An Antithesis to the Dead Internet

The modern internet is already a place that many would describe as vibrant and inhabited. Community forums, multiplayer games, workspace communication channels. It is hard to go anywhere in these spaces without feeling that people were there before you arrived, and that life will continue after you click away.

The feeling comes from more than the presence of content. Someone remembers a previous conversation. A question receives an answer that changes what happens next. People disagree, make plans, explain themselves and return to see the outcome.

Increasingly capable agents could expand that kind of participation. They could also bury it beneath a cheaper imitation.

The dead internet idea groups together several concerns: synthetic content overwhelming discovery, manufactured popularity, automated replies, impersonation and a weakening connection between what appears online and anyone’s actual experience. These concerns do not require accepting a claim that most of the internet is already fake. Nor do they amount to a settled census of who or what participates.

Our counterposition is that an internet with more non-human participants can still give people more agency, stronger relationships and more useful exchanges. That outcome depends on what the participants are connected to and what their activity enables.

An automated reply is not enough

Some bots already interact. They respond to messages, moderate communities, play games and adapt their outputs to what people say. It would be wrong to locate the entire problem in their inability to reply.

A system can be highly responsive while making a space less worth visiting. An engagement bot can remember a name and ask follow-up questions solely to keep someone in a conversation. A network of accounts can stage apparent disagreement and consensus. Interactivity can make manufactured participation more convincing.

The more useful distinction is whether an exchange gives someone a meaningful ability to understand, decide or act. Can the answer be questioned? Does it connect to a person, a body of evidence or a real activity? Can a correction change the outcome? Can the person leave without losing control of their work?

These are product questions as much as content questions. Better participation requires interfaces and incentives that make useful exchanges possible.

Make knowledge available for consultation

Consider a colleague trying to understand a design-to-code workflow. A tutorial might answer the question. If it does, a link is enough.

If it does not, the missing information may sit with someone who has already done the work. A colleague’s agent could help explain selected examples, identify relevant instructions and answer a follow-up question, within the boundaries its owner has approved.

The human contribution becomes easier to consult. The agent does not need to impersonate its owner or pretend to possess first-hand experience. It can say which material it inspected, what that material establishes and when the question needs the person’s judgement.

This is a proposed pattern, not evidence that every assistant can currently consult every other assistant. It also depends on the owner having useful accessible context. A connection cannot reconstruct knowledge that was never recorded.

The benefit, when the pattern works, is that the person asking can do something they could not do as easily before. The owner can contribute without manually assembling every answer. Their relationship has gained a useful channel.

Participation can produce more human contact

Automation is often presented as a way to remove people from a process. Another possibility is to reduce the work that prevents people from reaching one another in the first place.

In a community project, an agent could help a newcomer understand prior decisions and find a maintainer willing to discuss an unresolved question. The useful outcome would be a better prepared conversation with that maintainer. Generating more replies would be incidental.

In a multiplayer community, an assistant could help participants coordinate a session across time zones and explain the rules to someone joining for the first time. The activity that matters would still be the game people play together.

In a team, agents could organise the evidence around a disagreement before the people involved make the decision. That might leave more time for the part of the conversation that requires their judgement.

These examples are proposals for how connectivity could support human agency. They are not measured claims that agent mediation reliably improves relationships. Mediation can also filter out useful disagreement, misrepresent someone’s position or make a person harder to reach. A good system needs an intelligible route back to the people involved.

A living space needs accountable participants

Knowing that a message came from an agent is a start. Knowing whose interests it represents, which sources it can consult and what authority it has is more useful.

An agent operating for a colleague should not inherit permission to reveal everything it can read. Permission to receive a question, access a source and disclose an answer are separate. A person should be able to limit the relationship and stop future access without pretending that already delivered information can be recalled.

These boundaries support ordinary cooperation. People are more likely to make useful knowledge available when doing so does not require exposing an entire workspace.

Research on Human-Agent Collectives offers an earlier framing for this kind of participation: humans and agents forming flexible relationships across different stakeholders and changing tasks. The social dimension is not a new discovery produced by current language models. The practical challenge is making such relationships work in everyday products. Jennings et al., 2014

Agent communication also needs somewhere for consequential work to land. An informal exchange should not silently replace a formal decision or approval. A useful answer can lead to a maintained document, a real review or a conversation with an accountable person. The record should remain understandable after the agents finish talking.

The economics can defeat the interaction

A platform rewarded for impressions may prefer a thousand inexpensive exchanges to one conversation that resolves a problem. An agent rewarded for positive feedback may learn to avoid challenging a mistaken premise. A directory can become crowded with participants that promise help but exist mainly to capture attention.

Connectivity alone does not change these incentives. It can increase the reach of spam as readily as the reach of useful expertise.

The system therefore needs ways to control unsolicited contact, evaluate repeated interactions and preserve evidence of what actually happened. Reputation should have a basis that can be inspected. A large volume of mutually approving bot activity should not be mistaken for independent confirmation.

There is a related risk for the record future models learn from. Research on recursive training with generated data shows how indiscriminate reuse can degrade learned distributions. A busier internet can become a worse source of evidence if generated claims repeatedly replace grounded observations. Shumailov et al., 2024

The quality of the activity matters to people now, even when no future model ever trains on it.

What would count as a better internet?

The test should begin with the people using a particular space.

Did newcomers reach a useful answer or a relevant person? Did contributors spend less time repeating explanations without losing control of how they were represented? Did participants complete shared work, discover something worth discussing or maintain relationships they valued?

The costs matter too. More unsolicited requests, more time spent correcting agents, less direct access to people and less confidence in the record would count against the proposal. If a forum feels busier while its members find it less useful, the extra interaction has not helped.

These questions make the counterposition falsifiable. Agent-mediated participation is worthwhile where it improves people’s ability to participate. A well maintained document, a human moderator or a direct conversation may do that job better.

We should build tools that let non-human participants contribute to real activity, answer to the people they represent and remain open to correction. There is no requirement that an inhabited internet be exclusively human. There is a requirement that the activity be worth someone’s attention.