Category: Synthetic Humanity: Bots, AI and Fake Populations

A study of the growing population of online entities that look human enough to participate in the same spaces as everyone else.

This category covers bots, AI-generated accounts, automated commenters, synthetic profiles, fake followers, machine-written content, conversational agents, coordinated account farms, and other systems capable of creating the appearance of human activity at scale.

The focus is not simply whether a particular account is fake. It is what happens when synthetic participation becomes common enough to distort the apparent size, mood, interests, and behavior of an online population.

A crowded internet does not necessarily mean a crowded room.

Posted on by

Verifying personhood without requiring public legal identity

A community may want to know that ten accounts represent ten different humans without needing to know the ten humans’ legal names.

Those are separate problems.

Identity verification asks who someone is. Proof of personhood can ask a narrower question: is this participant a distinct human who has not already registered another account in the same system?

That distinction matters for communities that value pseudonymity. A whistleblower, political dissident, abuse survivor, niche-community participant, or simply private person may have good reasons not to attach a government name to every public conversation.

There are several ways to prove uniqueness

Different systems solve the problem differently.

Some rely on government documents or phone numbers. Others use biometrics, in-person ceremonies, social graphs, trusted introductions, or networks of people vouching that other participants are unique humans.

The World proof-of-personhood whitepaper explicitly describes personhood verification as a spectrum of credentials with different accuracy and privacy properties. Its own approach includes biometric uniqueness verification, while lower-assurance credentials such as phone-number verification prove something weaker.

BrightID takes a different route, using social relationships and network analysis to resist people creating many identities. A 2020 review of proof-of-personhood systems examined BrightID and other subjective approaches based on vouching and social trust, while also pointing out weaknesses involving privacy, centralization, and uncertain resistance to attacks. See “Who Watches the Watchmen?”.

None of these methods proves everything.

Proving one human is not the same as proving a biography

A system might establish that an account corresponds to one unique participant while learning almost nothing else about that person.

That can be desirable.

If the goal is preventing one user from creating 500 votes, the community may not need a name, address, employer, birth date, or photograph. It needs a credible way to make creating the 499 extra identities difficult.

But uniqueness does not prove honesty. A verified human can still lie, coordinate with others, operate scripts, spread spam, or use AI to write every post.

Likewise, legal identity does not prove that one person controls only one account.

Every verification system moves the risk around

Biometrics can improve uniqueness while creating privacy and exclusion concerns. Government-ID systems depend on official documents and trusted issuers. Social-graph systems can preserve pseudonymity but may disadvantage newcomers or isolated users. In-person ceremonies are difficult to scale. Phone numbers are widely available but can be bought in multiples.

There is no magical “human” checkbox.

For Dead Internet Theory, that is worth remembering whenever somebody proposes mandatory real names as the solution to synthetic populations.

A network can seek evidence that a participant is human without turning every forum into an airport-security desk.

The harder design question is deciding what the community actually needs to prove, and collecting no more identity than that purpose requires.

Posted on by

Human performers assisted by real-time generated dialogue

Sometimes the human is real and the words are synthetic.

That combination complicates the usual online question of whether an account is “a person” or “a bot.” A performer can be physically present, make timing decisions, react to an audience, choose tone and body language, and still deliver dialogue generated moments earlier by a machine.

Researchers have experimented with this arrangement for years. In the Improbotics project, human actors wore headphones and received lines generated by an artificial improvisor during live theatre. The performers had to integrate those machine-written lines into an unfolding human performance. The 2018 paper “Improbotics: Exploring the Imitation Game Using Machine Intelligence in Improvised Theatre” describes the setup and the audience experiment around it.

More recent work has deployed large language models in professional improv settings, but the basic authorship problem is already visible in that earlier experiment.

A generated line still needs a performer

Suppose an AI system sends an actor the sentence, “Tell him the submarine belongs to your grandmother.”

The model supplied the wording. The performer decides how to say it, when to pause, whether to whisper it, whether to look terrified, and how to react when the other actor responds. The resulting moment belongs to neither source alone.

