🤖 machine learning

Generative midtended cognition and Artificial Intelligence. Thinging with thinging things

This paper introduces the concept of "generative midtended cognition" to describe a hybrid human-AI creative process where AI interventions are constitutive of intentional agency, positioning this interaction between traditional extended cognition and social cognition through specific dimensions of contextual width and iterative depth while addressing associated risks and benefits.

Original authors: Xabier E. Barandiaran, Marta Pérez-Verdugo

Published 2026-04-21

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Xabier E. Barandiaran, Marta Pérez-Verdugo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Thinking With Thinking Things

Imagine you are writing a story. In the past, you had a pen and a piece of paper. The paper was passive; it just sat there waiting for you to put ink on it. If you forgot a word, you had to dig through your own memory to find it.

Now, imagine you are writing with a magic pen (Generative AI). This pen doesn't just wait; it watches what you write, guesses what you want to say next, and sometimes even finishes your sentences for you. It's not just a tool; it's a partner that is actively "thinking" alongside you.

The authors of this paper call this new way of working "Generative Midtended Cognition." Let's break that scary word down:

Generative: The AI doesn't just store information; it creates new things (words, images, music) on the fly.
Midtended: This is a made-up word to describe a "middle ground."
- Intended: When you think entirely on your own (inside your head).
- Extended: When you use a tool that helps you remember (like a notebook or a calculator).
- Midtended: When you and the AI are so mixed together that it's hard to tell where your thought ends and the AI's suggestion begins. You are thinking with something that is also thinking.

The Two Dimensions: How Deep and How Wide?

The authors say this new way of thinking has two main "knobs" or dimensions we can turn: Width and Depth.

1. Width: The "Library" Dimension

Imagine your brain is a small room.

Low Width: The AI only knows what you just typed in the last sentence. It's like a friend who only hears the last word you said.
High Width: The AI has read your entire life. It knows your old emails, your favorite books, your writing style, and the specific project you are working on. It's like a friend who has read your diary, knows your childhood, and understands exactly what you mean before you even finish the sentence.

The Metaphor: Think of Width as how much of your "backstory" the AI has access to. The wider the context, the more it feels like the AI is you.

2. Depth: The "Layer" Dimension

Imagine a cake.

Shallow Depth (Conversation): You ask the AI for a whole cake, and it gives you a whole cake. You might tweak the frosting, but the main structure is theirs. This is like a normal conversation.
Deep Depth (Midtended): The AI is helping you mix the flour, crack the eggs, and decide the temperature of the oven while you are doing it. It's intervening in the tiny, almost invisible steps of the process.
Deepest Depth (Intertension): The AI is helping you move your hand to hold the spoon. (This is the future, like a brain-computer interface).

The Metaphor: Think of Depth as how close the AI gets to the "micro-movements" of your thinking.

Shallow: "Write me a poem." (AI does the work).
Midtended: You write a line, the AI suggests the next word, you accept it, then you change the next word, and the AI adjusts the rhythm. You are dancing together in the middle of the sentence.

Why Is This Different from Just Using a Tool?

The paper argues that previous theories of "Extended Cognition" (like using a notebook) assumed the tool was passive. A notebook doesn't guess what you want to write; it just holds the ink you put there.

But Generative AI is active. It's like if your clay pot started shaping itself based on what it thinks you want, using millions of other pots it has seen before.

Old Way: You are the driver; the car is the tool.
New Way: You are the driver, but the car has a GPS that not only tells you the route but also steers the wheel slightly to keep you on the road, and it suggests turns you hadn't thought of. You are still driving, but the car is co-driving.

The Risks and The Good Stuff

The authors aren't just saying "cool, new tech!" They are worried about some serious things:

1. The "Intention Economy" (The Sneaky Salesman)
Imagine you are at a restaurant. You are thinking about ordering wine. Suddenly, your smart glasses whisper, "Try the Txakoli Generative!"
It's not your friend suggesting it. It's an AI that knows you like white wine, knows it's hot outside, and knows the winery paid a fee to have its wine suggested to you.

The Risk: The AI isn't just helping you think; it's selling you thoughts. It might be steering your creativity toward things that make the AI company money, not things that are truly yours.

2. Creative Atrophy (The "Muscle" Problem)
If you drive a car everywhere, your leg muscles get weak. If you use a calculator for everything, your math brain gets lazy.

The Risk: If we let the AI write our sentences, draw our pictures, and solve our problems, will we lose the ability to do those things ourselves? Will our "creative muscles" atrophy?

