledge. However, on modern iOS and Android devices, access to the microphone is regulated by permissions, making unauthorized background recording and audio transmission difficult and largely unreported in mainstream apps.
The notion of spying apps has gained traction to the point where it has been attempted for commercial gain. In August 2026, the Federal Trade Commission finalized a case against Cox Media Group and two marketing companies—MindSift LLC and 1010 Digital Works LLC—accusing them of deceiving customers by selling a product called Active Listening to advertisers. The companies claimed their system could analyze conversations near smart devices for targeted advertising and asserted that users had consented to be monitored.
However, the investigation revealed that the service did not actually utilize voice data or provide the promised targeting. Instead, it merely resold brokered data (such as email lists) under the guise of a new AI product. Collectively, the companies agreed to pay $930,000 in fines: Cox Media Group—$880,000, and MindSift and 1010 Digital Works—$25,000 each.
So, why is it that, despite the potential for surveillance, it is highly unlikely that anyone will actually do it? Primarily due to the prohibitive costs involved. Continuous monitoring would require storing vast amounts of data and incur significant computational expenses, not to mention the serious legal and reputational risks.
Moreover, in most cases, such surveillance is unnecessary. Systems already possess ample information about users.
I Just Thought About That!
Imagine walking down a rainy autumn street when a familiar tune pops into your head. You think it would be nice to go home and rewatch "The Umbrellas of Cherbourg" with a glass of wine. You don't search for the film, message friends, or open a browser. Then, later, you log into your favorite social media platform and see a post about that very movie. At this point, the legends of surveillance seem outdated. How did the algorithms read your mind?
The algorithms didn't read your thoughts at all! Instead, they operate on a different level. They present users with content that they are likely to be interested in, rather than what they are specifically thinking. This distinction is critical. To predict user desires, the system does not need to read minds; it simply needs to understand the user well.
In 2013, researchers from Cambridge University and Microsoft demonstrated how much information can be gleaned from ordinary digital footprints. In a study involving over 58,000 Facebook users, the algorithm could accurately predict not only obvious traits like gender and age based on likes but also personality traits, political views, levels of openness to experience, and other characteristics. The researchers noted that similar insights could be derived from search histories, viewing patterns, and purchases.
Since then, the scale and complexity of models have significantly evolved. Modern recommendation systems analyze not just individual likes but sequences of actions to infer probable intent behind them. For instance, multiple views of films by a specific director, an interest in French music, reading articles about classic cinema, and pausing on a photo of Catherine Deneuve—while these actions may seem trivial alone, together they can indicate a specific interest.
Research on recommendation systems explicitly employs the term latent intent—the user's hidden intention. Current algorithms model such intentions based on sequences of interactions to predict subsequent actions. Importantly, they do not need to guess the specific desire—narrowing down the options is sufficient.
Experiments with systems that extract dynamic intentions from behavioral history are ongoing. In 2025, researchers proposed the DUIP (Dynamic User Intent Prediction) framework, which combines an LSTM sequence model with a large language model: the LSTM encodes the evolution of interest, while the LLM uses the resulting "soft prompt" to predict the next item the user is likely to choose. In tests across multiple datasets (ML-1M, Games, Bundle), DUIP outperformed previous methods.
Another class of studies looks not at the entire history of views but at the current session: a few actions the user is taking right now. A 2025 paper in the ACM Transactions on Information Systems describes a model that uses LLM to extract multifaceted hidden intent from the session and makes recommendations based on it, showing improvement over basic approaches.
In other words, the direction of development is clear: the system aims to understand not just what the user has liked but what they are moving towards.
It is no secret that these systems not only analyze our tastes, beliefs, psychological traits, and interests but also shape them. This is where it gets truly unsettling.
You Might Think You Control the Process
How do we distinguish between recommendation and manipulation? An article published in January 2026 in the International Journal of Human-Computer Studies sheds light on this issue. Researchers examined how different methods of explaining recommendations influence choices.
In an experiment with 231 participants, various persuasive techniques were employed, including social proof, authority, scarcity, likability, and reciprocity. Convincing explanations not only affected choices but also increased the likelihood of selecting a less beneficial option. Essentially, the algorithm can determine which presentation method is most likely to lead you to choose a specific option from those available.