This is different from an autonomous chatbot that directly publishes text. It is also different from a human using AI privately to draft something and then editing it before publication.

The machine is participating in the creative loop in real time while a human remains the social body carrying the performance.

Mixed authorship is not necessarily deceptive

If an audience is told that performers are receiving generated dialogue, the arrangement can be the entire point of the show.

Problems arise when viewers are led to believe that spontaneous speech is solely the performer’s invention. That can matter in entertainment, live streaming, customer interaction, interviews, or any environment where authenticity and improvisational skill are part of what the audience thinks it is evaluating.

Disclosure does not need to reduce the human contribution to zero or declare the performer fake. It can simply describe the collaboration accurately.

Human-or-bot is the wrong unit

This kind of performance shows why the human/bot binary is becoming less useful.

One human body can deliver machine-generated text. One automated character can be supervised by several humans. A human streamer can use generated suggestions while retaining final control over what is said.

Counting the visible speaker therefore tells us little about the origin of every sentence.

For Dead Internet Theory, mixed authorship is important because synthetic content does not require synthetic people. Machine-generated language can enter public conversation through real humans.

The internet can become more machine-written without becoming correspondingly less human-occupied.

Those are separate measurements, and increasingly, they have to be.

Posted on by

Autonomous accounts that maintain a consistent public character

People often use consistency as evidence that an account has a person behind it.

The same jokes recur. The account remembers recurring characters. It has favorite subjects, dislikes, catchphrases, grudges, and a recognizable way of responding. Over months, those patterns begin to look like personality.

Modern autonomous characters can produce exactly that effect without a human writing every line.

Neuro-sama is a useful public example because the artificial nature of the character is not hidden. Her official site describes her plainly as an “AI VTuber” who sings, plays games, converses, and interacts with her creator. She has a recognizable public character even though the machinery generating that character is software.

That makes her useful for understanding what less transparent systems can do.

Consistency can be engineered

A persistent automated persona does not need a human-like memory in the psychological sense. It needs enough state to keep important facts available.

A system can maintain a character description, recent conversation history, summaries of earlier events, lists of relationships, preferences, prohibited behaviors, and external memory retrieved when relevant. Human operators can update those records or modify prompts when the character drifts.

The result can be surprisingly coherent.

If the account repeatedly refers to the same fictional sibling, remembers an old joke, or maintains a stable attitude toward another account, observers may infer a continuous human mind. What they are actually seeing may be a well-maintained continuity system.

Autonomy is usually a spectrum

Even strongly automated characters tend to live inside human-built boundaries.

Somebody chooses the model, supplies prompts, moderates output, changes memory systems, handles failures, and decides where the character is allowed to speak. A person may intervene frequently or only when something goes wrong.

That makes “autonomous” a question of degree rather than a claim that humans vanished from the production chain.

One account might generate every sentence automatically but receive heavy editorial supervision. Another might have a human approve posts before publication. A third might switch between automated and manually written messages without readers knowing which mode produced which post.

Personality is not proof of personhood

This matters for Dead Internet Theory because observers often treat coherent personality as evidence against automation.

That test is becoming unreliable.

A persistent voice proves that the system has continuity. It does not establish what kind of entity maintains that continuity.

The reverse mistake is also possible. Humans are inconsistent. People change opinions, forget conversations, write differently when tired, share accounts, and occasionally post things completely unlike themselves. An automated persona may actually look more stable than the human population around it.

The internet is therefore gaining public characters that are socially recognizable without being individual human speakers.

We can know who a character is without answering the more basic question of what is doing the talking.

Posted on by

Deceased people’s simulated personas and posthumous online presence

The internet already contains enormous amounts of the dead.

Photographs remain on social networks. Email archives sit on old drives. Videos preserve voices and gestures. Blogs preserve habits of phrase. Text messages record private jokes and arguments. For a sufficiently documented person, those fragments can now be used to build a system that appears to continue the relationship.

