What organizations perceive of their representation by AI systems

Introduction — a new question

For decades, organizations have built their presence: websites, publications, registers, content. The information that concerns them now exists across a multiplicity of environments — their own, and those they do not control.

Conversational AI systems now produce answers from these environments. When a system is asked about an organization, it does not access the organization itself: it assembles a representation from the information available to it, then returns it. What the system returns is not the organization — it is a reconstruction, faithful or partial.

A new question then appears: how are organizations represented through these systems — and what do they perceive of it?

This note documents the second part of that question: perception. It does not measure what the systems do; it reports what organizations understand, expect and fear about them today.

1. The emergence of a field of observation

OMINA was not born from a market study. It was born from a trajectory of technical exploration. After completing a degree in data science, its founder was looking to build a project that would demonstrate what she was technically capable of — and first asked herself whether she could build a language-model-related system of her own. As that exploration began, a more fundamental question than the initial project took hold: which informational paths do artificial intelligence systems take to construct their answers?

That question led to experiments, and the experiments to observations. Some took the form of anomalies, and some anomalies became deeper questions. One of them deserves to be stated for what it says about the nature of the field: a response can be perfectly coherent, precise and convincing while resting on a fragile or incorrect reconstruction. Nothing in the apparent quality of the response necessarily makes this visible.

An answer that appears correct is therefore not enough to establish that the system has reconstructed the organization correctly.

The initial question then shifted. It was no longer only a matter of knowing which paths lead to an answer, but of determining what can actually be observed in the reconstructions produced, what those observations legitimately establish, and how they should be interpreted. Each of these steps raised new questions that the analytical instrument, progressively refined, had to be able to address. From that movement came OMINA’s working framework — observation, interpretation, governance.

2. Why OMINA met with organizations

After observing the phenomenon technically, OMINA sought to understand how organizations themselves perceived this emerging situation.

Conversations were held with executives of organizations active in Switzerland, across varied contexts of sector and size. They followed a common set of questions, while leaving participants free to develop their answers. These exchanges distinguished two moments: what participants expressed before OMINA’s observation framework was presented to them, and their reactions once that framework had been laid out. The observations that follow are reported in aggregated form; none is attributable to any organization.

These conversations constitute neither a representative sample nor a validation: they confront a technically observed phenomenon with the way it is perceived by those it concerns.

3. What the conversations reveal

The question of sources.

Observed — One question returns insistently, often before it is even asked: where does a system draw what it says about an organization from? Through which paths? From which sources? The most advanced participants formulate the same question with greater precision; none of the participants had an answer.

Interpreted — The question of sources is the natural entry point into the subject: before wanting to act, organizations want to understand where the representations that concern them come from.

Publishing is not being understood.

Observed — Several participants base their confidence on their own publishing efforts: a clear website, a maintained presence, a history of visibility. That confidence rarely came with any verification of what a system actually returns about their organization.

Interpreted — When an organization has spent years building a clear website, publishing information and maintaining its online presence, it is natural to assume that this information will be taken up correctly by AI systems. But information being accessible does not mean it will necessarily be retrieved, connected and reconstructed in the same way by those systems. Without observing what a system actually returns, an organization cannot infer from the quality of what it publishes what has been reconstructed from it. Nothing in these conversations allowed anyone to confirm or refute that confidence.

The correction instinct.

Observed — Faced with the hypothesis of an inaccurate representation, the same first reaction returns: correct the system, “teach it”, report the error. More rarely, the correction is spontaneously directed at the information sources rather than at the system itself.

Interpreted — The reaction is understandable: the conversational interface gives the impression of a dialogue with a system that could simply be told about an error. But correcting an answer is not the same thing as understanding the information and relationships from which that answer was reconstructed. Once that distinction is made explicit, the question changes: it is no longer only about correcting the system, but about understanding what led it to produce that representation.

The absence of a direct relationship.

Observed — None of the organizations encountered has a way of observing what the systems return about it, nor a channel through which to respond. Verifications are non-existent, occasional or improvised; discoveries are sometimes made by accident, in the course of an exchange with a third party.

Interpreted — This absence of a loop — no view, no dialogue, no feedback — connects a large part of the perceptions gathered: uncertainty about sources, confidence based on what has been published, and the correction instinct.

Identity confusion, a concrete concern.

Observed — The possibility that a system might associate an organization with information about another entity — a namesake, a neighbouring entity, a better-documented one — appeared in the conversations in several forms: lived experience, stated fear, constructed example, risk assessed and then set aside.

Interpreted — Of all the phenomena in this field, identity confusion is the most immediately intelligible: it can be told in one sentence and requires no prior concept.

Being visible, being understood.

Being visible means that a system has enough elements to bring up an organization: its name is known, its website is accessible, information about it circulates, it can be mentioned in answers. Being understood raises a different question: when the system speaks about the organization, what does it actually reconstruct? Does it describe its activity correctly? Are its offers and services rendered faithfully — and connected to one another as they are in reality? Does it know whom they are intended for, what distinguishes the organization’s positioning, and which information belongs to it rather than to a neighbouring entity or one with a similar name?

An organization can be visible to a system and still be poorly understood by it.

The confusion between the two is understandable. Organizations have learned, over many years, to make their information available and to make themselves findable online. It is natural to conclude that a solid presence — abundant, accessible, well-maintained information — guarantees correct understanding. But the availability of information and its reconstruction by a system are two distinct phenomena: the second does not follow mechanically from the first.

This distinction did not come from the interviews. It was introduced during the conversations by OMINA’s observation framework. What was observed, rather, is how it was received: understood quickly, reformulated by participants in their own terms, it gave them a more precise way of describing concerns they had previously expressed more diffusely. A distinction that an executive makes their own within minutes says something about the phenomenon it names.

Being present in information environments is therefore not enough to establish how an organization will be reconstructed from those environments. That gap is what makes observation necessary.

4. What these observations do not establish

These conversations reveal perceptions and questions. They do not establish:

5. The questions this opens

The gap between what organizations perceive and what the systems actually do is precisely the field OMINA observes. The following questions outline its programme:

These questions no longer belong to conversation alone: they call for instrumented observation. They guide the research OMINA conducts and, when they can be translated rigorously enough into observable phenomena, contribute to the evolution of its analytical instrument.

OMINA — Observation note, 2026. The conversations mentioned were conducted in Switzerland in 2026 and are reported in aggregated form; no observation is attributable to any organization.