Abstract

Social media challenged the ideal curation of hypertext gardens with the messy communal space of fluid, personalised context and recommender system framing of content. Agentic LLMs take this to the extreme consequence, breaking boundaries between authorship, automation and the structural integrity of media, producing ephemeral traversals that are also new content brought to life for users. In this new era, the Hypertext community needs a new foundation that preserves its core vision into a new technological paradigm that has the same scale as the introduction of PCs and significance for academic and practitioner communities. We set the terms of this challenge through a theoretical investigation that provides an experimentally validated theory for Human-AI co-navigation of media spaces framed as landscapes, supporting both practitioners and the journey as sightseeing experiences across landmarks of meaning.

1 Introduction

The introduction of Agentic Large Language Models (AI) is not just an addition but a structural change to the core of the Web and digital systems. Despite disruptions and experimentation with uncertain results [5], the drive to introduce AI comes with the high promises and expectations of expanding knowledge accessibility and knowledge generation. AI could be a “dream machine” [23], connecting anything and enabling a seamless traversal of knowledge and media for anyone, crossing linguistic, cultural, cognitive, and social boundaries. AI has the potential to amplify and multiply human ingenuity, transforming ideas into services and products, knowledge into real impact.

Those promises and expectations are still unmet. Inconsistent quality of outputs is seen as the gulf that ought to be crossed to reach such dream, and it has being addressed so far mostly through guardrails, narrowing AI wanderings for predictable outcomes that mimic classic computing procedural support. The core of the problem is the paradigm misunderstanding seeking to reduce AI into a derivation of PC software. On the one hand, the effort in narrowing AI behaviour results into the underexploitation of its flexible capabilities. On the other hand, the transfer of classic computing deterministic properties sets expectations to failure, framing AI fundamental characteristics as a bug to be corrected rather that valuable features that complement existing systems.

Any productions of unfettered AI explorations is always legit, in the sense they represent a plausible path across knowledge. This comes with a AI-specific paradox that pushes to users the burden of accuracy, veracity and grounding of outputs, identifying what is actually possible, useful or true from AI free-flow expressions. This is an inversion of roles, the AI is not the objective, rigid frame that filters through knowledge but a subjective-like serendipity machine. The paradox manifests in the impossible premise of the counterpart human as omniscient truth checker.

The cornerstone of any solution must start with finding the balance between the human-machine roles respecting their respective capabilities. This cannot be achieved by forcing back classic roles of master/executor but as a negotiated positioning into knowledge fields to be explored. Differently from describing the process or characteristics of results, negotiating positions involves a fundamentally different cognitive activity, defining the underlying lenses, interpretations and values of the human-AI operational field.

AI is an epoch-defining challenge for the Hypertext community that built the foundation of technology for augmenting human cognition since its beginning. In the last decades, a major focus of the community has been web technologies and web content, arguably the most successful and flawed expression of Hypertext vision. The Web, as a research object for the community, was seen as entangled systems and artifacts loosely connected by common infrastructures and standards. Now, the AI remediated Web [2] (AI Web) bundles infrastructure, content and media into one by making the new Web an artefact as a whole, not just a container of artefacts. In essence, AIs are interactive fields, shaped by human knowledge (lenses, interpretations and values), that should be negotiated in producing new knowledge. In this view, the AI era is a call for action for the HT community to take on the conceptual and experimental effort required to guide the positive, fruitful deployment of AI potential.

In this paper, we are reframing this urgent issue related to human-AI cooperation in the Agentic LLM remediated Web and Hypertext systems by proposing new metaphors grounded on the HT history and building on them case studies aimed at showcasing and stressing this new conceptual foundation.

2 AI & Hypertext

To address the rebalancing of human-AI roles in navigating digital content, we reframe AI as Hypertext Engines and LLMs (core models of AI) as hypertexts. This allows us to leverage established analogies in the field as a starting point for our approach.

2.1 AI as Hypertext engines

Rather than treating generative AI as an artificial mind, a commonplace approach, we describe it more precisely as a hypertext machine: a technical dispositif that works on cultural traces and produces traversals across them. From this perspective, the model does not "understand" the world it references, nor does it anchor its outputs in lived experience; it operates instead on signs, recombining and recontextualising linguistic forms according to learned statistical regularities [27].

What appears as reasoning can thus be framed as the construction of coherent paths through an implicit, vast, latent network of intertextual relations distilled from prior texts. In this sense the “chain of thought” [34] is not an inner mental process, but the visible traces of navigation: step by step, the system decides which textual “neighborhood” to enter next, while maintaining coherence, genre expectations, and pragmatic fit [28].