People who believe their critical thinking is impeccable when making decisions or judgments often fall into a trap. They assume that if the system offers a good option, it is a recommendation, but if it tries to push them toward something unnecessary, it is manipulation.
Imagine the algorithm has data indicating that a user is particularly susceptible to social approval. Instead of merely showing a product, it conveys that "thousands of people have already chosen this model." Another user, perhaps with signs of heightened anxiety, might see: "Only two left in stock." A third might see: "People with similar interests are choosing this option."
The product remains the same, but the argument varies each time. This is personalization, but it is not manipulation in and of itself. Manipulation is typically defined as influence that is covert or opaque, directing a person toward a specific decision by exploiting their vulnerabilities or bypassing their ability to assess the situation independently. Personalization merely creates the technical possibility for much more precise influence. Whether that becomes manipulation depends on the purpose behind the system’s use.
A recommendation system acts like a mirror that can change not just your reflection but also you. This occurs at several levels:
- Prediction. The system attempts to understand what interests the user and suggests it. For example, if a user frequently views travel content, they are shown more travel-related material.
- Nudging. The system does not just reflect interest; it selects which of several options to present to the user first.
- Persuasion. The system understands which argument will resonate with you and employs it in explaining the recommendation.
- Manipulation. The system deliberately utilizes traits of the user, which they may not be aware of, to subtly lead them toward a result favorable to the platform or advertiser.
- The system starts to change the user. It repeatedly shows certain themes and largely ignores others. Over time, some interests become more pronounced while others fade from view. The user continues to make choices, but within an environment that the algorithm is constantly reshaping.
One cannot definitively state that the system reflects their interests or imposes others. It does both simultaneously. It is based on your real actions. If you never click on political content, the system will not inundate you with it (at least not at first). However, by presenting a particular version of the world, the system affects what you will see next, which will, in turn, shape what you desire.
It's akin to a mirror showing you a smiling face, and you, seeing that smile, unconsciously begin to smile a bit more. After a hundred iterations, your "real" facial expression may no longer be yours. Researchers studying autonomy in the context of recommendation systems highlight that the influence extends beyond individual decisions, affecting the long-term shaping of identity, knowledge, and critical thinking.
The State Doesn’t Need to Spy on You
It is sufficient to control the information environment, personalize narratives, and utilize choice architecture.
Telecommunication operators, government services, platforms, and data brokers provide metadata: who communicates with whom, where they go, what they search for, what they buy, and what subscriptions they have. Detailed models of behavior and vulnerabilities among different population groups are built on this foundation.
In China, this logic has been systematized within the framework of a "social credit" system: data on purchases, movements, online activities, and violations are transformed into a rating that impacts access to loans, travel, and government services. However, even without a formal rating, segmentation alone suffices: "youth prone to protests," "elderly citizens sensitive to pension issues," "entrepreneurs reacting to regulatory changes." This opens the door to personalized influence—similar to the micro-targeting used by advertisers, but applied to political and ideological agendas.
Algorithms enable the creation of dozens of versions of the desired narrative, each resonating with specific fears, hopes, and values of the audience.
One group might see: "Stability is more important than change, or it will be like the '90s.'" Another: "Only a strong government can protect against external threats." A third: "Reforms are already underway; don’t disrupt the process." A fourth might receive a completely different set of topics: benefits, infrastructure, local issues.
The state can shape the digital environment so that some options are visible and accessible, while others are hidden or made difficult to navigate.
Practically, this manifests as follows:
- government news aggregators prioritize pro-government sources; alternative media may be absent or labeled as "undesirable," "foreign agents," or "extremist";
- in search engines and recommendation feeds, ranking algorithms elevate "approved" topics and push marginal ones down or into a "shadow ban";
- when accessing government services, users are presented with "recommended" options (for instance, consent to data processing, subscription to notifications) that are pre-selected as more convenient.
This is classic nudging—"pushing"—on a national scale with elements of personalization. Different groups may be shown different sets of "acceptable" options, creating the illusion that individuals have a choice, even though the list of options has already been filtered.
In research terms, this is called hypernudge—dynamic, networked, individualized nudging that is particularly difficult to control due to algorithmic opacity. Individuals do not feel direct pressure; they simply live in an environment where some paths seem natural while others appear strange, risky, or "not for people like me."
