Inferring Gifted Potential from AI-Assisted Work: An Attributional Validity Framework for Gifted Education

Synthesis: An AI-assisted product can be excellent and still say almost nothing about the learner who produced it, and this conceptual article names that problem attributional validity for gifted education. Sak treats the relation from capability to product as many-to-one: learner competence, model capability, the learner's direction, human–AI fit, and context combine differently to yield equivalent artifacts. The framework separates four targets of inference — independent competence, intellectual Learner Agency, hybrid capability, and developmental carryover — and names three errors that follow when one target's evidence is read as another's: overattribution, developmental illusion, and presumed capability equivalence. Programs should name the capability they mean to identify, document how the work was produced, and require later evidence of transfer.

Key Findings

Product quality cannot recover learner competence. Hybrid performance varies with learner competence, AI capability, the learner's direction, human–AI fit, and context, so different configurations produce equivalent products.
Four targets of inference must be judged separately. These four answer different questions and need different evidence; mixing them is the attribution error.
Overattribution reads a system accomplishment as personal capacity. Disclosure names the assistance without revealing its contribution, and neither prompt counts nor word counts identify the learner's consequential decisions.
Developmental illusion mistakes supported performance for growth. AI can raise performance during support while learners show weaker later independent performance, and fluent assisted work inflates what they believe they can reproduce.
Presumed capability equivalence converts output similarity into learner similarity. AI-assisted products can converge across learners and narrow diagnostic differences, hiding exceptional problem framing in one learner while compensating for weakness in another.
Carryover requires evidence collected after support is reduced. Assisted improvement supports a developmental claim only when later independent performance shows reconstruction, retention, or transfer against an earlier baseline.
Identification records must describe conditions, not just products. Model and version, interface, permitted functions, adult mediation, and access conditions belong in the record.

Why a strong product is weak evidence

Assessments support interpretations about exceptional capacity and educational need rather than standing as scores in themselves. Generative AI breaks the transparency of that inference by contributing content, strategies, criticism, and revisions, so the observed product becomes an inverse problem: reasoning backward from an artifact to the capacities that produced it. The same tool can give different learners different intellectual contributions. Neither authorship statements nor disclosure of AI use repair that, and neither does reliability: a dependable product score can still support a weak conclusion about an individual's capacity. The remedy is to specify whether a claim concerns the learner, the learner–AI configuration, or development over time.

Four targets of inference

Attributional validity asks for evidence at four targets. Independent competence is what a learner can demonstrate without AI assistance; it is best established by a comparable unaided task in the same period, with task order and prior experience documented. Intellectual agency is the learner's control over what is asked, accepted, rejected, and revised — visible in problem framing, constraint generation, principled rejection, and integration, not in prompt frequency, which can signal dependence. Hybrid capability belongs to the documented learner–AI configuration and cannot be read as independent competence. Developmental carryover asks what changed in the learner, requiring later reconstruction or novel transfer under reduced support.

Three errors that follow

Overattribution reads a hybrid product as personal capacity: the product may be excellent while the learner's contribution stays unmeasured, and a disclosure line does not close that gap. Developmental illusion reads supported performance as growth: AI can lift performance during assistance while later independent performance falls, and fluent assisted work can inflate learners' estimates of what they can reproduce. Presumed capability equivalence reads similar outputs as similar learners, even though Creativity research shows AI can raise individual output while reducing diversity across outputs — a convergence that hides exceptional problem framing in one learner while compensating for weak understanding in another. The three coexist and produce both false positives and false negatives.

Designing identification around the claim

Three principles guide practice. Output equivalence: comparable AI-assisted products may come from different levels of competence, agency, contribution, and opportunity, so products cannot be ranked as measures of the learner. Carryover: assisted improvement supports a developmental inference only when later performance shows learning against a baseline or transfer to a novel task. Claim-specific convergence: a person-level attribution is strongest when sources converge on the same target, while disagreement defines a profile. Before Assessment, programs should name the target and match evidence to it — unaided tasks for independent competence, interaction records for Learner Agency, documented assisted tasks for hybrid capability, transfer tasks for carryover. Comparability requires documenting model and version, permitted functions, adult mediation, and access conditions, since uneven access can let differences reflect tools rather than potential.

