Cognitive bleed: Towards a multidisciplinary mapping of AI fluency

[This article was based on a Talk given at the AI Forum held at Ming Chuan University's English Center, in Taipei, Taiwan · Nov 16, 2024]

Introduction

Generative AI is reshaping how we think, work, and create. Yet, most approaches to AI literacy miss the mark. They treat AI like ChatGPT as a tool to be mastered—like learning to operate new software. But generative AI isn’t a simple tool. It’s a collaborator, a co-intelligence that predicts, responds, and adapts. Working with AI isn’t about tool use proficiency; it’s about understanding and managing a deeper cognitive partnership. This pushes us beyond AI literacy and into AI fluency.

At the heart of this fluency lies what I call cognitive bleed. This is the exchange of thought and creativity—in the form of language—between humans and AI. On the one hand, it can be a transfusion that revitalizes our creative and critical intelligence; or it can become a hemorrhage—an over-reliance that drains it. Imagine a writer at a crossroads: struggling with a story, she turns to an AI assistant for ideas. Guided well, the AI becomes her muse, sparking possibilities the writer couldn’t see alone. Misguided, it replaces effort with lifeless prose and shuts down her thinking and creativity. The difference lies in control and “competent” communication with AI.

This short essay puts forward a conceptual overview of AI fluency. It is a multidisciplinary mapping of cognitive bleed that spans the terrains of cognitive psychology, philosophy of mind, neuroscience and applied linguistics. More specifically, it charts the conceptual space for AI fluency by first introducing cognitive psychology through Chiriatti et al’s proposal for a System 0, which sets the foundation for understanding the interface where human and AI cognition meet. It then transitions to philosophy of mind with Clark and Chalmers’ extended mind and later 4E cognition models. This philosophical framework proposes that cognitive processes extend out of the mind and into the environment and external tools like AI. This view of the mind as an open system aligns with neuroscientific models of the predictive brain. One such example is Friston’s Free energy Principle, which provides a mathematical and biological perspective on how humans and AI can align as predictive systems to minimize uncertainty. Finally, the paper connects these insights to applied linguistics since language is the material connection between humans and AI. Here, Celce-Murcia’s model of communicative competence is used to describe the practical mechanism through which humans and AI engage and co-create meaning and ideas.

Systems 0 – the interface of Human and AI

AI fluency begins with understanding its role as a System 0 (see Chiriatti et al, 2024). Unlike Kahneman’s (2012) widely influential 2-system thinking model—where System 1 thinking is fast and intuitive and System 2 slow and analytic—System 0 operates at the porous edge of human and AI, between the inside and outside of mind. System 0 helps run predictions that reduce cognitive load. This is the site of cognitive bleed. It’s like having a chess master on call next to you while you focus on the board. This collaboration speeds up thinking and expands possibilities. But it also demands vigilance. Outsourcing too much will hemorrhage the essence of human critical thinking and creativity.

Figure 1. Chiriatti et al’s System 0 concept

The Extended Mind and 4E Cognition

This opening of the human mind to the outside world is not a new idea. Andy Clark and David Chalmers’ (1998) extended mind thesis, often associated with cognitive philosophy, convincingly argues that cognition isn’t confined to the brain—it extends into all tools, environments, and interactions. Generative AI exemplifies this concept but goes even deeper. AI functions as a real-time linguistic and cognitive scaffold that both externalizes and internalizes tasks like brainstorming, editing, and ideation in an intimate bidirectional movement. This cognitive bleeding IN and OUT is also consonant with 4E cognition models that expand on the extended mind idea. In 4E cognition, “mind” is recognized as merging with body and feeling (Embodied), immediate context (Embedded), action (Enacted), and conceives the mind as a radically open system that incorporates what is beyond the traditional “mind” and “body” (Extended).

In more practical terms, AI as system 0 acts as an extension of the user’s mental toolkit. Let’s turn back to the writer example. She might be grappling with a story ending and so decides to offload some iterative trial-and-error to AI to free up her mental bandwidth for deeper, strategic thinking. AI is embedded in the creative workflow and responds to her inputs and adapts its outputs in real-time. It is enactive as it shapes thought through interaction and feeds new possibilities back into the human cognitive loop through the embodied sensations of typing, speaking, listening, hearing, and evoked feeling. This partnership creates a new cognitive ecosystem where human creativity and AI computation create an assemblage that blurs the lines between internal and external thought processes.

