Umberto Eco – AI Prompting – Semiomatics

12 Giugno 2026

Written By Sebastiano Paolo Lampignano
Innovation & People Management Director

Abstract

This paper revisits selected studies by Umberto Eco to explore their relevance in understanding generative AI, ranging from his definitions of different types of Encyclopedias to the concept of textual cooperation, including all the elements that emerge in between – text, context, enunciation, model reader, and labyrinthine knowledge. These are interpreted, within Generative Artificial Intelligence systems, as regulative ideas capable of functional – albeit partial – forms of application.

However, this framework raises a fundamental question: are we still within the domain of semiotics, or are we facing a new epistemological threshold?

The proposal advanced here is that the current configuration of GenAI (Generative Artificial Intelligence) systems can no longer be analysed exclusively through the tools of generative and/or interpretive semiotics but instead requires an extension of the field of study – emerging in recent years – referred to as Semiomatics.

Encyclopedia as a Dynamic Architecture

Umberto Eco conceives the Encyclopedia not as a static repository of truths, but as a dynamic network of interpretants, including the true, the false, the imaginary, the controversial, and what is culturally sedimented.

In A Theory of Semiotics (Italian version: Trattato di semiotica generale, 1975), and especially in From the Tree to the Labyrinth (English version 2014, Italian version: Dall’albero al labirinto, 2007), the Maximal Encyclopedia emerges as:

potentially infinite;
not fully accessible;
accessible only through local reductions of complexity;
activated each time through situated interpretive practices

When Umberto Eco made his ideas on the Encyclopedia public, he could not have foreseen that the gap between his theoretical reflections and technological developments would gradually narrow. Over time, through ongoing reflection and refinement of his ideas, this line of inquiry culminates in 2007 (Italian version), with the publication of From the Tree to the Labyrinth. In this work, Eco formulates the statements cited above and, although he refers in some passages to the Internet, he could not have anticipated that, nearly twenty years later, Artificial Intelligence – also explicitly mentioned (Eco, Italian version, 2007, para. 1.6, p. 67; English version, 2014, pp. 16, 56) – would take a form so closely aligned with his own intuitions.

It is, of course, stated in several passages that the Maximal Encyclopedia is a regulative idea and does not possess content that is immediately accessible, except through processes of complexity reduction.

When the vast and undifferentiated world of the web is identified as one of the possible domains in which this type of Encyclopedia may reside, it is clearly understood that it consists of archives distributed across the globe, lacking any overall logical or physical integration, existing for different purposes, in different locations, languages, and formats, and produced with diverse objectives. Thus, the notion of content that is present yet not immediately accessible once again comes to the fore. At a certain point in his work, Umberto Eco hypothesized a possible proximity between the Maximal Encyclopedia and the World Wide Web, although he argued that the latter would have to be “a sort of World Wide Web far richer that the one to which we have access through the Internet” (Eco, Italian Version 2007, p.81, English Version 2014, p. 70). Nearly twenty years have passed since then, and today’s web is indeed a far richer world – so rich that it can only be meaningfully queried in partial and localized ways.

This vast body of information, embedded in a continuous process of expansion, enables the existence of textual units – of varying size – both as autonomous entities and in the form of a semiotic hypertext. That is, it can activate processes of semiosis, whereby a sign refers to and is understood through other signs, which in turn are understood by referring to further signs. However, to access this type of content – which exists in a state of almost chaotic invisibility – it is necessary to perform reductions in search complexity.

More specifically, as anticipated, the Maximal Encyclopedia proposed by Umberto Eco is a regulative idea, and this model of knowledge is not fully accessible in practice. Accordingly, he introduces two further forms of Encyclopedia – the Medium and the Specialized – as usable reductions available to human agents. The Medium Encyclopedia represents a form of reduction that contains information accessible both to specialists and to non–specialist users, whereas the Specialized Encyclopedia constitutes a reduction primarily useful for experts. The one does not exclude the other.

These reductions are not finite in number and can be composed according to local needs, where “local” refers to the specific context in which they are activated. What, then, is the instrument that enables both the reduction and the local use of such a vast, heterogeneous, and distributed body of content? Within the context of Generative Artificial Intelligence, two main modalities can be envisaged.