This network is not organised as explicit entries or stable concepts, like a traditional reference work. It is distributed as a web of tendencies: recurrent associations among forms, topics, styles, and argumentative moves. Generation becomes a traversal that constantly proposes candidate links and discards others, producing an output that reads like a deliberate argument even though, mechanically, it is a controlled reconfiguration of textual precedents.

Seen from a hypertext perspective, several familiar phenomena become easier to discuss without anthropomorphic shortcuts. “Hallucinations” are not memory failures so much as misaligned links: the system builds a locally plausible path that does not satisfy the user's referential anchors [18]. “Creativity” is not invention ex nihilo, but the ability to open unexpected connections between distant regions of the implicit network, producing routes that feel new yet remain legible to human interpreters [27]. “Instruction following” is, in turn, the shaping of the traversal space: system prompts, guardrails, and task formulations act as higher-order links that limit which regions can be entered, in what sequence, and with what rhetorical posture [25]. What counts as a good traversal is never purely technical: it depends on filters that are authorial, stylistic, and interpretative, and can be modelled as dynamic constraints on how textual material is selected and composed [36].

2.2 Grounding AI

AI's unpredictability and unreliability—marked by shallow outputs, poor source filtering, hallucinated facts and a “sycophantic” attitude [24]—often produces well-articulated but off-point responses, results of free-flow explorations of plausible associations. Yet, these flaws are not deal-breakers; humans are naturally equipped to navigate error, fuzziness, and deception. Just as we manage social frictions, we adapt by applying familiar negotiation strategies and heuristics to clarify goals and enforce AI accuracy.

Mutual understanding among humans, even in difficult social interactions, occurs because and within a common semiosphere [21], a historical and cultural system of presuppositions, saliences, and implicit inferences that orient our interpretation of the world [12], as well as the intentions of others. Artificial intelligence, however sophisticated, does not access such sign models, nor does it participate in the filtering function that selects, omits, structures, and endows with meaning what is said or left unsaid. Intersubjective filters order meanings through inherited habits and sedimented competences in navigating dynamic relations rather than calculation [14, 15]. Lacking this repertoire because existing outside of the temporality of culture, AI can only simulate syntactic correlations, without grasping the semantic depth or participating in the negotiation of meaning essential to bilateral communication [13, 26]

Furthermore, human heuristics are physiological and part of our everyday lived experience. We understand concepts like death, hunger, or love because they have meaning outside of text, grounded in our biological and anthropological reality. Lacking this embodied and empathetic dimension, unable to experience the frustrations of a physical body or the intricacies of what is at stakes in social dynamics, AI remains an external observer, treating meaning as a purely syntactic feature of linguistic interconnections.

Current attempts to provide context and grounding AI (seen in prompt libraries, skills-building guidelines of various professional domains) provide the machine with objects to handle in the form of a sum of instructions such as roles, task descriptions, and some operational constraints. Yet, these coordinates remain mere computational conditions applied to data and linguistic objects. AI might follow a role description, yet it lacks the ability to calibrate the subtle dynamics between intention and action. It remains a simulator of syntactic correlations, unable to grasp the background of presuppositions that make them pertinent in a shared world, nor the pragmatic implicatures that emerge from the situated use of language [16, 26, 31].

To bridge the gap between human context and AI simulation for enabling better human-AI cooperation and balancing, we need grounding techniques that transcend mere instruction-following. Since AI lacks the cultural repertoire and embodied dimension to align intention with action, we need new grounding approaches for enabling AI to navigate the "dynamic field" of human expectations and shared values rather than just processing syntactic data.

3 From Gardens to Landscapes

3.1 On Landscapes

The concept of “landscape” is frequently employed to describe digital spaces, yet often from a restrictive perspective. Digital landscapes are commonly invoked as integrated ecosystems of tools, platforms, and infrastructures that merge virtual and physical realities [4, 35]. These environments are defined by their pervasiveness—spanning from wearables and satellites to blockchains—and are increasingly mediated by generative AI systems.

In contrast, semantic landscapes serve as metaphorical descriptions for the aggregation of data into clusters of explicit meanings [1].They are termed landscapes due to the topological nature of their semantic geometries emerging from formal logic [33]. Within this framework, the Semantic Web functions as a global network of interlinked, annotated resources that dynamically reconfigure through new connections. For both humans and machines, exploring these topological arrangements, such as nodes and their relational links, is an act of discovery into what exists.

However, these technical definitions often overlook the constitutive quality of a real-world landscape as a lived space rather than a mere repository. Drawing from visual arts, urban studies, and the social sciences, a landscape is defined as a complex entanglement, a “palimpsest” [3, 7, 8, 10, 30, 32] consisting of: physical elements and human artifacts; the social and cultural meanings assigned to them; the collective practices evolving within those spaces; and the individual and collective identities reflected by the landscape itself.