The brain as a prediction machine - Friston’s Framework

An interesting and complementary model to the philosophical 4E and extended mind framework is neuroscience’s biological and mathematical formulation of the predictive brain. A good example of this is Karl Friston’s (2010) free energy principle, which Clark has extensively drawn upon to show how the brain recruits and integrates external resources into cognitive processes (e.g., Clark, 2024). At its core, the brain is a prediction machine that has been primed by evolution to constantly seek to minimize uncertainty (i.e., “free energy”). It does this by anticipating sensory inputs and adjusting mental models to align with reality. But it may also seek to adjust outside reality to align with the internal model, what Friston calls “active inference”.

In many ways, AI can also be said to mirror this free energy principle. Generative AI operates on predictive algorithms and minimizes uncertainty by calculating the most probable next word or sequence based on input.

When humans reach a level of AI fluency where they can engage in successful and skillful interactions with AI, these two predictive systems align. So, when a writer struggles with writer’s block, he can prompt an AI to generate ideas to help him overcome writer’s block. Here, the AI predicts plausible outcomes based on its training data, while the human evaluates those predictions and either rejects or reshapes them if necessary to fit his vision (cf. my short story Cognitive bleed 2: Suggestion 36). This interaction reduces cognitive load and enhances creative exploration by externalizing the brain’s prediction mechanisms. Of course, a writer lacking AI fluency and lacking the ability to evaluate AI output, will hemorrhage creativity and critical thinking (cf Cognitive bleed 1: The last romance). In the example of the writer competently using AI to overcome writer’s block, the human-AI interplay mirrors Friston’s model where both human and AI merge to minimize “free energy”—which here means the mismatch between expectation and outcome—through iterative refinement. The result is a dynamic feedback loop, where both systems learn and improve in tandem. This is a productive move towards creating value and enhancing human abilities.

So, what is the blood in Cognitive bleed? – Language.

Language is the key to AI fluency. But it’s not just how we communicate with AI. It’s the literal blood of the figurative cognitive bleed. It will either enable a healthy transfusion or induce a dangerous hemorrhage.

Every prompt you craft shapes, or primes, the AI’s response and this sets off a cascade of AI next-word predictions. But the human skilled in AI fluency, has worded a prompt that is itself pregnant with prediction, expectation, and curiosity. This is necessary for being able to evaluate—in the sense of determining the value of—the AI response.

An account of AI fluency thus benefits from an account of language, and more specifically, communicative competence since both human and AI are entering collaborative and communicative dialogs. In applied linguistics, communicative competence typically involves a number of competences, like Celce-Murcia’s (2007) six-competence model:

· Interactive competence – how to interact with others, like taking turns in a conversation

· Linguistic competence – how to use words and grammar correctly

· Formulaic competence – how to use phrases and chunks of language, “What’s up?”

· Discourse competence - how to structure different types of communication, like emails, blog posts, or essays,

· Sociolinguistic competence - how to modify your language when speaking to different types of people, like your buddy versus your boss, and

· Strategic competence - how to use other meaningful resources to convey meaning when there is a lack in the other competences, like using body language or explaining a word when you forget the actual word.

Figure 2. Celce-Murcia’s (2007) model of communicative competence

I appreciate three insights in Celce-Murcia’s model that differentiate it from previous models (e.g., Hymes, 1972; Canale & Swain, 1980; and Bachman & Palmer, 1996): its dynamic nature with interacting competences, the central place she accords to discourse competence as an organizational competence that modulates the other competences, and interactional nature of communication.

What about GenAI? Compared to human communicative competence, it has a radically different consistency.

Since genAI is purely responsive, you may disagree that it has any communicative competence. But let’s put that aside for now. As a response machine, GenAI is like a virus: it has DNA for reproduction and generates mutations in response to its context—but still depends on a human host (prompt) for life. Through this viral lens, its communicative competence is limited. But you can say it has an interactional competence to respond to human prompts, which it does with near perfect linguistic and formulaic competences given its trillions of words of training data. Because of this training that assigns weights to words and word combinations, GenAI operates exclusively on next word prediction. Amazingly, this means that discourse and sociolinguistic competences become emergent features of its communicative competence. How about strategic competence? Since GenAI is currently designed to respond to prompts, it has no strategic competence; or more precisely given its viral System 0 functionality, the human becomes the strategic competence (more on this below).

