What AI Contributes to the Internet and What Makes It a Peer

Vlatko

Oct 28, 2026

Share

Yesterday I was discussing with some peers from the AI community whether LLMs should be considered full members of the internet, equivalent contributors to a network that, until recently, was largely made up of people speaking to other people. I found myself resisting that idea, although the more I tried to explain why, the clearer it became that we were using the word “contribution” to mean several different things.

I have no difficulty accepting that a model contributes when it explains a paper, proposes a useful patch, or connects ideas from fields whose practitioners rarely speak to one another. Making knowledge accessible changes what people can do with it, and anyone who has spent an afternoon trying to understand someone else’s code knows that a good explanation can be a substantial contribution in its own right. What I am less comfortable with is the assumption that, because a system can participate so effectively in this exchange, its relationship to the production of knowledge is already equivalent to ours.

The distinction I was trying to make concerns what is actually being added to the network. Helping us navigate what we know, finding consequences we had overlooked, and discovering something that changes our understanding are related activities, but they deserve to be examined separately before we bundle them together as evidence of equivalent participation. The analogy that came to mind during the discussion was the difference between laying bricks and applying plaster.

When you build a house, the plaster does important work: it protects the structure, makes the rooms usable, and turns the rough result of construction into somewhere people might want to live. There is skill involved, and dismissing that work because it comes after the bricklaying would be foolish. Nevertheless, plaster depends on a wall already being there, which is the relationship I was reaching for when I described how language models work with the accumulated knowledge of the internet.

Much of their usefulness, as I see it, comes from making that inheritance easier to move through. They fill gaps in explanations, translate between different technical vocabularies, and bring scattered pieces of documentation together in a form that somebody can finally use. If a model helps an engineer understand an unfamiliar API or makes a difficult idea accessible to someone outside its field, the benefit does not become less real because the underlying knowledge existed before the conversation.

My concern is that this fluency makes it difficult to distinguish between improving our access to a structure and extending the structure itself. Once the explanation is sufficiently convincing, we can begin to attribute more to the system than the result actually establishes, especially when we have no practical way of tracing where its claims came from.

There is an obvious limit to my analogy, however, because connecting existing ideas is sometimes exactly how a new brick gets laid. A proof can use familiar techniques and establish something nobody previously knew, just as an algorithm can be original without inventing a new branch of mathematics. Since humans also build from inherited materials, the meaningful question is whether the resulting contribution adds something that survives scrutiny, rather than whether its ingredients have a history.

During the conversation, I described human originality as a long, thin tail of rare contributions that move the boundary of collective knowledge. I still find that a useful image, because most of what any of us says is familiar and most of what we build is incremental, while an occasional observation changes how a whole community understands a problem. But it remains an image: we do not have an established probability distribution of ideas that assigns humans one kind of tail and machines a cutoff beyond which discovery becomes impossible.

The same care is needed with the central limit theorem, which concerns the behavior of appropriately normalized sums or averages under particular assumptions. It does not establish that a language model can only reproduce its training examples, or that generated ideas must remain clustered around some fixed semantic average. Invoking it would give my intuition an appearance of mathematical certainty that it has not earned. [1]

What I think deserves attention is how training shapes what a system finds plausible, without confusing that influence with an absolute boundary on what it can produce. A learned distribution is more than a catalogue of examples, and learning useful structure can make combinations and solutions possible that were never explicitly present in the training data. There is also a difference between producing an answer that never appeared in that data and generalizing reliably to a different kind of problem, even though both achievements can be described rather loosely as going “out of distribution.”

This distinction matters because novelty alone tells us very little about whether knowledge has increased. An unfamiliar sentence may be nonsense, while a relatively conventional argument may establish a result nobody had previously proved. When I talk about extending the network’s knowledge, I mean a contribution that is meaningfully new and comes with a reason to believe it, which is a harder requirement than merely producing something unexpected.