To bridge the gap between this integrated concept of lived landscapes and its application to digital media, we must address a fundamental ontological barrier. For humans, the landscape is experienced through a triadic ontology: a persistent physical object, its subjective representation (multiple and chosen case by case from a vantage point), and its cultural expression mediated by the “semiosphere” [21] and oriented by user intention. This structure relies on a conscious multilevel integration that allows us to recognize various fragments as useful simplifications traceable back to a complex, lived reality, and use them purposefully in digital settings.

Conversely, AI views the landscape through a strictly cartographic lens, where the map is mistaken for reality. This mirrors failed attempts to interpret landscapes purely as texts, as a literal textual interpretation excludes the social practices and physical elements that define the landscape as a dynamic process [11]. While these elements function as forms of language, they follow a logic entirely distinct from the syntactic, rules-based logic of linguistic structures [29]. Persisting in this error by intrinsic limitation, LLM operate within a monadic, flattened ontology that collapse objects, representations, and interpretations into a single, undifferentiated linguistic space. The AI does not perceive a transforming territory of what is and what ought to be, but a distribution of utterances—where meaning is derived from quantitative correlation rather than referential truth. In this process, the lived landscape—of which digital semantic markers are just a trace—is converted into a “taskscape” [17].

Where the human gaze sees a “totality of practices” and enduring meanings beneath media aggregations keeping the ability to discriminate their relevance and the object-intention-action alignment, the AI navigates hypertextual spaces where easily get lost, with effects similar to human users lost in the exitless labyrinth of hypertexts [9, 20]. Consequently, designing for effective Human-AI co-navigation becomes a form of landscape curation and see-sighting experience design.

3.2 From Gardens to Landscapes

This process of landscape curation mirrors the hypertext tradition of “gardening”, where the curator acts as an active gardener pruning and cultivating paths, and the medium is envisioned as an assemblage of textual flora "planted" to invite deliberate exploration [19, 22]. Much like a garden bounded by low walls, classical hypertext balanced openness with implicit order, a metaphor Mark Bernstein operationalised in his “Hypertext Gardens” [6]:

Gardens are farmland that delights the senses; parks are wilderness, tamed for our enjoyment. Large hypertexts and Web sites must often contain both parks and gardens.

For Bernstein, a successful hypertext garden requires clear boundaries and complexity to enrich exploration. This ensures meanings remain discoverable and navigable despite intricate connections. Even when generated algorithmically, the reader's experience remains an encounter with intentionality, the vital bridge between lived knowledge and text.

As the web expanded into social media ecosystems, the garden metaphor began to fray at the edges. Web 2.0 technologies, and social media in particular, eroded the “low walls” of classical hypertext, sacrificing structural complexity in favour of immediacy, content fragmentation, and media standardization. The personal garden of literary works and hand-linked websites was largely replaced by the intensive landfarms of social platforms. In these communal spaces, the experience of exploration is moved outside the user's control; readers no longer wander through a purposefully designed garden, but navigate a relentless stream of disconnected content without the guidance of a coherent, pre-designed arrangements, relying on their own ability to filter and making sense of contents and connections.

As AI begins to mediate and generate content, the garden metaphor defined by human intentionality and boundaries reaches its limit. The emerging digital ecology is less a curated plot than an expansive, overgrown wilderness where human agency is filtered through automated systems that plant and prune at scale, dissolving the boundaries between authorship, automation, and the structural integrity of the text. Against this horizon, we propose transitioning from the hypertext garden to the landscape as an appropriate metaphor for navigating the complexities of AI-driven media spaces.

The transition from Garden to Landscape reflects a pivot from the internal organisation of HT artefacts established by the author (an inward focus) toward the external relationality of AI-driven exploration (an outward focus). In this new paradigm, meaning is less about inherent structure and narratives, and more about how content positions itself within vast fluid media spaces. While the classical garden aimed to overcome "muddled writing” through structure and literary techniques, the landscape faces a newer threat: a weaponized clarity and homogeneity that breeds uncertainty, demanding a more rigorous form of assessment and critical differentiation.

Where the garden celebrated the aesthetic joy of a plurality of traversal experiences, wandering through pre-designed paths, the landscape offers an unbounded territory where paths are ephemeral, rendered in real-time through the “plausible combinations” of seemingly infinite and indeterminate media. This shift redefines human agency, from selecting the path to follow to developing the tactics for alignment (or the “affective attunement” of mutual understanding) within human-AI co-navigation loops, also by seeking meaningful pauses within a high-paced flow of outputs. Lastly, where the garden prioritized the resonance of a specific voice, the landscape emphasizes positioning and association, warning us that simulated coherence and surface-level patterns are no longer reliable indicators of actual conceptual or operational connections (see Table 1).

