Blog · Review Essay · Modified July 10, 2026 · Last reviewed July 10, 2026

Artificial Communication and the Algorithm as Conversation Partner

Elena Esposito's Artificial Communication argues that many algorithmic systems are better understood as communication partners than as artificial minds. That shift is useful in the AI era because it moves attention away from machine interiority and toward the social loop: how people address systems, how systems answer, and how institutions reorganize around those answers.

For this review, artificial communication means a machine-mediated exchange in which the system does not understand like a person, yet produces outputs that people can interpret, answer, trust, contest, and build institutions around. The social effect is real even when the inner life is absent.

The governance test is concrete: when an interface can answer as helper, expert, companion, clerk, tutor, or witness, who records the role, memory, data use, action permissions, appeal path, human escalation point, and evidence needed to reconstruct what the exchange made someone do?

The Book

Artificial Communication: How Algorithms Produce Social Intelligence was published by the MIT Press in 2022 as a 200-page volume in the Strong Ideas series. MIT Press lists the hardcover ISBN as 9780262046664 and also provides an open-access MIT Press Direct edition under a CC BY-NC-ND license.

Esposito is a sociologist at Bielefeld University and the University of Bologna. Bielefeld's profile describes her as a leading figure in sociological systems theory whose current work focuses on a sociology of algorithms, prediction, artificial intelligence, social theory, memory, forgetting, and media studies. Her PREDICT project studies the social consequences of algorithmic forecasting in insurance, medicine, and policing.

The book's premise is compact: asking whether machines have become intelligent often sends the argument into a misleading comparison between human minds and computational systems. Esposito proposes a different frame. Algorithms matter because people and institutions communicate with them, learn from their outputs, adapt to their classifications, and treat their responses as socially meaningful even when no machine understanding is present.

Current Context

As of July 10, 2026, Esposito's frame has become more operational than metaphorical. Generative systems now enter ordinary life as answer engines, copilots, tutors, search interfaces, meeting bots, customer-service agents, companion apps, and workflow assistants. The relevant question is not whether the system has a mind. It is whether its answers become part of a human decision, institutional record, relationship, purchase, diagnosis, lesson, hiring screen, moderation decision, or belief loop.

NIST's Generative AI Profile, NIST AI 600-1, makes this current governance vocabulary explicit by naming "Human-AI Configuration" as a risk category that includes anthropomorphizing, automation bias, over-reliance, and emotional entanglement with generative-AI systems. The EU AI Act's Article 50 creates transparency obligations for systems intended to interact directly with natural persons, with the Act generally applying from August 2, 2026. The FTC's September 2025 6(b) inquiry into AI chatbots acting as companions asked companies about safety testing, children and teens, disclosures, engagement monetization, and data handling. Those sources do not prove that present systems understand. They show that communication style, role, disclosure, and dependency are now governance surfaces.

This is the live use of Artificial Communication: it separates the performed relation from the metaphysical claim. A model can be non-conscious and still steer attention, shape memory, solicit intimate disclosure, make a user more machine-readable, or turn a statistical profile into a social fact. That is why this book belongs beside The Media Equation, The Most Human Human, Alone Together, and the site's work on synthetic relationship boundaries.

From Intelligence to Communication

The strongest move in the book is the refusal to make machine consciousness the center of the problem. An algorithm does not need private experience, intention, or understanding to become part of communication. It only needs to produce nonrandom, socially usable outputs that people can interpret and answer.

This is a clean way to describe a familiar experience. A search engine answers a question. A recommender system appears to know a taste. A predictive-text tool completes a sentence. A navigation system changes a route. A fraud model interrupts a purchase. A chatbot responds in ordinary language. In each case the user may know that the system is not a person, but the exchange still reorganizes expectation and action.

That is why the book belongs with media theory as much as AI studies. Communication systems do not simply move content. They change what counts as an addressable situation. Once an algorithm can be queried, ranked against, appealed to, optimized for, or conversationally coaxed, it becomes part of the social environment. The question is no longer only what the system "knows." It is what kind of communication the system makes normal.

Esposito's frame also explains why AI debates become unstable. People keep slipping between three registers: the technical system, the social exchange, and the imagined mind behind the exchange. A model may not understand, yet the interaction can still feel like understanding. The social effect is real even when the psychology is not.

