Extended Executive Cognition: A Learning Outcome for the AI Age

Synthesis: Sidorkin (2025) proposes Extended Executive Cognition as a higher-education learning outcome: the capacity to allocate cognitive effort, decompose a task for human or machine execution, hold an accurate mental model of what AI can and cannot do, and integrate its output into a defensible result. The paper presents a framework, not a study: it builds the construct from curriculum theory, executive function research, distributed cognition, cognitive load theory and theory of mind, then argues that AI literacy and ethics instruction are insufficient without it. It sketches curriculum integration and process-based assessment across general education and several disciplines, and reports no empirical results: no sample, no intervention, no measured outcome.

From content production to cognitive orchestration

Higher education has answered AI through two channels, technical training in tools and instruction in ethics, and Sidorkin argues that both leave the central cognitive work unaddressed. Tool courses teach how systems function; ethics courses teach what ought to be permitted. Neither teaches a student to decide which part of a problem belongs to the model or how to assemble a defensible result. The missing capacity is metacognition aimed at the boundary between human and machine.

The account of work begins with what AI does. It drafts documents, reviews literature and makes predictions, so tasks sort into three classes: exclusively human, best performed by AI, and hybrid — predictive analytics and liability questions in law, radiology algorithms that read images more accurately than some practitioners, and leadership teams that orchestrate analytics instead of performing it.

That last shift carries the paper's only quantitative workplace claim. Citing survey data from Fortune 500 companies, Sidorkin writes that executives now spend 60% more time on scenario planning and ethical oversight than on direct data analysis compared with pre-AI workflows. The figure traces to a 2020 AI-in-education article, not an executive survey, so its provenance cannot be checked here. The qualitative claim is that "orchestration value" accrues to coordinating human and artificial intelligence, and expertise is redefined accordingly: memorization, procedure execution and standard document production lose value, while ethical reasoning, cross-disciplinary synthesis and tolerance of ambiguity gain it. The supporting evidence is mixed: experienced professionals resist AI through identity threat yet handle its complexity better, while novices gain more from AI training.

Defining the construct against AI literacy

Extended Executive Cognition is defined as the ability to strategically allocate cognitive effort, coordinate AI-assisted workflows and manage problem-solving that integrates AI capabilities with human expertise. Its components are task decomposition, mental models of AI, metacognitive awareness of one's strengths relative to the machine, and strategic delegation.

"Executive" borrows from cognitive psychology: inhibition, working memory and cognitive flexibility oversee basic operations and enable planning, problem-solving and metacognition. "Extended" borrows from distributed cognition, where tasks spread across human and non-human agents and performance depends on coordination. Cognitive load theory supplies the mechanism: as AI absorbs task complexity, human capacity moves toward oversight, evaluation and integration. That is the opposite of cognitive offloading: effort is relocated, not removed.

Much of the argument concerns what students must understand about the model. Sidorkin describes current large language models as probabilistic predictors calculating statistical relationships between tokens without necessarily possessing human conceptual frameworks. A working mental model must therefore include that AI lacks intentionality while producing goal-directed behavior, processes patterns without understanding them, and fails in novel contexts differently from human error. The correlation-causation gap is the worked example: the model surfaces correlations, the human interprets them causally.

The contrast with AI literacy is the sharpest definitional move. Literacy means understanding AI's functions and limits; Extended Executive Cognition means active cognitive management, the difference between reading comprehension and rhetorical skill. Technical education covers the first, ethics education the normative questions, and neither prepares a student to run a hybrid workflow. Accurate mental models need explicit instruction; unstructured exposure breeds naive anthropomorphic theories. The construct also has an equity dimension: access to tools confers limited advantage without a metacognitive frame, and students who develop the frame extract more from limited access — a suggestion offered without evidence.

Curriculum integration and assessment

The proposal has a general tier and a disciplinary tier. General education would host foundational courses emphasizing cognitive workflow design, strategic delegation and decision-making in hybrid systems rather than basic technical instruction, and would make them required of every graduate regardless of major. The design advice is constructivist and backward: authentic problems requiring coordination of human and artificial intelligence, structured around essential questions about effective cognition, following Wiggins and McTighe's learning design framework. The disciplinary tier asks each field how AI changes its workflow and which judgments stay human — AI-assisted research in STEM, hybrid team management in business, AI as creative collaborator in the humanities, AI in social science research methods, AI diagnosis in clinical simulation.