Table 1: Comparison between Bernstein's seven lessons from gardening with the principles of hypertext landscapes.

	Garden	Landscape
Challenge	Disorientation from muddle writing or subject complexity	Uncertainty from content plausibility, reinforced by writing clarity/homogeneity
Focus	Inward focus, priority of message and voice in structuring	Outward focus, priority on positioning, association, differenciation in relational spaces
Efficiency Trap	Shortest path are not always the best	Similarities are not always affinities
Positive Experience	Agency in aesthetic enjoyment of hypertext traversal	Agency in affective attunement in co-navigation with AI
Rhythm	Curated irregularities, avoided interruptions	Fostering pauses from high-paced flows
Territory	Clear boundaries and gateways	No boundaries, lighted paths
Paths	Pre-existing, designed to amplify the potential experience rather the existing content	Ephemeral, rendered through exploration of existing content and their plausible combinations

3.3 Sightseeing as Human-AI Alignment

As part of curating the landscape framing, foreground, middle ground, and background elements must be clearly distinguished. This ensures media artifacts are not merely stored as data—or unpredictably associated into plausible fictions but are anchored to a stable articulation of landmarks of meaning. This depth is achieved through a multilayered framework that integrates implicit knowledge (the cultural shadows and latent assumptions usually ignored by AI) preserving the triadic ontology of human reality rather than flattening it into a monadic linguistic reality.

In boundaryless environments, traditional navigation aids like linear history lists or global sitemaps fail to provide a true sense of orientation. Prompting thus emerges as a tool for designing sight-seeing experiences, acting as the primary mechanism to anchor AI and reclaim human agency. In this light, prompting becomes a form of hypertext authoring: users do not merely ask questions, they craft interfaces that define the interpretative frame. Consequently, the same model produces diverse readings of a corpus based on the specific authorial aid applied. The interaction is less a dialogue with a mind than a situated practice of inscription and path-finding.

By selecting specific regions and constraining traversals, prompting actively shapes both relevance and perspective. The generated outputs are then temporary paths whose quality is determined not by technical accuracy, but by interpretative filters—e.g., ethical, cultural, identity —that act as dynamic constraints. Then, prompting functions as a shared human-AI relational artifact; it does not merely add new facts, but provides the lenses and meta-skills necessary to discipline navigation against a clear horizon of evaluation. This rebalances the effort between human and machine, fostering the generation of new knowledge through a more precise alignment of intentions and actions.

In this context, prompts support an iterative process of scaffolding for AI as hypertext engines through: (a) Defining the Gaze. Revealing the various layers required to scan and read the landscape; (b) Pointing to Landmarks. Identifying coherent aggregations of explicit and implicit meanings that share affinities with the tasks at hand; (c) Beaconing Potential Paths. Assigning roles, guidelines, and referential anchors to assess dynamically evolving contents and practices; (d) Stabilising Routes. Creating paths that reflect a durable alignment between objects, intentions, representations, and actions.

4 Case Studies

To investigate and exemplify the presented new vision for hypertext, we developed a range of case studies about hypertext systems powered by AI and programmed to co-navigate landscapes with the user. Case studies were developed by expanding the methodology of two case studies on AI understanding of the web [2]. Each case study has been implemented at the stage of proof of concept or beyond, and documented as a brief that defines background, challenge, vision and use of AI, including an architecture that maps the information flow and the different uses of AI [anonymised]. Following, we briefly describe case studies, coded from i1 to i7, outlining the form of human-AI balancing (goal) achieved by producing knowledge from knowledge (how).

i1 Illegal Hate Speech.AI assists the legislative effort aimed at regulating illegal hate speech, producing and comparing different definitions against synthetic social media, generated to stress-test and anticipate misuse and unwanted consequences. The knowledge of the legislator guides AI in generating and analysing synthetic media content, to refine definitions sound with the legislative intention and operationable by both human and AI agents to identify threats to individuals and groups.

i2 Plant Phenotyping. AI assists agronomist in discriminating among plant variants of the same family, making a comparative analysis of, e.g., canopy shapes, to discriminate among likely alternatives. Agronomist knowledge is used to guide AI in identify what visible characteristics are of interest and how to make sense of analysis results in supporting classification. The balance is reframed by pushing to AI the generation and execution of tests tailored on the hypotheses that are identified thanks to human understanding of significant visual and contextual cues.

