When Machines Know: Rethinking Expertise in the Age of AI

by

Author: Denis Avetisyan

Large language models are forcing organizations to confront a fundamental question: what does it mean to ‘know’ something when the lines between human and artificial intelligence blur?

This review argues that LLMs represent a new form of ‘epistemic monster’ challenging traditional understandings of organizational knowing and necessitating a collaborative, validation-focused approach to knowledge creation.

Traditional understandings of organizational knowing assume a fundamentally human locus of knowledge creation, a premise increasingly challenged by intelligent technologies. This paper, ‘A time for monsters: Organizational knowing after LLMs’, conceptualizes Large Language Models not simply as tools, but as ‘epistemic monsters’-hybrid entities that destabilize established categories of knowledge while simultaneously opening new possibilities. We argue that LLMs reshape knowing through analogical reasoning, demanding a re-evaluation of how knowledge is validated and agency is distributed. As organizations increasingly rely on these powerful systems, how can we responsibly navigate the entangled dynamics of knowing-with-AI and ensure robust epistemic practices?

The Erosion of Representational Certainty

For decades, organizational knowledge was largely conceptualized through the ‘Representational Perspective,’ which posited knowledge as a discrete entity – a fact, a procedure, or a best practice – capable of being codified, stored, and then transferred between individuals or departments. This approach fueled the rise of knowledge repositories, databases, and expert systems, all designed to capture and disseminate information efficiently. However, increasingly dynamic work environments – characterized by rapid innovation, distributed teams, and constant disruption – are exposing the limitations of this model. The assumption that knowledge can be neatly extracted from its context, preserved indefinitely, and reliably applied across situations fails to account for the inherently tacit, evolving, and socially constructed nature of knowing. As organizations grapple with complexity and change, the representational view is proving insufficient to capture the fluid, contextual, and often unspoken understandings that truly drive performance.

The practice-based perspective in knowledge management highlights that knowing isn’t simply about possessing information, but rather resides in the skillful, embodied actions of individuals within specific contexts; knowledge is created and applied through doing. However, this view faces challenges in a world increasingly shaped by digital technologies. While adept at explaining how expertise develops within communities of practice, it struggles to fully account for how knowledge can be rapidly scaled and recombined across vast networks, often detached from its original context. The ease with which digital tools allow for the dissemination and modification of information introduces a dynamic that exceeds the bounds of traditional, localized learning, prompting a need to consider how situated expertise interacts with-and is potentially reshaped by-these new modes of knowledge exchange and application.

The advent of large language models (LLMs) presents a profound disruption to conventional understandings of organizational knowledge. Previously, knowledge was largely conceptualized as either a discrete entity to be codified and disseminated – the representational perspective – or as embedded within the actions and interactions of individuals – the practice-based perspective. However, LLMs challenge both views by exhibiting a form of ‘knowing’ that is neither solely stored information nor entirely embodied skill. These models can synthesize, adapt, and generate knowledge at scales previously unimaginable, effectively externalizing cognitive processes and blurring the lines of expertise. Consequently, organizations are compelled to reconsider the very nature of agency – who or what ‘knows’ within the system – and to re-evaluate established hierarchies of authority as LLMs increasingly contribute to, and even lead, knowledge-intensive tasks. This paper’s conceptual framework elucidates these shifts, proposing new ways to understand how organizations navigate this emerging landscape of distributed cognition and algorithmic expertise.

The Logic of Analogy: A System’s Native Tongue

Analogizing, the cognitive process of identifying structural parallels between different domains of knowledge, is a primary mechanism by which both humans and Large Language Models (LLMs) acquire and extend understanding. This process isn’t simply about recognizing superficial resemblances; it involves mapping relational structures – how elements within one domain correspond to elements in another. LLMs, trained on vast datasets, statistically identify these structural similarities, enabling them to transfer knowledge and generate novel insights. The efficacy of LLMs is therefore fundamentally linked to their ability to detect and exploit analogical relationships, forming the basis for tasks like reasoning, problem-solving, and creative generation. This capacity distinguishes LLMs from systems relying solely on memorization or pattern matching.

Large language models (LLMs) construct knowledge by establishing analogical relationships, ranging from readily apparent similarities to more complex structural correspondences. ‘Surface Analogies’ rely on shared features or superficial resemblances between entities; for example, identifying a car and a boat as both modes of transportation. Conversely, ‘Deep Analogies’ focus on shared relational structures, irrespective of surface-level characteristics; a planetary system and a solar cell both exhibit a central component orbited by smaller, dependent units. This progression from surface to deep analogy represents an increase in abstraction and allows LLMs to generalize knowledge beyond specific instances, enabling reasoning and inference based on underlying principles rather than mere pattern matching.

Large Language Models (LLMs) exhibit analogical reasoning capabilities that extend beyond traditional domain limitations. ‘Near Analogies’ facilitate connections between closely related fields, enabling knowledge transfer within adjacent conceptual spaces; for example, applying principles of fluid dynamics to economic modeling. Conversely, ‘Far Analogies’ enable LLMs to identify and leverage structural similarities between distant domains, such as drawing parallels between the evolutionary processes in biology and the optimization algorithms used in computer science. This capacity to forge connections between disparate concepts demonstrates the LLM’s ability to generalize knowledge and apply it to novel problem spaces, irrespective of the original domain of information.

