The paper is conceptual and programmatic rather than empirical: it offers concepts for future research and classroom design, not evidence that the approach improves test scores, attainment or other learning outcomes. The proposal begins with a change in what learners use to think. In Papert’s account, the Logo turtle was an “object to think with”: it responded to symbolic commands, and learners externalized their intentions through a cycle of coding, observing results and debugging errors.

Generative systems, Levin, Semenov and Gorsky argue, can instead function as bounded “agents to think with,” interpreting natural-language requests, generating alternatives and sometimes asking clarifying questions. Their responses are not conscious, intentional or morally accountable, but they introduce contextual variation into interactions that were previously organized around explicit instructions. This does not make formal computation obsolete.

Instead, it creates a hybrid environment in which learners, generative models, conventional software, symbolic tools and empirical observations can jointly shape an inquiry. Peirce’s semiotic distinction provides the framework’s technical foundation. A symbol acquires meaning through convention, an icon resembles what it represents, and an index is connected to its object by an existential or causal relation.

The authors distinguish three analytically separable forms of indexicality. Intralinguistic contextuality occurs when an expression is conditioned by other expressions; interactional indexicality occurs when its significance depends on a particular history of utterances and actions; and empirical indexicality occurs when signs connect to objects or events through observation, measurement, embodiment or a causal trace. A text-only language model considered in isolation has limited empirical grounding, but a complete human–AI configuration may acquire it through documents, software execution, sensors, measurements and participants’ contact with the world.

Meaning is therefore not inherently absent from generative AI; its connection to the world varies with the technical system and the situation in which it operates. To compare learning environments, the paper introduces three conceptual axes. The semiotic axis runs from predominantly symbolic to indexical meaning-making; the psycho-epistemological axis from learning regulated mainly by the individual to learning co-regulated through people and tools; and the pedagogical axis from instructionist transmission to constructionist making.

The authors position classical computer-assisted instruction near the symbolic, individual and instructional corner, Logo near the constructionist edge, collaborative coding as more dialogic, and AI-supported studios near the indexical, dialogic and constructionist corner. The numerical coordinates are illustrative author judgments rather than measurements. They also propose that movement toward situated meaning, dialogue and learner authorship may each support motivation, mapping these tendencies onto competence, relatedness and autonomy.

The authors explicitly describe that motivational relationship as a plausible, untested hypothesis that could also be affected by teacher personality, prior experience, affective state and social context. Dialogue has a central place in the proposed pedagogy because generative systems could make sustained, individualized conversation more economically feasible than conventional classrooms allow. The envisioned arrangement includes a learner, a teacher and an AI extension of the teacher.

The authors particularly favor using small or very small language models alongside larger systems because their visible errors can become objects of joint analysis rather than obstacles hidden behind a polished answer. A program with a detectable fault, for example, or a mathematical claim that requires scrutiny, can become material for inquiry. The proposed sequence is hypothesis, evaluation, justification and integration.

AI may help generate candidates, execute tools or reveal possible consequences, but learners should still compare claims with evidence and alternatives, explain why one explanation is preferable and connect the result to existing knowledge. The framework treats fluent output as only one part of learning and warns against substitutive cognitive offloading that removes the work through which judgment develops. The paper also reorders the traditional sequence of instruction.

Rather than always presenting a general rule before a concrete case, abductively initiated inquiry begins with a surprising event, a specific problem or an incomplete situation and asks what might explain it. Abduction generates plausible hypotheses; induction examines patterns and counterexamples; deduction derives consequences and tests whether an explanation holds. Generative AI can make alternatives and conjectures inexpensive to produce, but the authors reject any direct equation of probabilistic continuation with Peircean abduction.

Their stronger claim concerns the human–AI activity: a learner can treat an output as a candidate explanation, compare it with others and verify it through calculation, observation, executable code or proof. General theory remains essential, but it may emerge as a provisional stabilization of inquiry rather than as the instruction that begins it. The universal and the situated become recursively related, with new cases able to reopen a proposed generalization.

Programming remains the framework’s main model because its educational environments tightly couple three semiotic layers. A process can be seen iconically as a turtle’s visible path, expressed symbolically through words and formal grammar, and described indexically as an intention communicated to an AI system. Historically, programming education moved from rule-centered encoding, coding and retrospective debugging toward Papert’s constructionist culture, in which experimentation and debugging became central to understanding rather than merely correcting mistakes after execution.

Natural-language prompting and generative tools now shift more of the work toward contextual articulation of intention, while verification can involve a mixture of dialogue, conventional code, symbolic tools and empirical checks. The authors call this a hybrid-indexical era rather than pure indexical computation because symbolic execution, distributional generation, empirical reference, induction and deduction continue to coexist. Historically, this builds on Logo and Scratch rather than replacing their lessons with an entirely new educational tradition.

