The study, published in Discover Education by Alexander M. Sidorkin, examined a single 16-week semester of a course in child and adolescent development. Thirty-six students enrolled, most of them juniors and seniors at a regional public university serving a substantial Hispanic-serving and first-generation population. Most arrived with no prior coursework in qualitative research, no familiarity with academic publishing conventions, and little experience writing extended scholarly manuscripts. Yet the course required each of them to produce a full-length qualitative research manuscript of at least 5,000 words and to submit it to a peer-reviewed journal, with confirmation of submission as part of the final assignment. This was not a simulation of professional work; it was the work itself.

The pedagogical gamble was that generative AI could absorb the procedural burden that historically made such ambitious assignments impractical for novices. Authentic assessment, the idea that students should be evaluated on tasks resembling real professional practice, has struggled for decades with a built-in tension: authentic tasks demand competencies students have not yet developed, and the scaffolding needed to bridge that gap often dilutes the authenticity it is meant to support. Sidorkin’s argument is that a capable language model changes this calculus. A student who previously could not construct a conventional methods section or navigate disciplinary terminology can now obtain that support on demand, shifting the locus of difficulty away from routine production and toward decisions about evidence, interpretation, and methodological fit.

Crucially, the course did not treat AI as a threat to be policed. Students used a custom course-specific tool, the CHDE 111 Class Companion, a configured GPT that offered navigation, concept clarification, assignment protocols, and rubric-based draft review while explicitly instructed to support learning without providing completed assignments. They were also free to use ordinary ChatGPT or other major platforms without limitation, and the syllabus placed responsibility for the quality, veracity, and originality of the final product squarely on the student. For most assignments, students submitted complete logs of their AI conversations alongside their work, making part of the human-machine interaction visible for feedback and analysis.

To measure whether unrestricted AI access compressed student outputs, the study applied three rule-based indicators to the 33 final manuscripts. The Methods Specificity Index, a 0-to-6 scale scoring the presence of six methodological components such as a named qualitative approach, a concretely specified data source, sampling logic, analytic procedure, ethical safeguards, and a rationale connecting method to question, showed dramatic upper-bound compression. The median score was the maximum of 6, nearly 70 percent of manuscripts received the full score, and 97 percent scored either 5 or 6. Methodological specification, in other words, had become something AI scaffolding could reliably deliver.

The other two indicators told a different story. Claim-Evidence Coupling, the proportion of interpretive statements locally accompanied by specific evidentiary support such as a quotation, citation, or reference to the study data, ranged from 0.000 to 0.750 with a median of 0.333. Theory-Interpretation Linkage, the proportion of interpretive statements that explicitly invoked the manuscript’s own theoretical framework, ranged from 0.000 to 0.917 with a median of 0.444. Both remained substantially dispersed across the corpus. Students differed enormously in whether their interpretive claims were tethered to actual evidence and whether theory genuinely informed their analysis, dimensions that AI can assist with but cannot resolve, because warranting a claim depends on the actual data and argument of each individual project.

This differential pattern is the study’s central empirical finding. Compression appeared where structural scaffolding suffices; dispersion persisted where situated judgment is required. The theoretical framework behind the study, drawing on the five-dimensional authentic assessment model of Gulikers and colleagues, holds that when professionals routinely use AI, an assessment that categorically excludes it may actually resemble professional work less faithfully. Student contribution, in this view, is not independent text production but the direction, evaluation, integration, and revision of work produced within a human-AI system, a capacity the study terms extended executive cognition, paired with discerning thinking, the ability to judge AI output for substance and fit rather than equating fluency with quality.

Detailed comparison of four contrasting students, two from the top and two from the bottom of the course-points distribution, illuminated what those differences look like in practice. The stronger students orchestrated their AI use: they specified tasks precisely, supplied information the model could not infer, such as which social media discourse threads belonged in the analysis, repeatedly asked the Companion to audit sections for coherence, and pushed through multi-turn revision cycles until the conceptual fit was right. One student rejected a Companion-generated framework summary as misaligned with her research question and iterated across several turns until satisfied. The weaker students, by contrast, engaged in sparse, one-directional interactions, accepting template language and structural outlines without constraining them to the specifics of their actual studies. Their manuscripts could look competent at a glance, but theme statements were thin on evidence and theory was named rather than deployed.