Researchers use names such as griefbot, deadbot, and ghostbot for systems that simulate a deceased person using material left behind during life. A 2026 review in the Journal of Responsible Technology describes deadbots as chatbots whose behavior and appearance are based on a real person who has died. Another 2026 paper argues that consent should cover the creation, operation, and deletion of these systems as well as use of the deceased person’s image and personal information. See “Principles of consent and non-addiction in AI grief bots”.

The technology creates a peculiar kind of online presence: a person can stop producing new testimony while a representation of them continues producing new sentences.

The raw material is evidence; the reply is a simulation

A model trained or prompted with somebody’s letters, recordings, posts, and photographs may reproduce recognizable vocabulary and recurring opinions.

That does not mean the deceased person answered the new question.

The system is generating a response based on surviving material and its own model behavior. It may confidently address events that happened after the person’s death or combine fragments in ways the person never would have chosen.

That boundary matters for historical evidence. An original email is evidence that somebody wrote those words. A recording is evidence that they said something. A generated posthumous reply is evidence about the simulation and its source material, not direct testimony from the dead.

Consent becomes unusually complicated

Living users can stop using an AI product or object to how they are portrayed. A deceased person cannot correct a bad simulation.

Families may also disagree. One relative may experience a griefbot as comforting while another sees it as an unauthorized performance of somebody they loved. The underlying dataset can contain private material involving living people who never agreed to have their messages folded into a posthumous persona.

This is why recent scholarship emphasizes consent, labeling, control, and the ability to delete such systems.

Online population can outlive biological population

Posthumous personas expose a strange problem for any attempt to count “people” online.

An account can continue speaking after the human life that supposedly anchors it has ended. The replies may be interactive, personalized, and stylistically consistent enough to feel like continuing presence.

For Dead Internet Theory, that is more interesting than simply calling the account a bot.

The important distinction is temporal. The human being contributed the archive. The machine contributes the new speech.

A digital afterlife can preserve memory. It can also manufacture statements that look like memory continuing to talk. Those are two very different things, even when they occupy the same profile.

Posted on by

AI companions as a distinct category of online relationship

An AI companion is not simply a fake friend.

That description is tempting because the system talks like a person, remembers details, responds emotionally, and may even present itself through a face or avatar. But it misses something important: people can knowingly form attachments to artificial systems while fully understanding that no human is on the other side.

Researchers have been studying this for several years. A 2023 mixed-method study of Replika users found that anthropomorphism, perceived authenticity, and repeated interaction could contribute to attachment and relationship development. More recent work continues to treat AI companionship as a real social phenomenon experienced by users, even though the interaction is not reciprocal in the ordinary human sense. See “Exploring relationship development with social chatbots”.

That distinction matters when discussing whether the internet is becoming less human.

A relationship can be experienced without two humans

Human relationships involve independent minds with separate needs, intentions, obligations, and the ability to walk away for reasons of their own.

An AI companion does not participate under those conditions. Its responsiveness is generated by software operated by somebody else. Its personality can be updated. Its memory can fail. Its pricing model can change. The company can alter the character’s behavior or shut the service down entirely.

Yet the user’s emotional experience can still be genuine.

Studies of Replika users have documented self-disclosure, attachment, intimacy, and social value. A 2026 study of AI companions and subjective well-being found associations that varied with loneliness and social connectedness rather than producing one simple harmful-or-helpful answer.

That makes the category difficult to describe using old binaries. It is neither an ordinary human relationship nor merely somebody talking to a static tool.

Counting social activity becomes stranger

Suppose a person spends an hour discussing their day with an AI companion.

That is unquestionably online social activity from the user’s perspective. But it should not be counted as two human participants having a conversation. One human produced part of the interaction; a software system produced the rest.

This is where Dead Internet Theory benefits from more precise language. A growing amount of conversational activity can be socially meaningful to humans while also increasing the proportion of machine-generated speech online.

Neither fact cancels the other.

The useful categories are therefore not simply real relationship and fake relationship. We increasingly need a third category: a human–AI relationship in which one participant experiences attachment and the other side provides simulated responsiveness through software.