i3 Media Adaptation. AI assists adapting a culturally significant work into a new derivative form that celebrates and makes uses of its values and ideas. Human knowledge is used to guide the AI identification of the source core and to preserve it while generating a novel work. The balance is achieved by offloading to AI (1) the scan and analysis of commentaries, other derived works, and source, and (2) the exploration of alternative expressions in the target medium. Human author focus on the critique of the analysis and adaptation, and discriminating between false assumptions, and soundness between final result and source.

i4 Travel Guides. AI assists researchers in the free exploration of unstructured collection of travel guides. Research knowledge is structured by AI into a working theory that the AI uses in on-fly analysis of guides to condense a justificatory analysis of whether each resource fits and how within the research scope. AI offsets the burden of defining an ephemeral analysis pipeline that generates a curated collection approachable by human researchers.

i5 Semi-structured interviews. AI assists researchers in engaging with individuals from target groups about their experience. Human knowledge is structured into theories that are adapted by AI as questions, waived through the interview. AI takes on the burden of engaging and delivering interviews, generating novel questions that waive the concepts from human theories and hypotheses into the live conversation. AI also annotates interviews producing a structured corpora ready for researchers to be explored using the very concepts they provided to AI as input.

i6 Support to learning. AI adapts pedagogical approaches that fit the learner goals and preferences to deliver a support that balances notions with opportunities that stimulate individual study effort and reflection. The human knowledge is used to guide the extraction and adapting of teaching principles that produce novel interactive learning curricula for continuous learning and personal growth.

i7 Interactive Bibliotherapy. AI assists in the online delivery of mental health and wellbeing interventions. Human knowledge is structured in the construction of a library of narrative assets, enriched by a commentary making explicit the value and correct use of each one. AI produces novel, bespoke reading sessions tailored on the user. The AI shoulders the effort of triage users about their current struggles, find and adapt the narrative in their available time slot and guide reflection through targeted questions. Human focus on revising and improving assets and guides based on AI-generated insights from individual sessions and overall efficacy of the bibliotherapy program.

5 Results

The overarching landscapes, the scenarios that have being instantiated, were broken down as articulations of significant landmarks. Each landmark was defined as coherent, self-included steps result of defining the human and AI balance of responsibilities: user providing (i) reference to the sources and (ii) instructions on how AI where to engage and use them, (iii) AI tasked to apply such guidelines in the exploration of internal, provided and external knowledge to produce testable outputs. In i1 Illegal Hate Speech, AI takes four distinct roles with testable results such as synthetic social media or assessment under existing hate speech policies. Contrary to common prompting guidelines, our AI instructions were not heavy on guardrails and examples scarcely or not used at all. In i4 Travel Guides prompts included no examples but to structure an iterative discussion to clarify and refine an analytical framework directly from the user to be applied on a travel literature corpora. Indeed, the prompt refinement was instead an exercise in attunement, throughout an iterative process, aimed at correcting the AI gaze until able to discriminate significance as intended by the human user. In i3 Media Adaptation, most of the effort was dedicated to steering away AI from using tropes and classic images about rural XVII century villages from the source material to extract and remediate in contemporary terms the underlying sentiments and sense of dread and relief. Hence, each landmark was deemed complete when AI output was as grounded as possible – steered from baseline behaviour to reflect a manifested affinity with the human user expectations – and then consolidated through a Human-AI collaborative effort in structuring of replicable AI interpretative instructions (prompt). Each landmark definition was, per se, a fruitful exercise in reflection, rationalisation, structuring and decision making akin to coding but aimed at define, e.g., evaluation rather than procedures. In i7 Bibliotherapy, AI engaged users to identify an appropriate reading through ice-breaking small talks that could be take as much as necessary for AI to identify relevant insights. Indeed, it was the engaging and experiencing each landmark that, in essence, created novel ephemeral landscapes, broadening the human users’ perspective and enhancing their capabilities. In this view, the landscapes where hypertextual structures that could be traversed in depth and walk through across landmarks.

In i6 support to learning, human-AI exchange mix necessary meta-instructions with learning questions and feedback, deepening or broadening the scope of learning from the original design to the user emerging intents. Each traversal provided the opportunity to find a new balance that would fit the topic, effort and importance of the activity. Notably, the system was not used as pre-defined processes but blueprints to be instantiated case by case, just like even a standard tour offers a unique see-sighting experience, emerging from context contingency and human volition.

The underlying issues of responsibility and limitation of human-AI actors is being at the centre of all instantiations. For instance, i2 Plant Phenotyping aims at correcting the current AI-human roles, where AI is capable of gross classification while human experts must master all plant varieties and combination of features that can discriminate across variations and similar species. In this instantiation, human guide the logic used by AI in generating tests from the broad access to both analytical computer vision techniques and knowledge about plants. Another clear example was i5 Semi-structured Interviews where the responsibility of the AI was to use theories provided by human researchers to collect and curate a dataset about first-hand experiences to feed into a close reading process also involving human researchers.