The first, oriented toward specific and professional domains, is represented by the use of selective datasets for training the algorithm. In this case, a reduced version of the overall competence of the GenAI system is obtained. The second, applicable to public general–purpose systems such as ChatGPT, Claude, or Gemini, etc., is represented by the use of prompting, with its various techniques, its constraints, and yet its expressive freedom. It is this second modality that will be the focus of the remainder of this study.

Through prompting, it is possible to request that the AI (Artificial Intelligence) system select portions of knowledge – of any desired scale – and subsequently amplify such portions or, conversely, activate mechanisms of narcotization (attenuation) of certain elements in favor of others (Eco: Italian version 1979, English version 1984). It is worth noting that, under certain conditions and after a sequence of interactions with the algorithm, prompting effectively entails the application of a principle of forgetfulness – that is, the deliberate restriction of the set of collected information in order to foreground what is of interest.

Medium and Specialized encyclopedic elements are activated and combined according to the type of prompt provided as input to the algorithm. As soon as the text composing the prompt is entered, a process of textual cooperation begins between the user and the algorithm. It is worth noting – although this may generate some unease – that the algorithm, in turn, activates textual cooperation with the writer through what is written.

[…]

In this sense, prompting can be interpreted as the operational locus in which distributed agency emerges: the user does not merely query the system, nor does the system passively respond, but both participate in the co–construction of meaning through the activation of partial and situated encyclopedic pathways.

At this point, a fundamental question arises: who is cooperating with whom?

Textual Cooperation or Semiotic Simulation?

A more critical juncture in the relationship between the studies of Umberto Eco and Generative Artificial Intelligence concerns the notion of textual cooperation between user and algorithm. From Eco’s perspective, textual cooperation entails:

a Model Reader;
an interpretive strategy;
a shared encyclopedic horizon;
a system of signs intelligible to the participating agents.

In the case of GenAI, however, this cooperation appears structurally asymmetrical.

The algorithm:

does not possess a Self;
lacks embodied and lived experience;
does not produce interpretants, but simulates sign–based statistical correlations;
does not interpret, but computes plausibility, albeit through refined strategies of linguistic–semantic chaining

Certain limitations of traditional semiotics emerge when addressing this form of cooperation. From this perspective, GenAI systems do not participate in semiosis, but rather simulate its observable effects. It is, in fact, as if a genuine form of cooperation were taking place, yet several problematic aspects arise – such as the definition of a Model Reader, which, in this context, is effectively constructed by the Empirical Reader through the prompt. The algorithm, in its function as Author, adapts the complexity of its text to the request and does not project or anticipate its Model Reader; rather, it accommodates one that has already been implicitly configured. This asymmetrical cooperation – between a living body and a digital system – cannot always be adequately interpreted through the lenses of traditional semiotics, which was developed to account for the organization of meaning in human beings, endowed with both material (the body) and immaterial (the Self) dimensions. If we were to further increase the complexity of cooperative interaction, one might imagine two automata communicating through human languages. The principal dimensions of knowledge – such as belief, wisdom, and experience – would take on forms and contents fundamentally different from those associated with human cognition. Accordingly, these terms themselves may require profound reconsideration considering emerging technological developments.

How, then, could we conceptualize the organization of Knowledge and Meaning in such a scenario? One might even adopt an ironic stance, suggesting that GenAI knows that it does not know, yet behaves as if it did; or that it knows without truly knowing; does not know that it knows, yet knows much of what we ask it to know. But what kind of knowledge is this? How can we study the awareness of knowledge if the agent that expresses it lacks a Self–capable of perceiving it? It becomes clear how elusive and slippery this issue is.