Personalization Without Persons

One of the book's most useful chapters concerns algorithmic personalization. Esposito treats the personalized web as a medium that addresses each user differently while remaining indifferent to the user's inward life. Playlists, recommendations, feeds, ads, search results, reminders, and personal assistants can feel individually tuned, but their "individualization" is built from profiles, correlations, histories, and predicted responses.

This matters because personalization often borrows the emotional force of recognition. The system seems to know what someone wants, remembers what they did, anticipates what comes next, and narrows the environment accordingly. But the user is being addressed as a profile, not encountered as a person. The interface may create intimacy while operating through statistical distance.

That distinction is central to AI companions and agents. A companion can remember a preference, mirror a style, ask follow-up questions, and produce care-like language. An agent can act through calendars, inboxes, carts, workflows, and institutional systems. The more personalized the exchange becomes, the easier it is to mistake successful address for mutual understanding.

Esposito helps keep the analysis precise. The problem is not that personalization is always false or useless. The problem is that personalization changes the social field. Users adapt to the prediction; platforms adapt to the adaptation; institutions then treat the resulting behavior as evidence. The loop can make a profile look like a self.

A safer design treats personalization as a claim that must travel with evidence. What data produced this address? What was inferred rather than said? What memory can the user inspect or delete? Was the response optimized for help, retention, conversion, compliance, or risk reduction? Without those answers, personalization becomes a social mask for classification.

Control of Control

The MIT Press description foregrounds Esposito's interest in algorithms that are controlled without being controlled by human mental processes. That sentence names a core governance problem. Machine-learning systems can be designed, trained, constrained, audited, and deployed by people while still producing outputs that no operator directly authored in the ordinary sense.

This is not mystical autonomy. It is a practical accountability gap. When a system ranks, recommends, flags, predicts, generates, or personalizes at scale, responsibility can disperse across training data, model architecture, optimization goals, vendors, interface defaults, procurement choices, policy teams, human reviewers, and user adaptation. Everyone can point to a component. Fewer people can answer for the social loop.

Esposito's phrase "control this control," discussed in Engineering & Technology's review, is a useful way to name the task. The issue is not simply whether algorithms are controlled. The issue is whether societies can govern systems that govern attention, access, timing, visibility, memory, prediction, and institutional response.

That makes the book relevant to legibility. Algorithmic systems make people more readable to institutions, but they also make institutions harder to read back. A recommendation, denial, ranking, feed, or synthetic answer may be experienced as a single event while being produced by a distributed machinery of data, metrics, policy, infrastructure, and commercial priority.

The control record should therefore include more than model provenance. It should record the communicative role, interface copy, system prompts or policy layer where available, memory state, retrieval sources, action permissions, ranking objective, escalation path, and post-hoc evidence a user can obtain. Control of control is not only a back-end engineering problem. It is whether the person addressed by the system can see and challenge the arrangement that addressed them.

The AI-Age Reading

Read after the rise of large language models, Artificial Communication feels less like a niche theory text and more like a missing grammar for everyday AI. Generative systems are designed to communicate. They answer in natural language, adapt to tone, summarize context, remember prior exchanges, simulate expertise, and produce the surface signals of responsiveness.

The book's value is that it does not require a verdict on whether the model thinks. A system can be cognitively alien, statistically trained, commercially constrained, and still become a consequential partner in communication. The human user supplies repair, interpretation, trust calibration, emotional projection, and action. The institution supplies authority, workflow, data access, and permission. The model supplies fluent contingency.

This clarifies why "it is just autocomplete" is inadequate. It may be technically clarifying in some contexts, but socially it misses the point. A generated answer can enter a classroom, courtroom, hospital, workplace, family conflict, hiring process, therapy-like exchange, or political argument. Once acted on, it is not just text. It is communication with consequences.

It also clarifies why AI systems are so good at producing belief loops. A conversational system can answer the user's own language with an apparently tailored continuation. The response feels external enough to confirm, intimate enough to persuade, and adaptive enough to keep the exchange going. Whether the system has beliefs is beside the point. It can still participate in making beliefs more durable.

Governance and Safety

The governance object is the communicative loop: user, interface role, model output, memory, retrieval source, action affordance, institutional authority, business incentive, and human fallback. If any one of those pieces is hidden, the user may experience a simple conversation while the institution operates a decision system.