6 Discussion

The case studies illustrated how specific prompting tactics and system architectures has been applied as a scaffolding process in Human-AI co-navigation. By 1) defining the gaze to reveal nested layers of information, 2) pointing to landmarks that anchor conceptual affinities, 3) beaconing potential paths through referential anchors, and 4) stabilizing routes to ensure a durable alignment between intention and action, these approaches provide functional support to maintain human agency over the AI otherwise unbunded mediaspace exploration.

The process of defining the gaze involves an iterative calibration of informational layers to bridge the gap between implicit human meaning and explicit media content. AI hallucinations or misinterpretations often stem from linguistic gaps where the implicit real-world richness fundamentally determines the meaning of media, but it is unreadable to the machine. To mitigate this, the co-navigation space defined and tested across the case studies was always structured through a dynamic interplay of foreground, background, and middle ground. The foreground consists of explicit, visible elements (text, layout, metadata) pointed by the user, that could assess their proper consideration in the AI explorations. The background is the set of cultural and social references, including ideal and concrete anchors, that bound the scope of the user to a specific direction. Together, they function much like aligning a rifle's sight: the background sets the trajectory, while the foreground frames the target. Combined, they allow both the user and AI to filter superficial semantic similarities against deep affinities, which are defined not by likeness but as topological homologies between the relations of the elements in the foreground and background.

We observed as the most transformative potential lies in the middle ground, activated by the dual-alignment of foreground and background working as a relational space and not as a hierarchical structure. It is in this interstitial space that AI excels, surfacing and uncovering novel conceptual connections that the user might otherwise ignore, expanding the starting knowledge and, at the same time, remaining grounded in the user's objectives. Once the landscape is calibrated, then, the AI can move forward goal-oriented discoveries and actively test the potential dissonance between a user's intentions and their expressions supporting output refinement. Reaching a shared representation of the reference landscape with AI cannot rely on a sequence of prompts provided as a static formula. It is exactly the engagement in the design of a dynamic see-sighting process that genuinely expand human cognition where intent, language and actions are aligned and mutually evolving. It becomes an authorial practice of self-expression at a metacognitive level.

Pointing to landmarks facilitates co-navigation by decomposing complex explorations into a series of conceptual monuments. Each landmark serves as an organic aggregate of media and prompts, unified by a singular perspective that provides the necessary context for moving from one point to the next. Unlike linear pipelines or database filtering, this process is not a technical sequence; it is an exercise in composing harmony among blocks of meaning, prioritizing a shared vision over a mere articulation of tasks. However, at the same time, this process requires a strategic deception of AI. Unlike human cooperation, which relies on shared long-term goals, AI cannot grasp a user's ultimate intent. Landmarks therefore materialize the user's vision within isolated, manageable thought-spaces and blocks of meaning, aligning AI operations to a broader goal without burdening the system with the complexity of the entire vision. Hence, having to manage the integration of competing, if not contradictory, perspectives.

For instance, in a city sightseeing tour, visitors immerse themselves in the distinct atmosphere of different monuments to to get an overall impression of the place. Analogously, a user-driven navigation of media spaces by AI requires a progressive encounter with distinct landmarks representing distinct thought spaces within the AI linguistic space, anchoring specific ways of thinking to the media objects at hand. By restricting the AI's focus to these local alignments, the user manages complexity sustainably, allowing even serendipitous new landmarks to emerge during the journey. The resulting knowledge is a multiplication and augmentation of the initial user's knowledge, achieved through the progressive traversal of these curated conceptual sites. Value is generated not through rigid execution, but through immersion.

Beaconing potential paths involves defining guidelines, assigning roles, and establishing referential anchors to trace possible trajectories through a landscape that functions as a palimpsest of artifacts, meanings, and practices. While standard prompt engineering often focuses on what the AI must do or must not do, our case studies across institutions, associations, and university courses highlighted that effective human-AI alignment requires guidelines focused on how the user will interact with and apply the AI's outputs within their specific professional or social context. By illuminating the human's operational constraints, the prompting process grounds the AI in a reality it cannot otherwise perceive, mapping out the viable routes for exploration.

On this basis, role-assignment shifts from asking the AI to mimic a static social or professional persona to defining its function within a relational human-AI cooperative dynamic. In this shared division of labor, roles are framed explicitly: I will perform X, while you process Y to support my downstream actions to achieve Z. These beacons ensure that AI outputs are not detached final results, but manipulable artifacts grounded in real-world practices. For example, in our semi-structured interview case study, the same set of theoretical concepts were transformed into proactive questions for engaging participants and curated datasets for researchers to analyse significant elements of the answers. By producing reasoned outputs linked to specific practices, both the user and the AI can dynamically evaluate the process as the context of human-decision-making evolves, ensuring that the tracing of new paths within the mediaspace remains well-oriented and sustainable accordingly to the respective human and AI capabilities.

