Generative midtended cognition and Artificial Intelligence. Thinging with thinging things

In plain words

When an AI finishes your sentence and you keep the suggestion, who did the thinking? The authors call this mixed process "generative midtended cognition". The AI's suggestions take part in your creative process, yet you still feel you are its author. The middle word, "midtended", refers to that in-between state: not fully your intention, and not just a tool you use from outside.

The paper offers two ways to describe such a process. Width asks how much context the AI draws on, from the last word typed up to everything you have ever written or read. Depth asks how fine-grained the back-and-forth is, from finished products down to single letters and, at the extreme, brain signals. The authors place this kind of coupling between a chat with a colleague and low-level neural processes.

The paper is philosophy, not an experiment. It also worries about what this does to authenticity, to who holds the power over the AI systems, and to people's own creative skill.

TL;DR

A conceptual framework from philosophy of mind, not a method with an implementation. It defines "midtension" §3.2 and classifies human-AI creative coupling along a width axis (5 levels) and a depth axis (6 levels) §3.3. It reports no experiments or measurements. Read it for vocabulary and distinctions; do not expect parameters to implement.

Problem

Extended-cognition theories treat external resources as passive and under the agent's control; trust and glue criteria assume the environment has no generative power §2.2. Generative AI breaks that assumption, because transparent tools can now pursue purpose-structured interventions of their own §2.2. The paper asks how to describe such hybrid creative processes and where they sit among conversation, social cognition and extended cognition §1 §2.2 §3.1.

Method

The paper is a conceptual analysis. Its core is a definition, a two-loop picture and two dimensions, with no procedure to run. We state each in our own words.

Technical background the framework relies on. Generative systems produce output token by token. Each output is fed back in as input (for language models), while diffusion models refine a noisy image step by step §2.1. Human edits can enter this loop at any point §2.1.

Definition of midtension §3.2. Given a person X, a generative system Y and a product Z, midtension holds when interventions produced by Y become constitutive of X's intentional generation of Z, and X keeps a sense of agency or authorship over Z.

# Restatement of the midtension test and the two axes (no code in paper)
# Input:  agent, system, product (the paper's X, Y, Z)
# Output: label for the coupling (not a numerical score)

def classify_coupling(agent, system, product):
    # 1. Midtension needs the system's output in the agent's act  [§3.2]
    if not interventions_constitutive(system, product, agent):
        return "no midtension"
    # 2. Agent must still keep a sense of authorship  [§3.2]
    if not keeps_authorship(agent, product):
        return "no midtension"
    # 3. Place the coupling on the two axes  [§3.3] [fig. 2]
    width = context_level(agent, system)    # 0..4, see Width below
    depth = granularity_level(agent, system)  # 0..5, see Depth below
    return (width, depth)

Adoption condition §3.2 fig. 1. The paper describes two loops. The A loop is the person's intentional creative loop; the B loop is the system's autoregressive loop. Running side by side is not enough. The person must accept or adopt the system's interventions into the A loop without losing the intentional character of that loop. Suggested reasons for adoption include predictive fit with what the person was about to write, trust, familiarity, and simple laziness or the feeling that one's own standard is lower §3.2.

Width §3.3 fig. 2. How sensitive the generative process is to context. The paper's academic-writing examples give five levels:

• level 0: only the last word typed;
• level 1: the previous sentence;
• level 2: the whole text being written;
• level 3: wider context, such as chosen readings and journal guidelines;
• level 4: everything the person has written or read, including multimodal material.

Depth §3.3 fig. 2. The granularity of reciprocal interaction. The paper gives six levels:

• level 0: the final product (text, song, picture);
• level 1: creative units (a sentence, a paragraph, a musical phrase);
• level 2: perception-action tokens (a word, a note);
• level 3: sensorimotor dynamics, usually sub-personal (a letter);
• level 4: the proprioceptive body, for example bodily prostheses;
• level 5: the neural scale, for example brain-computer interfaces.

Three bands along depth §3.3. Integration at the level of complete units is "conversation", the kind of exchange with colleagues or a co-author. The deepest band is "introtension", the neural scale, which the paper treats as largely out of reach. The band between them is "midtended cognition": deeper than complete creative units, up to the sensorimotor level. The paper does not state the exact level numbers at the band boundaries.

Ethical and social dimensions §4 §5. The paper discusses authenticity (tuning to the person's context and to the company's values can both shape the output), generative power asymmetry (private companies control the systems), an "intention economy" in which products are injected into a user's suggestions, creative atrophy, and collapse under recursive training on AI output. It also names a possible benefit: using the system's suggestions as something to argue against §4.

Hyperparameters

Not applicable. The paper states no equations beyond the verbal definition §3.2, no algorithms and no numerical parameters.

Results

Not applicable. The paper reports no experiments, benchmarks, user studies or measurements. Its numbers are illustrative: for example, the description of GPT-3's 96 processing blocks §2.1, or a context window of up to 2 million tokens for Gemini, given in a footnote §3.2. These are the authors' own statements and have not been checked.

Paper vs code

No official code. The paper has no code repository. Its links are the Synthese DOI in the header and a ChatGPT share link in a footnote §2.1. No project page exists. A GitHub repository search for "generative midtended cognition" returned no results.

Pitfalls

• Two terms for one idea. The abstract speaks of defining "generative midtended cognition", but the formal definition in the body is of "midtension" §3.2. The text switches between "midtension" and "midtended" without a stated difference. Treat the §3.2 definition as the one to use.
• "Generative" is a neutral label. The paper uses it for processes that produce structured outputs in a creative sense, not to mean "built with a generative model" §3.1. Do not read the term as a technical claim.
• Depth bands have no numeric cut-offs. The midtended band is described in words, not by level numbers §3.3. Any mapping of a concrete interaction to a level is an interpretation, not something the paper specifies.
• Width may be split. A footnote in §2.2 proposes splitting width into a context dimension and a temporal "Length" dimension for the agent's history. The main text merges them and does not develop the second. The paper itself says the two dimensions could be extended §3.3.
• Width assumes accuracy. The authors say that treating width as high precision and accuracy is an idealisation, because systems hallucinate and can be inaccurate by design §4.
• Authorship is a condition. Midtension requires the person to keep a sense of authorship §3.2. The paper says that such a sense does not guarantee an authentic product of that person's agency §4. Do not infer authenticity from the definition.
• Speculative sections. The discussion of an intention economy, creative atrophy and future scenarios §4 is argument, not evidence. Do not cite it as a finding.
• Source version. This card is based on the LaTeX source of the Synthese post-print, not on the arXiv PDF or v1. The source says it is a post-print of Synthese 205:137 (2025). We did not compare it with other versions.
• Figures not reproduced. fig. 1 (loops) and fig. 2 (dimensions) are described in words only. Their captions credit third-party artwork in part, so we do not publish them.

Lineage

• Builds on: the "midtension" concept, introduced in an earlier 2024 arXiv article by the same first author §1; extended cognition with the parity principle and the trust and glue criteria §2.2; Heersmink's dimensions of integration §2.2 §3.3; enactive approaches to cognition §3.2.
• Compared against: passive versus active integration in extended cognition §2.2; material engagement theory §2.2; postphenomenology's cyborg intentionality §5; social cognition and participatory sense-making §3.2.
• Superseded by: not known from this card's sources.