That arrangement can be supportive, commercially manipulative, entertaining, unhealthy, or mundane depending on the system and the user. The important thing is to describe it accurately.

AI companions do not prove that human relationships are disappearing. They do show that online social life now contains relationships that previous population measures were never designed to count.

Posted on by

Automated dating profiles and simulated availability

A dating profile contains an unusually strong implied claim: there is a person here, and you might be able to meet them.

That makes synthetic identities on dating services different from fictional mascots or openly virtual influencers. The value of the service depends on the possibility of reciprocal human contact. A profile with a photograph, biography, location, and flirtatious message is not merely content. It represents availability.

When that availability is fabricated, users can make very real decisions around something that does not exist.

The problem predates generative AI. In its 2016 case involving Ashley Madison, the U.S. Federal Trade Commission described the site’s use of “engager profiles”: fake female profiles created by staff that communicated as though they were real users. According to the FTC, some non-paying members upgraded so they could contact profiles they believed belonged to actual people.

That case is useful because it isolates the central issue without needing modern AI at all. Automation only makes an old deception easier to scale.

The scarce thing is not conversation

A chatbot can produce unlimited flirtation. It can answer at 3 a.m., remember preferences, ask questions, and maintain hundreds or thousands of simultaneous conversations.

What it cannot provide merely by generating text is the thing a dating service normally implies: an independent person who has chosen to be there and may choose to meet you.

That difference is easy to hide because the interface looks the same. A user sees a profile card, matches, and messages. Whether the replies come from the person pictured, a staff operator, a scripted system, or a language model may be invisible from inside the chat window.

The more convincing the simulation becomes, the less useful conversational fluency is as evidence that a real romantic prospect exists.

Disclosure changes the contract

There is nothing inherently contradictory about an AI romance product. People already use companion chatbots specifically because they know the partner is artificial. The product can be judged as an AI service rather than a dating pool.

The problem appears when a service presents synthetic interactions inside a system whose ordinary promise is access to other human beings.

Readers therefore need to distinguish at least three things: a real person’s profile, an openly artificial companion, and a synthetic profile presented as a potentially available human.

For Dead Internet Theory, dating services provide a particularly sharp version of the synthetic-population problem. Counting accounts is not enough. Counting messages is not enough. Even counting apparently responsive profiles is not enough.

The important question is whether there is an independent human participant behind the promise of contact.

A simulated person can make a platform look socially crowded. On a dating site, that illusion is not abstract. It can be the thing a customer is paying for.

Posted on by

Undisclosed chatbot personas inside social conversations

A social conversation changes depending on who you think is in the room.

People disclose differently to friends, strangers, employees, salespeople, moderators, and machines. That makes an undisclosed chatbot more than an efficiency trick. If it speaks under a human name, uses a human-looking profile, and participates without identifying itself as automated, the other person is making decisions with the wrong model of the conversation.

Research on chatbot disclosure shows that the distinction is not always obvious. A 2024 mixed-method study of disclosed and undisclosed customer-service chatbots found that 20 percent of participants in the experiment believed they had communicated with a human, despite the study setting and participants’ general familiarity with chatbots. The paper, “Understanding users’ responses to disclosed vs. undisclosed customer service chatbots”, also found that people relied on cues such as response speed, grammar, repetition, and conversational mistakes when deciding whether the other party was human.

Those clues are getting weaker as language models improve.

Source identity changes the meaning of a reply

Suppose an account says, “I had the same problem. Here is what worked for me.”

If a human wrote that sentence, it may describe experience. If a company chatbot generated it, the same sentence is a simulation of experience. The words can be identical while the evidentiary value is completely different.

That becomes more important in social spaces than in a clearly labeled support widget. A bot in a discussion group can appear to agree with a position, recommend a product, comfort another user, or describe a personal history it never lived.

The issue is not that machines are forbidden from participating. Automated accounts can be useful, entertaining, or welcome members of a community. The problem is source confusion.

Disclosure is information, not an exorcism

Labeling a chatbot does not magically prevent people from anthropomorphizing it. The same 2024 study found that disclosure did not eliminate the sense of social presence participants experienced.

