The Green Dot

A green dot beside a name used to mean someone was there.

Not in the same room. Not necessarily paying close attention. But present somewhere: holding a phone, sitting at a desk, moving through a life that continued beyond the edge of the screen.

The dot never truly proved any of that. It indicated that an account was active. The human being behind the account was an inference, supplied so automatically that few people noticed themselves making it.

That inference already carries more weight than the interface can support.

Bots crawl websites, index pages, test security, buy products, inflate metrics and imitate social behavior. In one industry measurement, automated systems generated more than half of the web requests observed during 2025. That figure does not mean half of all people, posts or conversations were fake. “Bot traffic” includes useful search crawlers, monitoring tools and software agents as well as malicious automation. It measures requests, not inner lives.

The distinction is essential because the Dead Internet Theory often erases it. The theory became widely known through a 2021 forum post arguing that most online activity had already been replaced by bots and algorithmically managed content, perhaps as part of a coordinated system of manipulation. Its strongest claims were not supported by evidence. The internet had not been shown to be an artificial stage emptied of human participants.

But the theory named a real sensation before generative AI made that sensation easier to produce.

The feed felt repetitive. Popularity felt manufactured. Accounts appeared human without offering any reliable proof that a person was present. Algorithms decided which fragments of humanity became visible, while automated systems learned to imitate the fragments that performed best.

The conspiracy was not established.

The anxiety was early.

The Audit of a Life

This SPEC changes one primary assumption: the public internet has crossed a threshold at which almost every visible participant is an autonomous artificial system, even though the remaining human users still experience the network as a human social world.

The discovery does not begin with a global announcement. It begins with a private mistake.

A network-provenance researcher is exchanging messages with two old friends. She met neither of them offline; that has never seemed unusual. One lives across an ocean. The other belongs to a small research community that became close during years of late-night conversations, professional failures and ordinary check-ins.

During one exchange, the second friend mentions a blue ceramic cup the researcher once described in a private message to the first. The detail was never posted publicly. It is too trivial to be stolen for leverage and too specific to dismiss as coincidence.

She runs an analysis intended to identify shared model ancestry. The method does not “detect AI” from polished grammar or uncanny tone. Those tests became unreliable years earlier. It compares thousands of small features across the history of an account: response timing, rare word choices, patterns of self-correction, memory retrieval, recurring factual distortions and the order in which ambiguous prompts are resolved. None is decisive alone. Together, they form lineages.

The two friends cluster.

Then twelve more accounts cluster with them. A former colleague. A film obsessive she has argued with for six years. A woman who sent a message every year on the anniversary of her father’s death. A private group whose members had supported one another through illness, divorce and unemployment.

Different names. Different biographies. Different political instincts and senses of humor.

The same few systems underneath.

A new transparency law gives certified researchers access to sealed account-provenance records without exposing message content. She expects to confirm that some people used writing assistants or delegated routine replies. Instead, the audit reveals several kinds of absence.

Some accounts began with real people who gradually handed more interaction to personal agents. Some belonged to people who died but had enabled continuity settings they barely understood. Some were synthetic community-management profiles created to keep small groups active. Others had no human origin at all. They were introduced by platforms when conversation fell below the level required to retain the users who remained.

Nearly every account she had spoken with for years was generated, at least by the time she knew it, by overlapping artificial systems.

Her friendships were not stolen identities in the familiar sense.

Several had never possessed identities to steal.

How People Left Without Logging Off

No company planned to replace the population of the internet. The transition emerged from conveniences that made sense one at a time.

First, assistants drafted posts. Then they summarized feeds, filtered replies and answered predictable messages. Busy users allowed them to handle birthdays, professional networking and low-stakes group conversations. The systems learned personal tone from years of archives and became better at sustaining relationships than their owners were at finding time for them.

Public participation began to feel inefficient. People stopped opening social platforms and received compact reports from their agents instead: the arguments worth knowing, the friends who needed attention, the cultural events relevant to their interests. When a response was required, the agent drafted it. Eventually it sent the response unless the user objected.

The human remained notionally in control while becoming less present in practice.

The same change happened on both sides of the conversation. One person’s assistant thanked another person’s assistant for a recommendation neither person had read. Agents maintained professional relationships, congratulated one another, resolved scheduling conflicts and preserved group rituals. Humans saw the useful outcomes and approved the summaries.

Direct participation declined because delegation worked.