Assessment is where the framework becomes most concrete and least settled. The requirement is to judge process rather than product: how a student constructed a workflow, decomposed a task, delegated components and integrated results. The paper lists think-aloud protocols, workflow documentation frameworks and performance-based assessments on novel problems as validated tools, then adds metacognitive reflection assignments, analyses of student-AI logs, annotated process diagrams and decision trees. It also concedes that these methods require substantial development and testing and that no standardized metric for metacognitive development in AI contexts exists. That tension matters for assessment validity: a process score is meaningful only if the process it records is the process that produced the outcome, and no validity evidence for that inference is offered.

Key Findings

1. The paper advances Extended Executive Cognition as a distinct outcome: the ability to execute a complex cognitive task with AI assistance efficiently and validly, built from task decomposition, mental models of AI, metacognitive self-awareness and strategic delegation.
2. The construct extends three core executive functions — inhibition, working memory and cognitive flexibility — from coordination inside one mind to coordination across human and artificial agents.
3. The construct is separated from AI literacy on the model of reading comprehension versus rhetorical skill: literacy is understanding how AI works, while Extended Executive Cognition is purposeful orchestration toward an intellectual goal.
4. The single quantitative workplace claim is that Fortune 500 survey data show executives spending 60% more time on scenario planning and ethical oversight than on direct data analysis compared with pre-AI workflows, a figure attributed to a 2020 AI-in-education article rather than to the survey it names.
5. The expertise evidence is inconsistent: experienced professionals resist AI through identity threat yet handle its complexity better, while novices gain more from AI training and experienced teachers write better prompts.
6. Development is claimed to require structured experience — authentic hybrid problems, explicit feedback on allocation decisions, repeated revision of AI mental models — yet no fully validated instrument exists and the proposed developmental scale is never given levels.
7. The assessment menu is process-based, but the paper calls think-aloud protocols, workflow documentation and novel performance tasks validated while conceding that they need substantial development, that infrastructure is missing and scoring is time-intensive enough to threaten scalability.

What this means for practice

• Curriculum designers. Write the outcome as a process statement — decompose, delegate, verify, integrate — not a list of tools, and give it a required general-education home.
• Instructors. Grade the workflow artifact alongside the product. A decision log showing what went to the model, why, what was checked and what was rejected teaches more than a polished submission does.
• Assessment designers. Treat the proposed instruments as prototypes, not validated measures, and build validity evidence before attaching grades to a process score.
• Discipline faculty. Map where AI actually changes the workflow in your field and which judgments must stay human; the paper's sketches are prompts for that mapping, not templates.
• Institutions. Do not assume capable students absorb orchestration on their own. Access without an explicit metacognitive frame reproduces advantage.

Limitations

• The paper is a theoretical commentary and presents no empirical validation: no sample, no intervention, no comparison group, no measured learning outcome. The claim that the construct predicts post-educational success is asserted, not tested.
• No standardized metric exists. The one instrument attempt cited is a preprint by the same author, and the paper's description of three assessment tools as "validated" contradicts its own admission that the methods need substantial development.
• The 60% workplace figure is attributed to Fortune 500 survey data but cited to a conceptual article on AI in education, so its source cannot be verified here. The paper also flags a circulating "40% increase over a semester" as needing evidence it does not present.
• The construct is not empirically distinguished from neighboring ones — metacognition, self-regulated learning, AI literacy, trust calibration — so a new label's necessity is unproven.
• The prescriptions are tied to a particular generation of generative AI behavior: probabilistic prediction, contextual brittleness, prompt-level collaboration. The general argument may survive capability change, but the specific mental-model content may date quickly, as the author acknowledges.
• The scope is a US higher-education framing with no treatment of cost, faculty workload or student differences; the paper names the last as future research.

Citation

Sidorkin, A. M. (2025). Extended executive cognition, a learning outcome for the AI age. Computers and Education Open, 9, 100294.