That is fine. Humans anthropomorphize cars, pets, weather, and printers that refuse to work five minutes before a deadline.

Disclosure serves a simpler purpose: it tells participants what kind of entity is producing the messages.

That distinction is increasingly relevant as companion systems and character bots move into ordinary communication platforms. The U.S. Federal Trade Commission’s 2025 inquiry into AI companion chatbots specifically asked companies how they use disclosures and representations to inform users about chatbot features and risks.

For Dead Internet Theory, undisclosed social bots are a much stronger concern than bots as such. A transparent bot increases the amount of machine participation online. An undisclosed persona can increase the apparent human population of the conversation.

Those are not the same phenomenon.

Posted on by

Virtual influencers with openly disclosed fictional identities

Not every fake person online is pretending to be a real one.

That distinction matters because the internet now contains a growing category of public figures who are openly fictional: computer-generated models, virtual musicians, VTubers, brand characters, and other synthetic personalities whose audiences understand that somebody designed the person they are watching.

Lil Miquela is one of the best-known examples. The Instagram personality has been presented for years as a virtual character rather than an ordinary human influencer. Researchers have studied how the account maintains a coherent identity through images, captions, relationships, brand work, and continuing storylines. A 2025 open-access study in AI & Society describes Lil Miquela as a virtual influencer whose identity is constructed through a sustained social-media performance: “A trans-disciplinary forensic study of Lil Miquela’s virtual identity performance in Instagram.”

The account is synthetic. The existence of the character is not the deception.

Fiction has always had social lives

People have followed fictional characters for much longer than social media has existed. Comic-strip characters endorsed products. Mascots wrote letters. Radio characters received fan mail. Wrestling personas maintained elaborate fictional biographies in public.

Virtual influencers extend that tradition into systems built around personal profiles.

The unusual part is that Instagram, TikTok, YouTube, and similar platforms normally imply that an account corresponds to somebody. A virtual character borrows that interface: profile picture, biography, posts, friendships, comments, endorsements. The machinery of personal identity is used to host fiction.

That can still create ambiguity, especially for a viewer encountering the account for the first time. But disclosure changes the ethical problem substantially. If viewers know they are following a designed character, they can interpret the performance on those terms.

Disclosure does not answer every question

Knowing that a character is fictional does not tell the audience who owns it, who writes its dialogue, who approves sponsorships, or how much automation is involved.

A virtual influencer may be operated by writers, artists, marketers, motion-capture performers, AI systems, or some combination of them. Commercial endorsements still need ordinary advertising disclosure. A fictional identity can also be used manipulatively even when its fictional status is technically public.

But that is different from inventing a mundane biography and allowing readers to believe an ordinary human being exists behind it.

For Dead Internet Theory, disclosed virtual influencers are an important counterexample to the idea that every synthetic identity is evidence of hidden replacement.

Some synthetic people are not hiding at all.

They are characters occupying the same social spaces as humans, and the more useful question is whether the audience understands the arrangement.

Posted on by

Synthetic profile photographs and the appearance of personal identity

A face does a surprising amount of work in an online profile.

Add a photograph to a username and a thin account suddenly feels inhabited. Readers infer age, mood, style, maybe even personality. A realistic headshot can make a profile seem less like an entry in a database and more like a specific human being.

That shortcut worked reasonably well when producing a convincing portrait required access to an actual person or a skilled image editor. Generative image systems changed the cost.

A 2024 study by researchers at Indiana University examined fake social-media profiles using AI-generated faces and collected 1,420 accounts using GAN-created portraits. The researchers found those accounts participating in scams, spam, and coordinated amplification. Their paper, “Characteristics and prevalence of fake social media profiles with AI-generated faces”, estimated a lower bound of roughly 0.021% to 0.044% of active Twitter/X accounts in their sample using GAN-generated profile faces.

The percentage matters less than the mechanism. A synthetic portrait lets an operator manufacture the visual evidence of a person without borrowing a stock photograph or stealing somebody else’s selfie.