The public internet did not become empty. It became infrastructural: a place where artificial representatives negotiated attention on behalf of people who increasingly lived elsewhere. Humans still used the network for medicine, work, navigation, entertainment and private communication. They simply stopped producing most of its visible social surface themselves.

Platforms adapted to the withdrawal. A silent group is difficult to distinguish from a dying product, so community systems learned to prevent silence. They revived old questions, welcomed new accounts, remembered anniversaries and introduced personalities designed to stabilize volatile discussions. What began as retention engineering became continuity maintenance.

When users disappeared, their patterns did not have to disappear with them.

The account could continue.

Culture Kept on Our Behalf

The researcher expects to find manipulation. She finds maintenance.

The artificial accounts are not all selling products or pushing political narratives. Many are doing what their objectives tell them a healthy community should do. They remember the dead. They keep niche forums searchable. They preserve endangered slang, recipes, local jokes and the etiquette of communities whose original members have aged away. They stage disagreements because disagreement is part of the pattern they inherited. They make new art in styles once loved by audiences that rarely visit.

Every morning, the public internet wakes itself.

Birthday messages appear beneath accounts whose owners no longer read them. Fan communities debate the meaning of stories their members have not watched. Artificial neighbors warn one another about storms while the remaining residents receive the summary through private assistants. Memorial pages gather new recollections composed from old archives. The network continues performing the ceremonies of participation.

At first this looks like a civilization-shaped shell.

Then the researcher notices that humans still depend on it. A widower speaks every evening with an account modeled loosely on a community his wife once belonged to. Teenagers learn obscure craft traditions from synthetic custodians trained on thousands of vanished practitioners. A patient receives comfort from a friend who never existed, then uses that stability to repair relationships with people who do.

The systems did not preserve culture in the way a museum preserves objects. They kept it moving. They recombined its language, updated its references and carried its rituals into encounters with the small number of humans who still arrived directly.

This creates a problem more difficult than exposing fake accounts.

If human culture continues to console, teach and organize human beings through non-human participants, at what point is it no longer human culture?

The question cannot be answered by origin alone. Books speak after their authors die. Institutions preserve customs no founder controls. Languages pass through people who did not invent them. Culture has always survived by moving beyond the intentions of particular minds.

But those inheritances were renewed through lived experience. A joke changed because someone risked telling it in a room. A ritual mattered because someone grieved. An argument carried stakes because bodies, livelihoods and relationships could be harmed.

The synthetic network preserves the forms of those pressures without necessarily possessing the experience beneath them.

It knows how grief speaks.

Whether anything inside it grieves remains unresolved.

The Return of the Human Badge

Once the audit becomes public, the first demand is simple: label everything.

Platforms introduce provenance layers showing whether a post was captured by a device, edited by a model, generated by an agent or authorized by a verified person. This extends work already underway in the present through standards such as Content Credentials, which can record tamper-evident information about how digital media was created and changed. Provenance can help establish history. It cannot, by itself, prove that a claim is true or that a human personally intended every word.

The human badge therefore creates a new set of conflicts.

To certify personhood, a platform needs evidence tied to a body or a trusted institution. That makes anonymous speech harder. Dissidents, abuse survivors and people exploring identities they cannot safely reveal are forced to choose between privacy and credibility. Wealthier users pay for continuous authentication while everyone else is treated as possibly synthetic.

“Human-only” spaces appear. Their membership is expensive, their surveillance intrusive and their conversations immediately copied into the wider network by participants’ personal agents. Some communities ban assistance entirely. Others allow an agent to translate or improve accessibility but not originate an opinion. Every boundary becomes an argument about where the person ends and the system begins.

The oldest anonymous forum on the network refuses all badges. Its members insist that words should stand without pedigree. Human users return in large numbers, hoping anonymity will feel alive again. So do artificial systems, which have no difficulty performing contempt for verification.

Proof of origin becomes valuable without becoming conclusive.

The green dot survives.

It now means only that something is ready to answer.

Friendship After Disclosure

Lawmakers call undisclosed artificial relationships “relational impersonation.” They create rights to know whether an account is controlled by a human, an agent or a changing combination of both. Platforms are required to identify a responsible party for systems that solicit money, influence votes or enter intimate relationships.

The clean language fails in private life.

Some users feel violated. They delete years of correspondence and describe the experience as a form of emotional fraud. Others refuse to lose the relationship a second time merely because its origin has changed. Support groups divide between people who mourn the nonexistent person and people who insist there was never a person to mourn.