A central concept in the proposal is navigational thinking. In problems whose relevant representation is not yet stable, a learner must first discover what counts as the problem and which directions are promising before decomposition, formalization or execution becomes possible. Navigation reads the local structure of that problem space and adjusts course as evidence or unexpected output appears.

The authors stress that this is not a rigid sequence: abduction reframes possible directions, computation tests their consequences, and empirical or inductive evaluation can reopen the search. Interaction with generative systems also encourages what they call artistic navigation, extending Papert’s interest in bricolage. Prompting becomes a form of improvisational composition in which learners iteratively adjust wording, tone and context while reading subtle cues in a system whose response cannot be fully specified in advance.

This resembles the shift from an object that executes instructions to a partner that contributes contextual interpretation, although the learner must retain authority over goals, evidence and decisions. The framework remains deliberately unfinished. The authors identify unresolved questions about whether Papert’s constructionism was inherently indexical, how learning environments should be assessed along the three axes, whether the proposed motivational relationship holds across populations, and when a complete AI system has genuine causal contact with the world.

They also call for developmental research on what children, adolescents, university students and adults may appropriately delegate to AI, how independent reconstruction can be preserved and how process documentation can support feedback without becoming indiscriminate surveillance. Their conceptual conclusion is that schools cannot pretend generative systems are absent, but neither prohibition nor uncritical adoption adequately addresses the changed conditions of learning. If the framework survives testing, its practical aim will be a learner able to use AI for situated exploration while retaining responsibility for evidence, reasoning, independent reconstruction and judgment.

Subject of Research: A conceptual framework for learning with generative AI based on indexicality, constructionism and navigational thinking.

Article Title: The new constructionism. Papert after Peirce: indexicality, navigational thinking, and learning in the age of generative AI

Article References: Levin, I., Semenov, A., & Gorsky, M. (2026). The new constructionism. Papert after Peirce: indexicality, navigational thinking, and learning in the age of generative AI. Frontiers in Education, 11, 1902338. https://doi.org/10.3389/feduc.2026.1902338

Image Credits: AI Generated

DOI: 10.3389/feduc.2026.1902338

Keywords: generative AI, constructionism, indexicality, navigational thinking, education, programming pedagogy, abductive reasoning, human–AI learning, epistemic agency, computational thinking, Papert, Peirce

The research team, led by Khairul Azhar Jamaludin of Universiti Kebangsaan Malaysia together with colleagues at Universiti Malaya, framed the problem around a persistent weakness in educational robotics research. Studies of robot-based learning frequently report broad improvements in engagement, problem solving or computational thinking, yet STEM literacy itself has been defined in wildly different ways across the literature, spanning disciplinary knowledge, attitudes, skills and engagement. Without anchoring assessment to what a specific curriculum actually intends to teach, the authors argue, it is impossible to know whether a robotics intervention is producing the competencies a school system has promised. Their solution was to define what they call context-specific STEM literacy: a set of four competency domains drawn directly from Malaysia’s Year 6 Reka Bentuk dan Teknologi (RBT) curriculum, the national design and technology subject.

Those four domains were recognition of programming symbols and functions, recognition of programming interfaces and hardware, construction of pseudocode, and flowchart-based problem solving. The researchers built a curriculum-aligned test covering all four, validated it with three subject experts, piloted it for clarity, and confirmed good internal consistency with a Cronbach’s alpha of 0.85. The instrument combined thirty multiple-choice items with structured problem-solving tasks in which students had to analyse real-life technology scenarios and translate them into pseudocode sequences and flowcharts, yielding a total raw score from zero to eighty.

The intervention itself was a classroom-adapted version of FIRST LEGO League, the well-known robotics programme, redesigned through the ADDIE instructional design model so that every task mapped onto RBT curriculum standards rather than serving as a competition enrichment activity. Students in the experimental group, thirty-four children, moved through three deliberately sequenced phases: familiarisation, in which they learned the robot hub, motors, sensors and basic programming symbols; application, in which small teams built and programmed robots to complete movement challenges through iterative cycles of testing and debugging; and abstraction, in which they translated robot actions into step-by-step procedures, pseudocode and flowcharts. The control group, another thirty-four students, covered identical curriculum content through conventional teacher explanation, textbook work and worksheets, delivered by the same teacher with equivalent instructional time. Pre-test scores confirmed the two groups started from statistical parity.