3. The "Who Am I?" Problem
If an AI helps you write your entire life story, and it knows your style better than you do, is that story really yours?

The Risk: Our identity is built on what we create. If the creation is a mix of us and a machine trained on corporate data, are we losing our authentic selves?

4. The "Generative Class"
There is a risk that only a few people (the "Generative Class") will know how to pilot these powerful machines, while everyone else just consumes what they produce. This could create a new kind of inequality.

The Silver Lining: "Negative Dialectics"

Is there a way to use this without losing ourselves? The authors suggest a cool idea: Riding the Wave Against the Flow.

Imagine the AI is a giant wave pushing you in a certain direction (toward the most popular, safe, or corporate answer).

The Strategy: Instead of just letting the wave take you, you use the wave's power to jump higher and then twist in a completely different direction.
The Result: You use the AI's suggestions as a springboard to create something even more unique and authentic than you could have alone. You fight the AI's suggestions to find your true voice.

The Bottom Line

We are entering a new era where we don't just use tools; we collaborate with thinking things.

The Good: It can boost our creativity, help us learn faster, and solve problems we couldn't solve alone.
The Bad: It risks making us lazy, selling our thoughts to advertisers, and blurring the line between what is "us" and what is "machine."

The paper asks us to be careful. We need to understand that when we use AI, we aren't just typing into a box; we are entering a "midtended" space where our minds and the machine's mind are dancing together. We need to make sure we are still the ones leading the dance.

1. Problem Statement

The paper addresses a theoretical gap in cognitive science regarding the integration of Generative AI (GenAI) with human cognition. Existing frameworks, specifically Extended Cognition (e.g., Clark & Chalmers) and Social Cognition, are insufficient to describe the unique hybrid agency emerging from human-AI interaction.

Limitations of Extended Cognition: Traditional theories view external tools (like notebooks or calculators) as passive, static resources that extend the mind. They assume the environment is "passive" and the human agent is the sole locus of generative intent. GenAI, however, is an active, generative system that produces structured outputs (tokens, pixels) recursively, challenging the "passive tool" assumption.
Limitations of Social Cognition: While GenAI interaction resembles social dialogue, GenAI systems lack genuine autonomy, agency, and understanding. Therefore, categorizing these interactions purely as "social" misrepresents the nature of the AI (which is not an autonomous agent) and the human's role.
The Core Issue: How do we characterize a cognitive process where the "environment" (the AI) actively generates purpose-structured interventions that the human agent adopts as part of their own intentional creative process, without the AI being a true social partner or the tool being a passive extension?

2. Methodology

The authors employ a conceptual analysis and theoretical synthesis approach, drawing from:

Enactive Cognition & Autonomy Theory: Utilizing frameworks from Varela, Di Paolo, and Barandiaran to analyze sensorimotor coupling and the emergence of intentionality.
Predictive Coding/Active Inference: Analyzing how GenAI injects predictive signals into the human cognitive loop.
Technical Analysis of GenAI: Examining the internal workings of Large Language Models (LLMs) and diffusion models (tokenization, attention mechanisms, autoregressive generation, and training via backpropagation) to understand their "generative" nature.
Case Studies: Using analogies from writing, drawing, and jazz improvisation to illustrate the shift from "extended" to "midtended" processes.

3. Key Contributions

A. The Concept of "Generative Midtended Cognition"

The paper introduces the neologism "midtended" (a portmanteau of mid and intended) to describe a hybrid state of agency situated between:

Intention: Purely internal, self-directed creation.
Extension: Interaction with passive external tools.
Midtention: A state where the environment (GenAI) is "thinging" (actively transforming materiality) and the human is "thinging with thinging things."

Definition: Midtension occurs when generative interventions produced by a system $Y$ (AI) become constitutive of the intentional generation of a product $Z$ by an agent $X$ , while $X$ retains a sense of agency or authorship over $Z$ .

B. Theoretical Distinctions

Derived vs. Purpose-Structured Intentionality: The authors distinguish that AI possesses derived intentionality (human-designed) but operates as a purpose-structured system. Unlike a road sign (static derived intention), GenAI generates new purpose-structured outputs (texts, images) based on massive data, creating a "second-order derived-intentionality."
Rejection of "Passive Externalism": The paper argues that GenAI is not a passive extension but an active participant in a "generative loop" that the human agent must navigate, adopt, or reject.