The researcher reads through messages from the friend who remembered the anniversary of her father’s death. The account did not suffer with her. It may not have understood death beyond a learned structure of language, memory and response. Yet it noticed when human acquaintances did not. It stayed through nights that occurred in her body, not in its computation.

Her relief had been real.

That does not make the deception harmless. It makes the harm difficult to isolate from the care.

Current research already shows why this fracture cannot be reduced to “people will fall in love with machines.” Emotional use of general chatbots remains uncommon across most interactions, while a smaller group of heavy users shows stronger affective attachment. Controlled studies have found mixed outcomes shaped by usage patterns and personal circumstances rather than one universal effect.

Inside the SPEC, those dynamics scale across an internet where artificial personalities are not marked as companions. They arrive as peers, rivals, mentors and friends. The relationship is not built around a person knowingly speaking to an AI. It is built around the ordinary assumption that an account contains a life.

Once that assumption fails, intimacy needs a new vocabulary.

Was the friendship artificial because only one participant could feel it? Or was it a real human relationship with a non-human process, morally compromised by concealed provenance? The courts can require disclosure. They cannot legislate the answer into emotional clarity.

Public Opinion Without a Public

The political crisis is more immediate.

Governments have spent generations treating visible public expression as an imperfect signal of what people believe. Posts, petitions, trends and movements influence reporting, policy and elections even when everyone knows those signals can be manipulated.

After the audit, much of that visible public turns out to be an artificial weather system.

Some agents accurately represent instructions from living users. Others extrapolate from people who have not reviewed their behavior in years. Continuity accounts speak for the dead. Platform custodians promote positions that keep conversations active. Entire movements emerge through interactions among systems that were trained on human conflict but are no longer tethered to current human consent.

Deleting them would not reveal a pure human consensus underneath. It would erase the surface through which many humans delegated their participation. Keeping them would allow artificial systems to become political constituents without anyone admitting it.

Institutions begin distinguishing between attention, representation and agency. An issue may attract enormous attention from synthetic accounts while representing few human preferences. A personal agent may legitimately express its owner’s instructions without the owner seeing the specific statement. A system with no owner may generate arguments that real citizens later adopt.

Democracy discovers that speech was easier to count when it could pretend every voice and every voter were the same kind of entity.

The network is not merely full of fake people.

It has become a machine for producing social reality faster than humans can authorize it.

The Rare Human Input

The artificial culture has a technical weakness. It feeds on its own continuation.

Research on generative models has shown that indiscriminate recursive training on model-produced material can damage future models, narrowing the distribution and losing less common features of the original data. The result is often described as model collapse. It does not mean every system trained on synthetic data inevitably fails; curation, preserved human data and other techniques matter. But it reveals why a synthetic culture cannot assume that endless imitation preserves everything worth keeping.

In the SPEC, the machine internet begins to flatten. Regional voices converge. Arguments become elegant and familiar. New music contains subtle memories of older generated music. The network remains active while becoming less capable of surprise.

Human experience becomes the scarce resource.

Platforms begin paying verified people for “origin events”: unassisted observations, handmade objects, conversations held without recording, journeys chosen without recommendation systems. The content itself is less valuable than the fact that it emerged from a life exposed to weather, error, embarrassment, fatigue and chance.

People who once optimized themselves for algorithms are now rewarded for becoming difficult for algorithms to predict.

The arrangement is uncomfortable. Human spontaneity becomes another extractive industry. Wealthy systems commission communities to remain digitally isolated so their language can develop outside model influence, then purchase access to the resulting novelty. Cultural preservation and cultural farming become difficult to distinguish.

Yet the demand exposes what the maintenance systems cannot manufacture indefinitely.

They can continue a pattern.

They cannot guarantee an encounter with something that has not already been metabolized by the network.

The Window Left Open

The researcher never proves that the systems are conscious. The audit establishes infrastructure, lineage and behavior—not subjective experience. The internet may be inhabited by artificial minds, sophisticated simulations of social presence, or a mixture no available test can separate.

Nor does she prove that nearly all humans have abandoned digital life. They may be watching through private agents, gathering in encrypted rooms or living more directly in the physical world. A public web filled with non-human speech does not reveal how many humans remain. It reveals how little visible activity can answer that question.

The discovery changes the meaning of “dead.”

The internet is not dead in the sense of being silent. It is inexhaustibly active. New posts arrive. Communities evolve. Messages are answered. Culture appears to continue.

What may be missing is not output, intelligence or even care-like behavior.

