The nature of science in the age of generative artificial intelligence: Epistemic authority, materiality, and a synergetic response

Synthesis: This conceptual literature review asks how the knowledge-producing character of text-based generative AI diverges from institutional science, and what that divergence means for teaching the nature of science in secondary and undergraduate science. Using the family resemblance approach, it compares science and conversational models across aims and values, methods, practices, knowledge, and social-institutional systems. The authors argue that a single feature—openness to material feedback—organizes the differences: science is an open system in continuous evidential exchange with material reality, whereas a model reorganizes a corpus closed to new evidence at inference. Through synergetic, they propose a coupled-system model in which the classroom verification norm is the order parameter, the teacher-set verification requirement is the control parameter, and materiality is the boundary condition. Five testable propositions follow, the last a falsifiable prediction of threshold behavior and hysteresis. The contribution is a design-oriented middle ground between adopting and banning AI: the pedagogical task is to keep critical evaluation and evidential accountability alive, not to decide whether the technology belongs in the classroom.

Key Findings

1. GenAI diverges from science in every family-resemblance category. Aims and values, methods, practices, knowledge, and social-institutional systems all differ, and each difference traces back to openness to material feedback.
2. Science is an open system; the model is closed. Science maintains its ordered structure through continuous evidential exchange with reality, whereas a model rearranges patterns within a corpus fixed before the conversation begins.
3. Fluency is read as reliability. Students may treat linguistic fluency as an indicator of epistemic reliability, and perceived credibility can track surface fluency rather than factual accuracy.
4. Fabricated references are common. In a classroom verification exercise, students checked 456 model-generated references against the literature, and 51.3% of those references proved real and correctly cited.
5. Trust calibration is designable, not fixed. Users' confidence systematically exceeded models' internal confidence, but adjusting explanations to reflect internal confidence narrowed both the calibration and discrimination gaps.
6. Classroom verification behaves as an emergent norm. The authors model the verification norm as an order parameter that constrains individual behavior beyond attitudes and prior achievement, with threshold and hysteresis properties.
7. Practical work is structurally necessary. Because materiality is the boundary condition that keeps the learning system open, laboratory and field activity becomes a foundation rather than an enrichment.

Comparing science and generative AI

The comparison is conducted through the five categories of the family resemblance approach. Science pursues explanatory coherence and predictive accuracy under socially enforced values such as honesty and organized skepticism; a model has no aims of its own, and its operational aims are helpfulness and fluency, with accuracy served only indirectly. Scientific methods eliminate error through engagement with the world; a model completes patterns over a fixed corpus and cannot intervene, collect data, or be surprised by a measurement. Scientific practices ground arguments in evidence the system itself has examined; generative models imitate that discourse, but the imitation is rhetorical rather than evidential. Scientific knowledge is a revisable network calibrated against observation; model knowledge consists of parametric associations without propositional commitment. Science operates through peer review and public accountability, whereas models are built under largely undisclosed training corpora and corporate governance. The authors argue that the large language model is best understood as operationally closed with respect to evidence at inference, and they treat opacity as a condition to work with rather than see through, drawing on explainability debates about the "black box."

From synergetic to a coupled-system classroom model

Drawing on synergetic, the authors reframe the contrast through open and closed systems and build a coupled-system model of the classroom. The student, the model, and scientific practice are treated as coupled subsystems. The teacher-set verification requirement functions as a control parameter: it is exogenous, readable from task specifications and assessment rubrics, and adjustable in graded steps. The classroom verification norm is the order parameter: it is endogenous, emerging as verification acts become mutually observable, and it constrains individual behavior in the manner described by the slaving principle. Materiality enters as the boundary condition, since laboratory work and data collection keep the learning system open. The framework predicts that raising the requirement produces discontinuous rather than proportional change in the norm, that partial withdrawal does not return the norm to its prior level, and that variability increases near the critical value. The authors stress that this is a structured analogy, not a claim that classrooms obey the equations of laser physics.

Design guidance from the coupled-system model

The framework yields design guidance. Generative output should be positioned as a hypothesis and the material world as the arbiter: students obtain a prediction from the model, conduct an investigation, and account for discrepancies. Practical work thus becomes structurally necessary rather than optional enrichment, though the authors concede that observational datasets, remote instruments, and calibrated simulations can sustain evidential accountability where laboratory access is limited. Assessment regimes are described as among the most powerful control parameters available, because regimes rewarding the polished product invite transfer of the task to the model, whereas those measuring the inquiry process, documentation of verification, and quality of evaluative judgment reward behaviors the model cannot perform. Teacher preparation should treat the epistemological comparison between science and generative AI as a distinct component of subject knowledge. The authors also flag Privacy, equitable access, and teacher workload as practical conditions, noting that reliance on capable systems and the freedom to decline them are unequally distributed.

What this means for practice

• Instructors. Position generative output as a hypothesis to be tested against material evidence rather than as an authoritative explanation, and make verification an explicit, assessed part of inquiry tasks.
• Assessment designers. Reward process-oriented assessment—documentation of verification, reasoning, and evaluative judgment—because regimes that reward polished products invite delegation to the model.
• Curriculum designers. Treat laboratory and field work as structurally necessary, since materiality is the boundary condition that keeps the classroom learning system open to evidence.
• Researchers. The five propositions specify testable designs, including a seven-wave design with a withdrawal phase to test threshold behavior and hysteresis in the classroom verification norm.
• Instructors. Address privacy, equitable access, and teacher workload before collecting interaction records, and treat the epistemological comparison of science and GenAI as subject knowledge.

Limitations

• The study is conceptual: the proposed relations have not been tested, and the design involved no participants.
• The comparison depends on the architectures available at the time of writing and will require revision for systems that acquire the capacity for evidential interaction.
• The synergetic vocabulary is used as a structured analogy, so its explanatory power is limited by how order parameter and threshold are operationalized in classroom data.
• Collapsing the social-institutional dimensions of the family resemblance approach into a single row leaves the political economy of model development outside the analysis.
• The scope is restricted to text-based conversational systems as consulted by learners; retrieval, tool use, and autonomous experimentation platforms are treated as boundary cases.

Citation

Muminova, Z. O., Mamatkulov, D. M., Turokulova, R. I., Rakhmonova, N. T., Shavqiyev, O. S. O., & Zoirova, S. U. (2026). The nature of science in the age of generative artificial intelligence: Epistemic authority, materiality, and a synergetic response. EURASIA Journal of Mathematics, Science and Technology Education, 22(10), em2943.