What this means for practice

Instructors. Judge AI-assisted work by how it was produced, not how good it looks: collect the interaction record, the consequential choices, and the learner's reasons before reading the product as evidence of advanced standing.
Assessment professionals. Name the target before assessing and match evidence to it — unaided performance for independent competence, process evidence for Learner Agency, documented assisted performance for hybrid capability, transfer tasks for carryover.
Administrators. Document model and version, interface, permitted functions, adult mediation, and access conditions whenever AI-assisted work informs nomination, placement, or acceleration.
Researchers. Treat the three errors as testable hypotheses, not findings: the framework derives them from theory and estimates no rate.

Limitations

The article is conceptual and reports no empirical validation: no sample, instrument, or estimated effect establishes that the three errors occur at any given rate.
Its mechanism evidence comes from other learner populations; the author says such research identifies mechanisms, not how strongly they operate in gifted education.
Presumed capability equivalence is a theoretically derived risk, not an established prevalence claim.
The framework sets no operational thresholds — how much later independent performance, after how long, counts as carryover remains open.

Connected Concepts

Connected Articles

Which inference is at risk? Assessment validity reasoning and generative AI — Which inference is at risk? Assessment validity reasoning and generative AI
Coauthorship integrity: Reconceptualising assessment validity for the age of generative artificial intelligence — Coauthorship integrity: Reconceptualizing assessment validity for the age of generative artificial intelligence
The LLM Fallacy: Misattribution in AI-Assisted Cognitive Workflows — The LLM Fallacy: Misattribution in AI-Assisted Cognitive Workflows
Towards Scalable Measurement of Durable Skills — Towards Scalable Measurement of Durable Skills
ChatGPT in Education: An Effect in Search of a Cause — ChatGPT in Education: An Effect in Search of a Cause
Cognitive Offloading in the Age of Generative AI: What Does It Mean for Students With Learning Disabilities? — Cognitive Offloading in the Age of Generative AI: What Does It Mean for Students With Learning Disabilities?
Reclaiming Epistemic Agency: A Critical Framework for Human-Generative AI Co-Agency in Education — Reclaiming Epistemic Agency: A Critical Framework for Human-Generative AI Co-Agency in Education
"Polished Artifacts, Fragile Engagement? Tackling the Challenge of Reduced Epistemic Effort in Human-AI Knowledge Construction" — "Polished Artifacts, Fragile Engagement? Tackling the Challenge of Reduced Epistemic Effort in Human-AI Knowledge Construction"
The Tragedy of the Cognitive Commons: How AI Could Disrupt the Regeneration of Professional Expertise — The Tragedy of the Cognitive Commons: How AI Could Disrupt the Regeneration of Professional Expertise

Citation

Sak, U. (2026). Inferring Gifted Potential from AI-Assisted Work: An Attributional Validity Framework for Gifted Education. EdArXiv preprint.

Metadata

Study design	Theoretical Analysis
Discipline	Learning Sciences
Level	K 12
Intended audience	Instructors, Assessment Professionals, Administrators, Researchers
Learning and instruction	Creativity, Transfer of Learning, Metacognition, Scaffolding
Technologies and techniques	Generative AI, Prompt Engineering
Assessment and measurement	Assessment Validity, Assessment, Educational Measurement, Evaluative Judgment
Equity, ethics, and responsible use	Equity, Differential Effects Across Learner Groups, AI Use and Disclosure Statements
Foundations of AI in education	Human AI Collaboration, Learner Agency, Theories and Frameworks, Limitations in AIEd Research
Page kind	Framework

Source History Edit Chat