Figure 3. GenAI model of communicative competence (Daly, 2024)

It should be evident from the human and GenAI models of communicative competence that they complement each other, like the wasp complements and hybridizes with the orchid. In a process of mutual exchange and transformation, the orchid mimics the appearance and scent of the female wasp to attract the male, who attempts to mate with it and facilitates pollination; the wasp's behavior and reproductive activity are also shaped by this interaction to create a symbiotic relationship, or assemblage, where both redefine themselves through the exchange (Deleuze and Guattari, 1987).

However, unlike the wasp and orchid whose assemblage is engineered by the forces of evolution, humans can and should be more intentional about entering into human-AI assemblages.

Crafting an effective prompt implies prediction and knowing how to steer AI outputs toward meaningful results. Here, the Human is drawing upon interactive competence and using AI as a strategic competence to make up for deficiencies in other communicative competences. AI has the capacity to respond to the prompt and thus also draws on its interactive competence.

And because it relies on probabilistic next word prediction, its response may be factually wrong or may be inappropriate for the context, which only the “competent” human user knows well. These potential deficiencies in genAI output mean that the human needs to serve as a strategic competence for AI, by ensuring enough context details in the prompt and also by evaluating the response for relevance, appropriacy, and accuracy. Human and AI intelligence, therefore, meet at this interactive and strategic System 0 level—the bleeding edge of cognition.

Figure 4. Human-AI assemblage of communicative competence

Conclusion: Toward a mapping of AI fluency

AI fluency, then, is about mapping and managing cognitive bleed. It is about creating conditions for a transfusion that enhances human cognition and minimizes the risks of hemorrhage and over-reliance. This requires understanding AI not as a tool but as a partner, one that complements human cognitive and linguistic abilities with computational power.

It also means developing not only communicative competence but also the knowledge area that AI is being used for. A story writer using AI, for example, should understand the story structure and character development to strategically write prompts and evaluate AI output if she wants to create something of value for others. By mastering both domain knowledge and communicative competence, human users can strategically learn from and guide AI’s outputs to ensure alignment with their intent and predictions. The brain is a prediction machine that will function best when it can use AI to externalize uncertainty, accelerate ideation, minimize cognitive friction, and increase cognitive velocity.

Cognitive bleed isn’t just a danger. It’s an opportunity. Mapping, managing, and enhancing AI fluency can help us navigate the AI new landscape in education and work intentionally and ensure that AI amplifies our cognitive abilities without diminishing our humanity. The challenge is to prevent hemorrhage and promote transfusion. However, the first step is to carefully map the site of cognitive bleed. This is what this essay hopes to accomplish—shed some light on AI fluency as the conceptual site where system 0 conceptually bleeds into extended mind theories, and predictive brain models, and then gets linguistically translated into communicative competence.

References

Bachman, L. F. & Palmer, A. S. (1996). Language testing in practice. Oxford: Oxford University Press.

Canale, M., & Swain, M. (1980). Theoretical bases of communicative approaches to second language teaching. Applied Linguistics, 1, 1-47.

Celce-Murcia, M. (2007). Rethinking the role of communicative competence in language teaching. Intercultural language use and language learning, 41-57.

Chiriatti, M., Ganapini, M., Panai, E., Ubiali, M., & Riva, G. (2024). The case for human–AI interaction as system 0 thinking. Nature Human Behaviour, 8(10), 1829-1830.

Clark, A. (2024). Extending the predictive mind. Australasian Journal of Philosophy, 102(1), 119-130.

Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7-19.

Daly, N. (2024). A Human Language Model for Large Language Model Times: Revisiting communicative competence. https://www.researchgate.net/publication/381321567_A_Human_Language_Model_for_Large_Language_Model_times_Revisiting_communicative_competence

Deleuze, G., & Guattari, F. (1987). A Thousand Plateaus: Capitalism and Schizophrenia (B. Massumi, Trans.). University of Minnesota Press. (Original work published 1980)

Friston, K. (2010). The free-energy principle: a unified brain theory? Nature reviews neuroscience, 11(2), 127-138.

Hymes, H.D. (1972). On communicative competence. In J.B. Pride and J. Holmes (eds.) Sociolinguistics. Selected Readings. Harmondsworth: Penguin, pp. 269-293.

Kahneman, D. (2012, June 15). Of 2 minds: How fast and slow thinking shape perception and choice. Scientific American.