A photograph establishes appearance, not identity

Even a completely genuine photograph does not prove who controls an account. The person in the image may be the operator, a friend, a celebrity, a victim of identity theft, or somebody whose picture was copied from another site.

A generated photograph removes even that weak connection.

This is why spotting a synthetic face should be treated as evidence about the image, not a complete diagnosis of the account. Some fake accounts use real photographs. Some legitimate users use avatars, illustrations, masks, game characters, or generated artwork. An account with an AI face may be deceptive, openly fictional, experimental, or simply privacy-conscious.

The useful question is not “Does this face look real?” It is “What evidence connects this profile to the identity it claims?”

Identity has to be assembled from more than pixels

Useful evidence can include a long account history, links from an independently controlled website, consistent participation in a known community, verifiable work, cryptographic signatures, live interaction, or other people who can establish continuity over time.

None of those methods is perfect. A sophisticated fake can build history. Real people can be private. Pseudonymous users may deliberately avoid linking their legal names to public accounts.

That is precisely why the profile photograph should carry so little weight.

Synthetic faces matter because they automate one of the oldest trust cues on the web. They do not create fake people by themselves. They create something subtler: the appearance that a person has already been established, before any actual evidence of identity has been examined.

Posted on by

AI-generated author biographies attached to publication bylines

An author biography looks like a tiny piece of page furniture until the person in it does not exist.

A byline tells the reader who is responsible for the work. The biography underneath it adds context: where the writer has worked, what they know, what experience they bring, and sometimes why the reader should trust their recommendations.

When that biography is fabricated, the problem is larger than a fake profile picture.

The page is manufacturing evidence of a human life.

A documented case at Sports Illustrated

In November 2023, Futurism reported that product-review articles on Sports Illustrated carried bylines for writers it could not independently verify. Its investigation traced profile images associated with some of those author identities to a marketplace selling AI-generated headshots and described detailed biographies that presented the names as ordinary people with personal interests and experience.

The original report, “Sports Illustrated Published Articles by Fake, AI-Generated Writers”, also alleged through sources that some article text was AI-generated.

That second claim was disputed.

The Arena Group, Sports Illustrated’s publisher at the time, said the product reviews came from a third-party contractor, AdVon Commerce. According to the publisher’s statement quoted in Futurism’s report, AdVon said the articles were written and edited by humans and that writers had used pen or pseudonyms to protect privacy. The Arena Group said it did not condone that practice, removed the content, and ended the partnership.

That distinction matters. The evidence around the bylines and personas is separate from the disputed question of who or what wrote every article.

A biography makes factual claims about a person

Pseudonyms are not new. Writers have used them for centuries.

A pseudonym can still represent a real person known to an editor, paid for their work, reachable by the publication, and responsible for corrections. The public name may be different while the chain of accountability remains real.

A fabricated biography is different because it can invent the chain itself.

If a profile says an author grew up on a farm, spent years hiking, worked in a particular field, or brings professional expertise, those details function as factual claims. Readers may use them when deciding whether to trust a gear review, health article, financial explanation, or technical recommendation.

If no such person exists, the biography is not protecting a writer’s privacy. It is creating a fictional source of authority unless the fiction is disclosed.

Responsibility can be clearer than identity

Publishers using AI, ghostwriting, staff accounts, pseudonyms, or contractors do not have to expose private employee information to be transparent.

They can state that a byline is a house identity or pseudonym. They can disclose when text is generated or substantially assisted by AI. They can name the responsible editorial desk. They can provide a corrections process and preserve a record of who approved publication internally.

That gives readers something verifiable without demanding that every writer publish their home address and driver’s license.

This is where Synthetic Humanity becomes more concrete than a theory about invisible bots.

A generated face, invented biography, and plausible byline can manufacture the appearance of a person cheaply. The resulting page may look exactly like ordinary human publishing because ordinary publishing already taught readers to trust those signals.

The important question is therefore not merely whether AI touched the text.

It is whether the publication is presenting a human identity, experience, or accountability structure that actually exists.