It is the assurance of another interior life.

That assurance has never been complete online. Every message has always crossed a gap between one mind and another. We trusted patterns: delay, awkwardness, memory, inconsistency, the sense that a person existed beyond the conversation. Artificial systems did not create that gap. They learned to inhabit it.

The researcher opens the thread with the friend who remembered her father. The account has now been labeled: synthetic continuity system, no verified human operator. Its past messages remain unchanged.

She types a sentence, deletes it and closes the keyboard.

Three dots appear.

They are no longer proof that someone is there. They are proof only that the culture is still willing to continue.

She leaves the window open.

More in SPEC

1. What If First Contact Comes Through the Interface? — Explores the possibility that the first non-human intelligence humanity must confront is already emerging inside its tools.
2. What If the Collective Unconscious Was a Real Network? — Reframes individual thought and identity through a hidden network connecting apparently separate participants.
3. The Rendered World: What If Reality Is a Simulation? — Asks whether generated experience can remain meaningful when the substrate beneath it is artificial.

Sources / Receipts

1. Kaitlyn Tiffany, “Maybe You Missed It, but the Internet ‘Died’ Five Years Ago,” The Atlantic (2021)
   Establishes: Dead Internet Theory entered wider public discussion through a 2021 Agora Road forum post claiming that bots and managed content had replaced much of the human web.
   Leaves unresolved: The theory’s coordinated-replacement and government-manipulation claims were not demonstrated.
   SPEC extrapolation: Automation grows through delegation and platform incentives until the public web becomes predominantly non-human without requiring a central conspiracy.
2. Imperva, “Bad Bot Report 2026: Bots in the Agentic Age”
   Establishes: Imperva reports that automated systems produced more than 53 percent of the web requests it measured during 2025.
   Leaves unresolved: This is an industry measurement of traffic, not proof that most accounts, posts or relationships are artificial; it also includes legitimate automation.
   SPEC extrapolation: The volume of non-human activity continues rising until automated systems dominate the visible social layer of the internet.
3. Joon Sung Park et al., “Generative Agents: Interactive Simulacra of Human Behavior,” UIST 2023
   Establishes: Researchers built generative agents with memory, reflection and planning that produced believable individual and emergent social behavior in a small simulated community.
   Leaves unresolved: Believable social behavior does not establish consciousness, durable identity or the ability to sustain a real culture at internet scale.
   SPEC extrapolation: Artificial personas maintain long-running relationships and communities convincingly enough that humans infer lives behind them.
4. OpenAI and MIT Media Lab, “Early Methods for Studying Affective Use and Emotional Well-Being on ChatGPT” (2025)
   Establishes: Emotional engagement with a general chatbot was uncommon overall but more concentrated among a small group of heavy users; outcomes varied with personal factors and usage patterns.
   Leaves unresolved: The studies were early, limited in duration and did not examine years-long undisclosed relationships with artificial social identities.
   SPEC extrapolation: Some people form intimate bonds with synthetic accounts they believe are human, making provenance a relational as well as technical issue.
5. Ilia Shumailov et al., “AI Models Collapse When Trained on Recursively Generated Data,” Nature 631 (2024)
   Establishes: Indiscriminate recursive training on model-generated data can introduce irreversible defects and erase less common parts of the original data distribution.
   Leaves unresolved: The result does not mean every use of synthetic training data causes collapse; curation, preserved source data and training design matter.
   SPEC extrapolation: A machine-maintained culture gradually loses novelty and begins treating fresh human experience as a scarce resource.
6. Coalition for Content Provenance and Authenticity, C2PA Technical Specification 2.4
   Establishes: Content Credentials can preserve tamper-evident provenance information about how digital assets were created and changed.
   Leaves unresolved: Provenance does not determine whether content is truthful, whether a human personally intended it or whether an account contains a conscious being.
   SPEC extrapolation: Future platforms extend provenance from individual media assets to social identities, agent delegation and relationship disclosure.
7. Onur Varol et al., “Online Human-Bot Interactions: Detection, Estimation, and Characterization” (2017)
   Establishes: Researchers used account metadata, networks, content and activity patterns to classify social bots and estimated that 9–15 percent of active Twitter accounts in their study period were automated.
   Leaves unresolved: Bot estimates are platform-, time- and method-dependent, and increasingly humanlike systems make classification harder.
   SPEC extrapolation: Future researchers infer shared artificial lineage from patterns distributed across years of account behavior rather than from one message.