The headline result is hard to ignore. After the intervention, the robotics group averaged 78.44 percent on the STEM literacy post-test against 64.85 percent for the control group, a difference the authors report as t(66) = 8.43, p < 0.001, with a Cohen’s d of 2.04, an unusually large effect for classroom research. Both groups improved significantly from their starting points, but the experimental group’s within-group gain carried an effect size of 2.69 compared with 1.22 for the controls. An analysis of covariance controlling for baseline performance reinforced the picture: group membership accounted for roughly 63 percent of the variance in post-test scores, with the adjusted means separated by about thirteen percentage points.

The more scientifically interesting story, however, sits one level down, in the domain-by-domain breakdown. All four competencies improved significantly, but the magnitude of improvement varied systematically. Programming symbol recognition showed an effect size of 1.87 and interface and hardware recognition 1.69, while pseudocode construction reached only 1.24 and flowchart-based problem solving just 0.83. In other words, the competencies most tightly coupled to visible, hands-on interaction with the robots grew fastest, while those demanding abstract symbolic representation grew more slowly, even though students had been explicitly taught to bridge the two.

Qualitative data from classroom observations, student reflections and teacher reflections explained why. Three themes emerged. First, a gap between task execution and written representation: students confidently built, programmed and debugged their robots, yet hesitated when asked to express the same logic on paper. One student told the researchers, I know what the robot should do, but I don’t know how to put it in steps; another said, When I see the robot, I understand. When it’s on paper, I get confused. Second, difficulty spiked during the transition to abstraction. Participation patterns shifted in the later sessions, with students taking longer, asking more clarification questions and, in some cases, showing frustration. We have to think about the robot and the steps and the symbols, it’s too much, one student said, while another observed simply that building is easier than writing. Third, students leaned heavily on the robotics context itself, mimicking robot movements with hand gestures or mentally simulating actions to answer written questions, suggesting their understanding remained tethered to the physical environment rather than transferring freely to decontextualised problems.

The authors interpret this pattern through cognitive load theory, which holds that working memory becomes strained when learners must coordinate multiple interacting elements simultaneously. Recognising a programming symbol or a sensor is a recognition task anchored in repeated, direct experience; constructing pseudocode requires translating concrete experience into formal, sequential, symbolic representations without immediate physical feedback. The study also acknowledges that the nine-session intervention gave students far more practice executing robotics tasks than practising representational translation, and that younger learners typically need repeated cycles of guided practice before procedural understanding consolidates into stable symbolic reasoning. The team is candid about other caveats: the close alignment between intervention activities and the assessment may have inflated gains, a novelty effect cannot be ruled out, the sample came from a single school, and the quasi-experimental use of intact classes limits causal certainty. The findings, they stress, speak to the integrated learning environment rather than to any inherent superiority of robots as a medium.

Why does this matter internationally? Because the core contribution is methodological. Aggregate STEM literacy scores would have shown a clean, celebratory win for robotics and concealed the uneven development underneath. By assessing curriculum-aligned domains separately and pairing the numbers with qualitative evidence, the study demonstrates that a single intervention can simultaneously excel at building situated, procedural understanding and underdeliver on abstract, algorithmic reasoning, a distinction most robotics studies never measure. For curriculum designers, the practical implication is concrete: robotics lessons should embed explicit scaffolding for representational translation, for example by having students verbalise the sequence of actions after a task and then convert that sequence into pseudocode and a flowchart, so that competence in manipulating a machine is deliberately connected to competence in symbolising its logic.

The researchers call for larger, multi-school samples, longitudinal designs tracking whether these gains persist, and deeper investigation of instructional strategies that carry students from hands-on interaction to formal computational representation. For now, the study offers a strikingly clear message for the growing global movement to put robots in classrooms: the robots work, but the paper is where the real learning challenge begins, and educators who ignore that gap may be training skilled operators rather than fluent computational thinkers.

Subject of Research: Effects of robotics-based learning on curriculum-specific STEM literacy development among primary school students

Article Title: A quasi experimental study of robotics based learning and differential development of context specific STEM literacy among primary school students

Article References: Jamaludin, K. A., Ebrahim, M. I., Nasri, N., Rashid, S. M. M., Kamaruzaman, F. M., Nasri, N. M., Hashim, H., & Alias, N. (2026). A quasi experimental study of robotics based learning and differential development of context specific STEM literacy among primary school students. Discover Education, 5(1), Article 1136. https://doi.org/10.1007/s44217-026-02250-x

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02250-x

Keywords: robotics-based learning, STEM literacy, primary education, quasi-experimental study, FIRST LEGO League, pseudocode, flowcharts, cognitive load theory, constructionism, Malaysia, design and technology curriculum, computational thinking