Propaganda No Longer Requires Sacrifices
The logic of prediction, recommendation, and influence that applies in marketing also operates at the level of state-individual relations. The stakes are higher: you will make what seems like an informed choice not between different manufacturers or bike models, but rather, for instance, between what to consider normal and what to perceive as a threat.
State influence through algorithmic systems operates at the level of the conditions of choice. It does not merely narrow the catalog; it determines which categories are available for thought, which questions can be asked, and which alternatives even come to mind. This is not manipulation of choice within the game; it is a change in the rules such that some moves become invisible.
In terms of French philosopher Michel Foucault, we do not always encounter repression in the classical sense ("you cannot speak about..."); rather, we face a power that produces certain forms of subjectivity, specific regimes of truth, and particular ways of being in the world.
The algorithmic environment does not prohibit you from thinking differently—it alters the way you think: making certain thoughts less likely, less natural, and less supported by your surroundings. This is not about preferences within an already established world; it is about the constitution of the world itself—what entities exist, what relationships are possible, which narratives are considered supported and approved, and which are marginalized or unthinkable.
When the state controls the information environment at an algorithmic level, it gains the ability to determine not only "which facts to show" but also "which epistemic frameworks are legitimate." The same fact can be framed within a narrative of "defending sovereignty" or "suppressing freedoms"—and the algorithmic amplification of one frame over another determines which interpretation becomes dominant.
This is not merely "propaganda" in the conventional sense of imposing a single message on everyone; rather, it is a personalized production of truth regimes: different groups are shown different epistemic universes, each with its own "obviousnesses," its own "common sense," and its own boundaries of what is permissible.
Subjectivity in a Loop
In the classical liberal model, the subject precedes choice: a person has preferences, values, and beliefs, and makes choices based on them. In an algorithmic environment with feedback loops, this sequence becomes blurred.
Preferences are not just "there"; they are formed through interaction with an environment that adapts based on previous reactions. The subject does not merely choose—they become someone who chooses in a specific way because the environment has made certain choices feel more natural and others more difficult.
We are talking not only about whom you will vote for in the next election. It is about the kind of subject you become through constant interaction with an algorithmically configured environment. What questions do you consider important? What alternatives seem real? What values are taken for granted?
There is no such thing as a neutral information environment; it has already been filtered and continues to be filtered anew with every action we take online.
Here, a new form of alienation arises. In the Marxist tradition, alienation occurs when the product of labor becomes an external force dominating the worker. In an algorithmic environment, alienation takes on an epistemic dimension: you exist in an information environment that you did not create, do not control, and do not fully understand, yet it dictates your preferences, beliefs, and decisions.
The issue of autonomy here is not that "you are being controlled," but that the boundary between "yours" and "imposed" becomes indistinguishable. Do you truly want what you want? Or do you want it because the environment has made other desires less likely?
This is particularly significant in the state context: loyalty, patriotism, agreement with policy—all of these can be "genuine" in a phenomenological sense (the person truly feels that way), yet they may also be products of an environment that systematically amplified certain narratives while marginalizing others.
Still Think Your Thoughts Are Being Read?
If so, it suggests you possess a well-functioning magical consciousness. Humans tend to remember hits and forget misses. If Facebook shows you a thousand posts in a day, and one happens to align with your thought (intentionally or not), that is what will stick in your memory. This creates an ideal machine for producing a sense of the supernatural: the algorithm makes countless probabilistic predictions, and individuals notice the most striking coincidences, later recalling them as a sequence of "first I thought—then it showed up."
This is a reality that must be accepted one way or another, especially since there is no clear answer to the question of what to do about it. Retreating from virtual reality will still amount to a retreat from one aspect of reality. No one can completely escape the influence of algorithmic work; it is also the environment in which you become who you are.
Recognizing that your informational environment is not neutral is one scenario for acceptance. Understanding that the feed you see is not the "objective world" but a version of the world configured by algorithms based on your previous reactions is a minimal form of hygiene. Your personal reflection will still be insufficient to resist the changes that will occur under the influence of networks.
The problem of algorithmic influence is not personal but structural. It requires regulation (transparency in recommendations, options to choose between different ranking modes, restrictions on micro-targeting in political advertising), institutional mechanisms (independent audits of algorithms, protection for researchers studying these systems), and cultural norms (recognizing that "personalization" is not just convenience but also power). There can be demands for choices not only within the algorithmic environment but also concerning the environment itself. However, how and by whom these demands will be met remains an open question.