One is reminded of the immense work of Maurice Merleau–Ponty (1945), in Phenomenology of Perception, which places the body at the center as the original locus of experience, overcoming the separation between subject and object. Perception is neither a mere sensory datum nor an act of consciousness, but a living intertwining of world and body. Artificial Intelligence has no body – not yet; it does not know the sensation of physical pain, nor does it possess any awareness of finitude or the end of life. For this reason, its abstract mode of thinking may appear similar to human thought, yet it remains profoundly different. It represents a mode of existence that is other and elsewhere. Certainly, this is a theme that will require extensive discussion in the future, particularly in light of the fact that the cognitive difference between human beings and Generative AI is not merely one of degree, but of nature – assuming that the very notion of a ‘cognitive difference’ is indeed applicable.

Certainly, it cannot be overlooked that much of semiotics, given the nature of its origins, is intrinsically tied to the human being. Its concepts and theoretical frameworks are grounded in human cognitive and enunciative strategies. What narrative programs, then, or what kind of understanding can a generative algorithm truly convey? When GenAI employs the pronoun “I” in its responses, what kind of “I” is being invoked?[1] What is its actual lived experience? And can such a condition account for a particular mode of being that is non–human – for instance, an automaton operating within a regime of radical abstraction? Indeed, it has no body (not yet), does not feel pain (not yet), and does not bear existential responsibilities (not yet). One might therefore envisage, for the future, the development of a new vocabulary capable of resemanticizing process–oriented terms such as thinking, reasoning, feeling, perceiving, and believing, so as to include agents that are not human yet, for the time being, behave as if they were. Alternatively, one might propose the introduction of an entirely new terminology, more consistent with the evolving scenarios of semiomatics.

From Semiotics to Semiomatics

The intervention of non–human entities and elements in the generation of meaning calls for a constructive reconsideration of semiotics, from which emerges the proposed field of study referred to as Semiomatics. It is acknowledged that many theories – and the terms used to articulate them – were developed in contexts in which the human being was both subject and object of inquiry. As a result, they often prove difficult to adapt to new realities. Terms such as understanding, thinking, reasoning, learning, and perceiving, for instance, may become problematic when applied beyond the human domain. In hybrid contexts – or even in purely artificial ones – can semiotics alone still serve as the framework capable of accompanying these new cognitive developments?

In some studies, it has been proposed to apply semiotics to Artificial Intelligence as if it were a “particular case of falsification” (Leone, 2023). Certainly, this may represent an interesting possibility that deserves further exploration. However, it may be useful – drawing on the profound work of Henri Bergson – to consider the current stage as merely one among many that have preceded it and will follow; otherwise, the risk that “we have loaded movement with immobility” (Bergson, 1938) would be considerable. It is therefore necessary to engage coherently with this ongoing transformation, within which the underlying objective – at times deliberately concealed – is not simply to simulate human intelligence, but to surpass it. This is a thought that undoubtedly provokes unease yet remains highly plausible.

Within this complex context, the term Semiomatics may be understood as an interdisciplinary field of study aimed at analysing:

the automated production of signs;
the circulation of meaning in hybrid scenarios (human and non–human), if not even in the absence of necessarily human subjects;
a new semiotics of knowledge;
the encyclopedic manipulation carried out by statistical machines;
the use of signs as operations rather than as interpretations.

It is evident that underlying some of the considerations presented is a fundamental question: what, then, is the difference between Semiomatics and Computational Semiotics? One might offer a concise answer by stating that Semiomatics, beyond the electronic processing of signs, aims to revise and/or identify semiotic theories and methods in which the human being is no longer the dominant subject, but rather one – not necessarily always present – among the actors involved. In this sense, Semiomatics could not merely extend semiotics into the computational domain, but would instead reconfigure its anthropological foundation, shifting the focus from human–centered semiosis to distributed and potentially non–subjective processes of signification.

It emerges at a specific juncture within Umberto Eco’s semiotics: where the sign continues to function – or appears to do so – while the human interpreter disappears. This new domain of inquiry must therefore question not only how signs operate, but also what kind of world they produce when detached from human embodiment. Drawing again on Eco’s work, it could be understood either as an extension of general semiotics or as one among the specific semiotics (Eco, 1984b). Future research will undoubtedly provide Semiomatics with its proper theoretical positioning.

[1] For a very interesting discussion of the theories and practices of enunciation, see Claudio Paolucci’s book: Persona, published by Bompiani, 2020.