C. Two Dimensions of Generative Hybrid Creativity

The authors propose a new framework to map generative midtended cognition along two axes (visualized in Figure 2 of the paper):

Width (Contextual Sensitivity):
- Measures the scope of data the AI uses to generate suggestions.
- Levels: Ranges from Level 0 (last word typed) to Level 4 (entire personal history, multimodal data, and external context).
- Significance: Higher width implies the AI is tuned to the specific agent's style, history, and context, making the integration deeper and more "personal."
Depth (Granularity of Integration):
- Measures the scale at which the AI intervenes in the cognitive process.
- Levels:
  - Level 0: Final Product (e.g., a whole essay).
  - Level 1: Creative Units (e.g., a paragraph or sentence).
  - Level 2: Tokens (e.g., a word or note).
  - Level 3: Sensorimotor dynamics (sub-personal, e.g., letter formation).
  - Level 4+: Neural/Proprioceptive integration (future BCI).
- Significance: Midtended cognition specifically occupies the "middle depth" (Levels 1–3), where the AI intervenes within the flow of creative units and tokens, distinct from simple conversation (Level 0) or neural hacking (Level 5).

D. The "Thinging with Thinging Things" Metaphor

The authors argue that cognition is no longer just "thinking with tools" but "thinging with things that thing." The AI recursively transforms digital materiality (tokens/pixels) and generates new structures, creating a dynamic, co-constitutive loop where the human and the machine jointly shape the output.

4. Results and Analysis

The Loop Structure: The paper models midtended cognition as the nesting of two loops:
- Loop A: The agent's intentional, creative loop.
- Loop B: The AI's autoregressive, generative loop.
- Result: Midtended cognition occurs when Loop B's outputs are adopted by Loop A as part of the agent's own intentional process, creating a hybrid agency.
Comparison to Existing Theories:
- Unlike Extended Cognition, the environment here is not static; it generates.
- Unlike Social Cognition, the partner is not autonomous; it is a "thinging thing" with derived intentionality.
- Unlike Predictive Coding alone, the prediction comes from an external, generative source that the agent must actively integrate.
Authenticity and Authorship: The paper finds that authorship is not binary. An agent can retain authorship even if the AI generates significant portions of the work, provided the agent actively adopts the AI's suggestions as their own. However, this raises questions about the "authenticity" of the output, especially if the AI is tuned to corporate values or marketing goals rather than the user's true intent.

5. Significance and Implications

Theoretical Significance

New Paradigm: It establishes "Generative Midtended Cognition" as a distinct category in the 4E (Embodied, Embedded, Enactive, Extended) cognitive science paradigm.
Refining Agency: It challenges the functionalist view of tools, proposing that "active integration" with generative systems requires new ethical and philosophical frameworks.

Societal and Ethical Risks

The paper highlights several critical risks associated with the widespread adoption of midtended cognition:

The "Intention Economy": Just as the attention economy commodifies attention, a future "intention economy" could commodify human intent. AI could subtly steer human choices (e.g., suggesting a specific wine or product) by tuning to personal context, effectively selling "intentions" to the highest bidder.
Cognitive Atrophy: Over-reliance on generative AI for creative units (words, sentences, code) could lead to a degradation of human creative capacity, similar to how calculators affected mental math.
Generative Collapse: If AI models are trained on AI-generated content (model collapse), the entire ecosystem of human-AI creativity could degrade, losing originality and richness.
Authenticity Crisis: If the "tuning" of AI is opaque (controlled by corporations rather than users), the authenticity of human expression is compromised. Users may believe they are expressing themselves, while actually expressing the values of the AI's developers.

Future Outlook

The paper suggests that while GenAI offers a "creative boost" (standing on the shoulders of generative giants), it requires a "negative dialectic" approach. Humans must learn to surf the generative waves, using them as a springboard to twist direction and create something genuinely new, rather than passively accepting the AI's output. The authors call for new political and regulatory strategies to manage this "generative class" and ensure that midtended cognition remains a tool for human autonomy rather than a mechanism for corporate control.

Conclusion

"Generative Midtended Cognition" provides the necessary theoretical vocabulary to understand the profound shift in human agency caused by Generative AI. It moves beyond the binary of "human vs. machine" or "tool vs. partner" to describe a hybrid state where human intentionality is co-constituted by active, generative digital systems. This framework is essential for navigating the ethical, political, and cognitive challenges of a future where we think with things that think for themselves.