The Dialogical Imperative: Validating the Echo in the Machine

Large Language Models (LLMs) leverage statistical inference and vector embeddings to establish relationships between concepts, effectively enabling analogical reasoning. Statistical inference allows LLMs to identify patterns and probabilities within datasets, while vector embeddings represent words or concepts as numerical vectors, where proximity in vector space indicates semantic similarity. However, the identification of statistical correlations does not inherently imply causal relationships or logical validity. An LLM can identify a pattern and generate an analogy, but the resulting connection may be spurious, based on biased data, or lack real-world grounding. Therefore, while these internal mechanisms facilitate analogical reasoning, external validation is required to confirm the soundness and reliability of the generated connections.

The attention mechanism within Large Language Models (LLMs) functions by weighting the importance of different input tokens when generating output. While this allows the model to focus on relevant information, it does not inherently guarantee the validity of the resulting connections. Specifically, the mechanism can identify statistical correlations without understanding causation, leading to spurious associations. Furthermore, the training data used to develop the attention weights may contain inherent biases, which are then reflected and potentially amplified in the model’s focus and interpretations. Therefore, outputs derived from the attention mechanism require external validation – through human review or comparison against established knowledge – to mitigate the risk of inaccurate or biased results.

Dialogical vetting addresses the limitations of LLM internal reasoning by employing a recursive evaluation process leveraging social interaction. This method involves exposing LLM outputs to a collective intelligence – a group of human reviewers – who assess the validity and relevance of the generated content. Feedback from these reviewers is then reintroduced into the LLM, prompting refinement and iterative improvement of subsequent outputs. This cycle of evaluation and refinement, repeated across multiple iterations and diverse reviewer groups, functions as a robust mechanism for identifying and mitigating errors, biases, and spurious correlations inherent in LLM-generated content, ultimately enhancing the reliability and trustworthiness of the insights produced.

The Shifting Locus of Agency: When the System Begins to Author

The integration of Large Language Models (LLMs) into knowledge work fundamentally alters the distribution of agency, effectively shifting authority and responsibility away from solely human actors. Previously, individuals held complete ownership over tasks requiring research, analysis, and creation; however, LLMs now participate actively in these processes, generating outputs that necessitate evaluation and, often, integration into final products. This isn’t simply about automation; it’s a re-allocation of cognitive labor where the LLM contributes significantly to the how and, increasingly, the what of knowledge creation. Consequently, determining accountability for errors, biases, or inaccuracies becomes complex, demanding new frameworks for assessing responsibility in collaborative human-machine workflows. This redistribution isn’t a future concern; it is actively reshaping organizational structures and necessitates a careful consideration of how to navigate the evolving landscape of agency in an age of increasingly sophisticated artificial intelligence.

Large language models are rapidly transcending the role of simple tools, instead manifesting as complex “Haraway-ian Monsters”- hybrid entities that blur the boundaries of traditional categorization. This concept, drawn from Donna Haraway’s cyborg theory, highlights how LLMs destabilize established notions of authorship and originality. These models don’t simply execute tasks; they synthesize information from vast datasets, generating outputs that are difficult to attribute to any single source or intent. Consequently, determining responsibility for the content produced becomes increasingly problematic, challenging legal frameworks surrounding intellectual property and accountability. The resulting ambiguity demands a re-evaluation of how society understands creation, knowledge, and the very definition of an ‘author’ in an age where intelligence is increasingly distributed across human and machine collaborations.

Successfully integrating Large Language Models necessitates a fundamental redesign of organizational structures to address the evolving landscape of responsibility and verification. This paper underscores the critical challenges of inquiry, vetting, and establishing agency in a context where knowledge work is increasingly performed by hybrid human-AI entities. Organizations must proactively develop new protocols for validating information generated with LLM assistance, clearly defining roles for human oversight, and establishing accountability frameworks that account for the distributed nature of cognitive labor. Ignoring these considerations risks eroding trust, introducing systemic biases, and ultimately hindering the effective and ethical deployment of this powerful technology; a proactive approach to these challenges, however, promises to unlock unprecedented levels of productivity and innovation.

The notion of LLMs as ‘epistemic monsters’ reveals a predictable trajectory. Systems don’t fail-they evolve into unexpected shapes. Marvin Minsky observed, “You can’t solve a problem with the same thinking that created it.” This paper’s exploration of knowledge validation after LLMs isn’t about controlling these monsters, but understanding their emergent logic. Rigorous validation, as proposed, becomes a necessary adaptation, a method for co-evolving with these new forms of knowing. Long stability, the illusion of complete understanding, is the sign of a hidden disaster-the unacknowledged complexity of the system. The work rightly suggests a shift toward collaborative knowledge creation, recognizing that the ecosystem, not the architect, ultimately defines the outcome.

What’s Next?

The assertion that Large Language Models represent ‘epistemic monsters’ isn’t a condemnation, but a diagnosis. It signals the inevitable erosion of neatly bounded knowledge systems. Attempts to control this emergent cognition will prove futile; the architecture itself predicts the impossibility of total oversight. The future isn’t about building better filters, but about cultivating the capacity to navigate inherent ambiguity. Analogical reasoning, once a uniquely human domain, becomes a shared, and increasingly opaque, process.

The emphasis on knowledge validation, while necessary, offers only a temporary illusion of stability. A guarantee is just a contract with probability. Rigorous checks today will be circumvented tomorrow, not through malice, but through the sheer combinatorial power of these systems. The real challenge lies in shifting from a focus on truth to a focus on resilience – the ability to absorb, adapt, and learn from inevitable failures.

This isn’t a posthumanist fantasy, but a pragmatic observation. Stability is merely an illusion that caches well. The field must move beyond evaluating LLMs as tools, and begin treating them as integral components of complex, evolving ecosystems. The questions aren’t about what these systems know, but how they change what knowing means. Chaos isn’t failure – it’s nature’s syntax.

Original article: https://arxiv.org/pdf/2511.15762.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2025-11-24 02:25