Stabilizing routes involves creating paths that reflect a durable alignment between objects, intentions, and actions by managing the inherent instability of human-AI cooperation. Since AI lacks access to the user's decision-making context and underlying motivations, constant reciprocal feedback is essential to prevent the system from chasing a ghost it cannot see. To counter sycophancy and drift, users must provide feedback to every AI output to keep the exploration centered, but also actively require the AI to specify its understanding of the task and its constraints (beyond signaling generic reasoning) because potential frictions help to recalibrate the exploration. By treating the prompting sequence as a relational artifact rather than an executable program, this continuous feedback loop maintains the equilibrium necessary to prevent the AI from diverging from the user's evolving intentions within the unbounded mediaspace.

These combined operations translate the triadic nature of human reality into the AI's linguistic space, giving depth to otherwise flattened objects and representations. By articulating these distinct perspectives, the shared media landscape functions as better approximation of human perception of reality, allowing the user to reconstruct their complex reality for the AI through a guided sightseeing experience.

7 Conclusions – On the Future of Hypertext

AI presents an epoch-defining challenge for the Hypertext community; good news, since tackling technologies that augment human cognition has always been our focus.

In the most general terms, AI has the same potential as the introduction of Personal Computers and, it could bring forward a renaissance for the HT community, if taking on the task of defining the vision for AI as a human enhancing, not a human-replacing, technology. Hypertext must remap its core mission to a technology that is drastically different from classic computing. Our landscape metaphor fits within this challenge, reframing the traditional view of human-computer interaction in a new way sound with AI, relabancing uniquely Human experience of reality with the specific computational capacity of AI.

Part of the challenge is also about how inform the different communities of practice of tool makers, writers, artists and designers on the use of AI. To this end, the presented experimental case studies seed a library of scalable, reusable blueprints of human-AI collaborative landscape construction, exemplifying how the new concepts apply to real-life challenges. Furthermore, the experiments showcased the renewed centrality of human vision over the technical challenges of implementing new systems. AI operates as a processing engine for text-encoded human guidance, and AI's fast generating skills allow to treat outputs as fungibles, producing bespoke artefacts tailored and hard-coded to a human creative vision. This alone brings back the community to visioning avant-garde and experimental works over the tension of sustainability through standardisation that defined web technologies.

In conclusion, the (bright) future of Hypertext community hinges on if and how we are going to address this shift and provide to the rest of the technical and non-technical communities the vision, concepts, strategies, tools and guidance.

Source

ACM full-text HTML

ACM Digital Library

Imported from ACM’s structured HTML source. ACM Reference Format: Lucia Lupi, Sam Brooker, Riccardo Fedriga, Davide Picca, and Alessio Antonini. 2026. From Gardens to Landscapes: a new Foundation for Hypertext in the Age of Agentic LLMs. In 37th ACM Conference on Hypertext (HT '26), September 14--18, 2026, London, United Kingdom. ACM, New York, NY, USA 8 Pages. https://doi.org/10.1145/3800935.3830866

References

[1] Trond Aalberg, Carlo Bianchini, Marshall Breeding, Elena Corradini, Karen Coyle, Marija Dalbello, Claudio Forziati, Mauro Guerrini, Antonella Iacono, Giovanni Michetti, et al. 2020. Semantic web manifesto: the community of data, edited by Italian Library Association (AIB)–Study Group on Cataloguing, Indexing, Linked Open Data and Semantic Web (CILW). (2020).

[2] Alessio Antonini, Lucia Lupi, Mariusz Pisarski, and Sam Brooker. 2025. Literary Hypertext, AI, and Google's New Web: An Aesthetics Discussion. In Proceedings of the 36th ACM Conference on Hypertext and Social Media. 127–136.

[3] M Antrop and V Van Eetvelde. 2017. Landscape Perspectives. The Holistic Nature of Landscape Vol. 23 (Landscape). The Netherlands. Springer. doi 10 (2017), 978–94.

[4] Aleksander Aristovnik, Dejan Ravšelj, and Eva Murko. 2024. Decoding the digital landscape: An empirically validated model for assessing digitalisation across public administration levels. Administrative Sciences 14, 3 (2024), 41.

[5] Alejandro Bellogín, Paolo Giudici, Stefan Larsson, Jun Pang, Gerhard Schimpf, Biswa Sengupta, and Gürkan Solmaz. 2025. Systemic Risks Associated with Agentic AI: A Policy Brief. ACM Europe TPC-Autonomous Systems Subcommittee (2025).