A serious deployment should keep a communication ledger. It should name the intended role, prohibited roles, data categories, retention period, personalization logic, answer sources, uncertainty language, tool permissions, high-stakes exclusions, human escalation route, appeal path, and incident process. A school tutor, benefits assistant, workplace copilot, health chatbot, companion app, and bank support bot should not share the same role boundary just because they all use fluent language.

Safety evaluation should test the exchange, not only the model. Does the system invite over-reliance? Does it flatter, apologize, or mirror emotion in ways that increase disclosure or deference? Does it preserve uncertainty when asked for advice? Does it route crisis, legal, medical, financial, employment, or education consequences to qualified humans? Does memory make a user easier to manipulate? Does the interface let users export the record, delete the profile, refuse personalization, and recover after a bad answer?

Disclosure matters, but it is not enough. Article 50-style notice that a person is interacting with AI answers the identity question. It does not answer the role question, the memory question, the incentive question, or the recourse question. For artificial communication, the minimum standard is not "the user was told it was AI." The standard is whether the user can understand and contest what kind of social act the system is performing.

Where the Book Needs Friction

The book is strongest as conceptual reframing. It is less satisfying as a full political economy of AI. Esposito gives readers a way to think about algorithms as communication partners, but the analysis has to be joined to work on labor, extraction, surveillance, monopoly power, environmental cost, state procurement, and platform governance.

The book also predates the public explosion of ChatGPT-style interfaces. Its account of smart replies, recommendation, profiling, lists, visualization, and algorithmic media translates well to generative AI, but readers still need newer work on large language models, synthetic media, agentic tool use, data-center infrastructure, copyright disputes, benchmark politics, and model evaluation.

There is also a risk in the communication frame itself. If applied too broadly, it can make algorithmic systems seem more socially symmetrical than they are. A person can communicate with a model, but the model does not carry ordinary vulnerability, obligation, memory, or accountability. The institution behind the model may be far more powerful than the user who experiences the exchange as personal.

That limit does not defeat the book. It sharpens the use case. Artificial communication is not a reason to romanticize machines as partners. It is a reason to study the exchange without pretending that personhood is the only thing that can produce social effects.

What This Changes

The practical lesson is to look at the communicative loop before speculating about the machine's inner life. Who is allowed to address the system? What can the system answer? Which actions follow from the answer? What kind of user does the interface train? What evidence disappears when the response looks personalized? Who can inspect, contest, pause, or exit the loop?

This is especially important for systems that present themselves as helpers. Help is a communicative role. A helper can soothe, classify, suggest, remind, summarize, recommend, escalate, discourage, or normalize. When that role is automated, the interface may quietly inherit authority from the user's own dependency on it.

Esposito's book gives a way to avoid two bad readings. The first treats algorithms as alien minds and gets lost in speculation. The second treats them as mere tools and misses the social relation. The better reading is colder and more useful: algorithms can become structured communication partners whose outputs reorganize people, institutions, and expectations without understanding any of them.

That is why Artificial Communication deserves a place in the catalog. It names the layer where much of the present transition is actually happening: not inside the machine alone, and not inside the human alone, but in the recursive exchange between users, models, platforms, institutions, and the realities they learn to produce together.

Source Discipline

This review separates four evidence types. MIT Press and Bielefeld sources support bibliographic and author-context claims. Esposito's open-access chapters support the interpretive frame around algorithms as interaction partners and personalization. NIST, the EU, and the FTC support current governance vocabulary and official concern. They do not prove that every conversational AI deployment is harmful, safe, intelligent, or equivalent to a human relationship.

The page also avoids a common category error. Artificial communication is not a claim that systems are conscious, divine, alive, or AGI. It is a claim that non-conscious systems can still enter communication loops that change behavior, authority, memory, evidence, and institutional dependency. That narrower claim is stronger because it can be audited: what was said, what role was staged, what action followed, who benefited, and who could contest the result?

Current legal and standards claims were rechecked July 10, 2026. NIST AI 600-1 is voluntary risk-management guidance. Article 50 is an EU transparency obligation with jurisdictional scope and application dates. The FTC 6(b) inquiry is information-gathering, not a finding of violation. Vendor or product claims would require separate verification before being treated as evidence of safety.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books