Information Without Ignorance

How AI fails to understand despite having all the facts, and how GEO might not be helping…

A 2026 paper published by UCL's AI policy department caught my eye, not only because of its ballistic title — Time Without Death: Finitude, Social Order, and What Machines Lack — but because it put into words a question I think a lot of AI science, be it in development or application, has been gearing up to answer.

What is the difference between how we think and how an ML system thinks?

Of course, blatant anthropotheisms (if that's even a word) aside, it's pretty clear that we do, or rather must, think differently. There are likely many, many reasons for this distinction, but none has caught my attention as much as Canhui Liu's description of Finitude.

As we see the steady emergence of Agentic Ecosystems, namely Claude's multi-agent systems, the question of how they transfer information across models becomes an ever more pressing one. Although much of the current state of the largest and most powerful agents is in a bit of a voluntary black box of its own (thank you, commercial privacy lawyers), there is little to no doubt that they are designed around building networks of specialised and highly integrated sub-models.

Canhui Liu argues that, despite developing sophisticated forms of coordination, the AI lacks a constitutive finitude. One that, according to the philosophies of Stiegler, Heidegger, Hegel, as well as many others, is fundamental to the making of cumulative human culture. The agent's knowledge can, in its own mind, be copied, restored, deepened or layered upon; but never irreversibly lost.

I want to take this a step further; perhaps it isn't loss per se that is generative but the incompleteness of knowledge. I think it's here we come to the real distinction, and one that has not necessarily been tackled just yet, retrieval and understanding are not obviously the same optimisation problem.

In that sense, the question that GEO and AI management need to answer is: At what point does additional specification, while measurably increasing an entity's visibility, decrease the model's ability to distinguish its essential characteristics from its adjacent ones?

Why Agents Suffer Under Excess

One of the clearest demonstrations comes from a paper tellingly titled Lost in the Middle. Nelson Liu and colleagues found that language models presented with long contexts often fail to use the information that matters most. Counterintuitively, simply adding more surrounding content — even when the context still contains exactly one fact that answers the question — can make a model perform worse than giving it no context at all, depending on where that fact happens to sit. So what's going on?

We can move forward with this idea; maybe it's just a matter of whether you want an AI to truly understand your brand's meaning; put it where the model is most likely to notice it. However, I believe this still only captures part of the picture.

Part of the problem is that optimisation has been geared towards one side of the equation, that of visibility. It's no surprise that, for most, if not all, of digital markets' existence, the actual content and communication of what an entity or brand is come from the human hands that build it; the only goal was visibility because the communication came after. Agentic systems fold those two worlds together. Visibility is defined by the quality of communication, a measure of quality that is, at least, not entirely coherent with the measures we have traditionally used.

It suggests there is a difference between possessing information and recognising significance. As context grows, every detail becomes increasingly well specified; the model is simply provided with more information to retain.

The issue that emerges isn't so much a problem with the volume of content exceeding the model's ability to process it. If that were the case, the fix would be relatively simple: larger hardware, better processing.

The problem is that, in privileging a kind of objectivity about the entities in question, every characteristic is presented as an objective property. On what grounds, then, can the model treat one as more constitutive than another?

Agents Increasingly Think In Communication With Others

This becomes even more interesting once we stop thinking of models as isolated systems. Techniques like ReAct, which interleave reasoning with external actions such as search and retrieval, have become foundational to modern AI agents. Rather than producing an answer in a single pass, the model now coordinates between tools, memories and specialised processes, continually passing information from one stage to the next.

That seemingly technical shift changes the way knowledge itself has to exist inside the system. Information can no longer remain a rich internal representation; it has to be communicated consistently and through a largely invisible process devoid of user input, where one might be able to interject and disassociate certain pieces of information.

This is quite different from the kind of conceptual development Canhui Liu describes through his engagement with social epistemology/philosophy. Human ideas do not simply accumulate descriptions of themselves. They develop through negation. We discard explanations, forget details and allow certain interpretations to recede until something more coherent emerges. Meaning is produced as much by what disappears as by what remains.

An AI has no comparable mechanism. Its architecture is built around preserving information by making it communicable, such that every potentially relevant detail can be retrieved, summarised and passed forward. Development, therefore, occurs through increasingly layered acts of description. The concept does not become clearer because there is no editing; it is forced to reconcile a massive degree of what it perceives to be highly relevant information. Layers upon layers of retrievable characterisation will not necessarily construct a communicable image of an entity's or a brand's identity.

On the other hand, I wonder how this relates to the increasing requirement of conversational prompt networks; the AI actively seeks more information, while this seemingly can only narrow down the scope of what the user wants. Each clarifying prompt can quietly make the earlier ones irrelevant — you're resolving your own meaning as you go. But the model has no way to discern that kind of loss as productive. Unable to treat earlier turns as superseded, it just keeps carrying them forward, and perpetuates its own confusion (I should like to mention that I have yet to see any studies on this explicitly, but I understand that the nature of conversational prompting is a fast-growing study in AI policy and GEO theory)

We end up, as we so often do, in a vicious loop. The architecture favours entities that stable attributes can describe; we then define our entities by increasing their density and the prevalence of those attributes, all to garner more visibility. This works well until you realise that the agent does the job of interpreting as well and is, arguably, moving away from guiding a user towards the brand's own communication altogether. As the heart of Canhui Liu's paper describes, description is additive; meaning is selective.

Retrieval vs Understanding: What Do You Want?

So the question that we are left with is, what is it we really want? Does optimisation still only mean retrieval? Or are we trying to optimise a brand's capacity to know the brand in question?

I think the answer, outside the context of GEO, is an engineering one: an integrated system of information negation, running on multiple concurrent models of credibility, could potentially mimic the seemingly productive exclusion of information that defines the resolution of meaning — at least in topics where purely qualitative representation only gets you so far. In any case, I have little doubt that the people building these systems are already wrestling with these questions (if only we heard more about the development of this 'shared future').

Within GEO, however, we don't get to redesign the architecture. We redesign the information. For the last two years, optimisation has largely meant making ourselves more legible to machines. The assumption has been that if we can simply tell AI more about ourselves, it will understand us better.

But perhaps that's the wrong assumption.

Claude Shannon famously argued that information isn't valuable because there is more of it. It is valuable because it reduces uncertainty. Although Shannon wasn't concerned with meaning in the way I am here, his point is still instructive: information only has value insofar as it differentiates.

As brands race to produce increasingly comprehensive, objective descriptions of themselves, they may also be making themselves progressively less informative. A description that tries to tell a model everything risks making it harder for the model to recognise what actually makes the brand a 'brand'.

There's nothing worse for the goals of optimisation than a flat landscape, and there's nothing worse for an AI's ability to communicate meaning than excess information.

That matters because GEO is no longer simply responding to AI. It is increasingly participating in the informational environment AI inhabits. Every optimisation teaches models not just what a brand is, but how brands should be described. If our instinct is always to replace ambiguity with exhaustive description, we shouldn't be surprised if future systems become extraordinarily good at retrieval while finding distinction increasingly difficult to compute.

If that's the challenge, the future of optimisation isn't about writing more. It's about deciding what deserves to remain meaningful after everything else has been communicated.

As your Art teacher has always told you, the mark of quality in communication is knowing what to leave aside, so the meaning stands out.