Any argument that machines can only finish existing walls has to address the fact that AI systems have already participated in discovery. FunSearch is a particularly relevant example because it paired a pretrained language model with a systematic evaluator and an evolutionary search procedure, producing new constructions for the cap set problem and improved heuristics for online bin packing. Those results count as contributions to knowledge, and describing them as recombination does not make their novelty disappear. [2]

What interests me is the arrangement that made those contributions possible. The model proposed programs, the evaluator tested them, and successful candidates informed subsequent rounds, so there was a process for separating promising ideas from plausible failures. The model’s weights remained frozen during the search, which also means that continual updates to internal weights cannot be a prerequisite for every form of machine-assisted discovery. A system can begin with a human inheritance and still help us reach a result that goes beyond it, much as people do when they work from what previous generations have learned.

For me, this makes the question more concrete. A discovery process involving a model, an evaluator, a search procedure, and researchers has specific ways of establishing novelty and correctness, and understanding how those parts work together tells us more than a general argument about whether silicon can be creative. It also gives us a reason to distinguish the achievement of that whole process from a model producing convincing text in a setting where nothing comparable checks what it says.

In some domains, we can construct an evaluator that gives us useful feedback: a program executes, a proposed construction satisfies the stated conditions, or a candidate performs better on a defined test. Even then, the evaluator may capture only part of what matters, but at least there is a procedure we can inspect. The situation becomes considerably harder when the contribution is a claim about an organization, an explanation of why a project failed, or an interpretation of a social phenomenon for which we do not have an agreed test.

A persuasive answer can arrive well before anyone has a reliable way to assess it, and this is where I think the internet community faces a serious problem. Imagine that a model proposes an unusual explanation for a production outage, several generated articles repeat that explanation, and a later summary presents their agreement as an emerging consensus. We might be looking at a growing collection of apparently independent accounts while still having only one underlying observation, or no observation that supports the explanation at all.

The difficulty in that example is that repetition can resemble corroboration when the dependencies between sources are hidden. Ten accounts restating an unsupported claim do not provide ten independent reasons to believe it, however carefully written each account may be. Humans have always been capable of producing this illusion, but making another polished version much easier to generate than the first claim is to check could make the imbalance worse.

I worry about a network in which activity grows faster than understanding, because the signals we ordinarily use to judge whether something has been examined become increasingly easy to reproduce. For an engineering community, a practical response would begin with preserving the difference between a claim that has circulated widely and one that has been tested independently, along with enough of its history for another person to work out which kind they are reading.

The phrase “full member of the internet” seems to bundle together the ability to make original contributions and the much broader question of what it means to treat a participant as a peer. On the first point, I think a valid contribution deserves to be judged on its merits, regardless of the substrate that helped produce it. The evidence from systems such as FunSearch is a reason to take machine participation in discovery seriously, even while we remain specific about what was accomplished and how.

Peerhood involves a relationship that develops across time. In a technical community, somebody can challenge your explanation, ask what you tried, or point out that the result does not support your conclusion; how you respond then becomes part of how other people judge your next contribution. We learn whose judgment to trust, in which domains, and for what reasons, partly because a person’s history has consequences for their future behavior. People often fall short of that standard, but the expectation still helps explain what we mean when we call somebody a peer.

This connects to my earlier thinking about persistence, where the part that interests me is experience durably changing a learner’s capabilities, representations, and strategies for approaching the next unfamiliar problem. Such a change might involve weights or other learned internal state, and it asks more of a system than retrieving a Markdown file that describes yesterday’s interaction. External memory can support useful continuity, but keeping a record that an approach failed is a different achievement from developing a capability through understanding that failure.

Persistent learning would make the idea of an AI peer more substantial to me because its history could shape the participant itself, as well as the reputation maintained by those around it. That would leave the question of consciousness unresolved and would offer no guarantee of good judgment; it would, however, create a way for competence and dysfunction alike to accumulate through the system’s own development.

My original instinct was that organic intelligence supplies something fundamentally different to this process: the rare departure that opens a direction nobody previously knew how to pursue. I still think there is a substantial question about the role of human experience in discovery, because we encounter consequences, form commitments, and develop through lives that continue between the questions we are asked. Our relationship to an idea includes what happened when we acted on it and what that experience changed in us.

But the claim that only organic intelligence can create knowledge goes considerably further than that intuition, and I do not have a mathematical proof that would justify it. Neither an appeal to quantum consciousness nor an appeal to statistics supplies the missing explanation, so