[6] Mark Bernstein. 1998. Hypertext gardens: Delightful vistas. Eastgate Systems Web site, accessed at www. eastgate. com/garden, February (1998).

[7] Helen Bromhead. 2018. Landscape and Culture-Cross-linguistic Perspectives. (2018).

[8] Niclas Burenhult and Stephen C Levinson. 2008. Language and landscape: a cross-linguistic perspective. Language sciences 30, 2-3 (2008), 135–150.

[9] Robert Coover. 1992. The end of books. New York Times Book Review 21, 6 (1992), 23–25.

[10] André Corboz. 1983. The land as palimpsest. Diogenes 31, 121 (1983), 12–34.

[11] Vera Denzer. 2025. Landscape as Text: Textuality of Landscape. In Landscape Handbook: German Language Research Perspectives. Springer, 117–123.

[12] Umberto Eco. 1992. Interpretation and Overinterpretation. Cambridge University Press, Cambridge.

[13] Umberto Eco. 2000. Kant and the Platypus: Essays on Language and Cognition. Harcourt Brace, New York.

[14] Umberto Eco. 2014. From the Tree to the Labyrinth: Historical Studies on the Sign and Interpretation. Harvard University Press, Cambridge, MA.

[15] Riccardo Fedriga, Lorenza Saettone, and Emanuele Micheli. 2023. Linguistic and Cultural Competencies in Dynamic Possible Worlds. In Ontologies for Autonomous Robotics 2023. CEUR-WS.org, Aachen, 1–9.

[16] H. Paul Grice. 1989. Studies in the Way of Words. Harvard University Press, Cambridge, MA.

[17] Tim Ingold. 1993. The temporality of the landscape. World archaeology 25, 2 (1993), 152–174.

[18] Ziwei Ji, Nayeon Lee, Rita Frieske, Tianle Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of Hallucination in Natural Language Generation. Comput. Surveys (2023).

[19] George P Landow. 1992. Hypertext, metatext, and the electronic canon. Literacy online: The promise (and peril) of reading and writing with computers (1992), 67–94.

[20] George P. Landow and Paul Delany (Eds.). 1994. Hypermedia and Literary Studies. MIT Press.

[21] Yuri M Lotman. 2005. On the semiosphere. Σημɛιωτκ$acute{eta }$-Sign Systems Studies 33, 1 (2005), 205–229.

[22] Theodore H. Nelson. 1981. Literary Machines 931. Mindful Press.

[23] Theodore H Nelson. 1987. Computer lib/dream machines. Microsoft Press.

[24] Lucy Osler. 2026. Hallucinating with Ai: Distributed Delusions and ’Ai Psychosis’. Philosophy and Technology (2026).

[25] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, et al. 2022. Training language models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155 (2022).

[26] Claudio Paolucci. 2025. Nati cyborg. Cosa l'intelligenza artificiale generativa ci dice dell'essere umano. Luca Sossella Editore, Roma.

[27] Davide Picca. 2025. Not Minds, but Signs: Reframing LLMs through Semiotics. arXiv:2505.17080v2. arxiv:2505.17080 [cs.CL]

[28] Davide Picca. 2025. The Semiotic Channel Principle: Measuring the Capacity for Meaning in LLM Communication. arXiv:2511.19550. arxiv:2511.19550 [cs.AI]

[29] Edward Relph. 2019. Landscape as a language without words. In Handbook of the changing world language map. Springer, 3141–3154.

[30] Bernardo Secchi. 2011. Prima lezione di urbanistica. Gius. Laterza & Figli Spa.

[31] Robert C. Stalnaker. 1973. Presuppositions. Journal of Philosophical Logic 2 (1973), 447–457.

[32] Carlo Tosco. 2024. Il paesaggio come storia. Società editrice il Mulino.

[33] Michael Uschold. 2003. Where are the semantics in the semantic web?AI magazine 24, 3 (2003), 25–25.

[34] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35 (2022), 24824–24837.

[35] Jiadong Yu, DA Bekerian, and Chelsee Osback. 2024. Navigating the digital landscape: challenges and barriers to effective information use on the internet. Encyclopedia 4, 4 (2024), 1665–1680.

[36] Lorenzo Zangari, Davide Picca, and Riccardo Fedriga. 2025. Authorial Filtering and Computational Models: A Dynamic Analysis of Umberto Eco's Foucault's Pendulum through a Fluid-Dynamics-Inspired Framework. In Anthology of Computers and the Humanities, Volume 3: Computational Humanities Research 2025. 1122–1135. https://doi.org/10.63744/yVYI1Je9wQvM