The study is candid about its limits. It is a single-semester case study without a comparison group, so it cannot isolate AI orchestration from prior preparation, effort, or research aptitude. The three indicators are heuristic measures applied with the assistance of ChatGPT rather than validated instruments, and the instructor, course designer, tool designer, analyst, and author are all the same person, a concentration of roles that creates real confirmation-bias risk despite safeguards such as fixed rules and independent case selection. The process evidence also covers only a late slice of the semester, so it cannot show whether orchestration practices developed through instruction or simply differed among students from the start.

Even with those caveats, the implications are significant for anyone designing courses in the AI era. The study suggests that authentic assessment can remain viable and diagnostically useful when instructors explicitly distinguish between dimensions AI can scaffold reliably and dimensions where student judgment remains consequential, then align tasks, feedback, and rubrics accordingly. It also suggests that AI orchestration itself should become an instructional object, modeled and taught rather than assumed, and that when AI can generate prose quickly, the bottleneck shifts toward reading, evaluating, and revising it. Whether the pattern transfers across disciplines and institutions remains an open empirical question, but the Sacramento case demonstrates that giving every student a powerful AI assistant did not make them interchangeable. The tool was common; the judgment was not.

Subject of Research: Authentic assessment and student judgment in an AI-integrated undergraduate qualitative research methods course

Article Title: Authentic assessment in an AI-integrated qualitative research methods course

Article References: Sidorkin, A. M. (2026). Authentic assessment in an AI-integrated qualitative research methods course. Discover Education, 5(1), Article 1153. https://doi.org/10.1007/s44217-026-02273-4

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02273-4

Keywords: authentic assessment, generative AI, qualitative research methods, undergraduate education, evaluative judgment, human-AI collaboration, assessment design, rubric, extended executive cognition, pedagogy, case study, academic writing

The problem the researchers set out to solve is one that has quietly plagued educational AI for years. Modern classrooms are saturated with information: a teacher’s vocal prosody, gestures, and board work; students’ gaze patterns, emotional states, and hand-raising rhythms; and the intricate back-and-forth of classroom discourse. Any single one of these channels tells only a fragment of the story. Yet most existing AI evaluation systems either ignore this multimodal richness or, more critically, produce accurate predictions without explaining themselves. A system that spits out a quality score of 74 with no justification is of little use to a teacher hoping to improve. The field, as the authors put it, has tended to prioritize prediction over explanation.

MFIR-ME tackles both problems at once through a five-branch architecture that treats five distinct categories of classroom signals as first-class citizens. An acoustic encoder built on a bidirectional LSTM with self-attention pooling processes vocal features such as speech clarity and pacing. A linguistic encoder based on multilingual BERT captures the semantic structure of classroom discourse. A visual encoder combines ResNet-50 frame analysis with a GRU to track gestures and movement, while a temporal convolutional network models the rhythm of the lesson. Finally, a graph attention network maps the web of student interactions, capturing engagement patterns that no other channel can see.

The real technical innovation lies in how these streams are combined and interpreted. A Cross-Modal Attention Fusion module, or CMAF, projects each modality’s output into a shared 256-dimensional latent space and uses eight attention heads to dynamically weight how much each channel should contribute in a given context. Rather than fixing the importance of, say, acoustic features in advance, the model learns that their relevance may shift depending on what is happening linguistically or behaviorally. The fused representation then feeds both a prediction head that outputs a teaching quality score from 0 to 100 and, crucially, an explanation pipeline.

That explanation pipeline is where the framework departs most sharply from prior work. Instead of relying on a single interpretability method, MFIR-ME runs four complementary feature attribution techniques in parallel: SHAP, which draws on cooperative game theory to compute each feature’s average marginal contribution; Random Forest importance, which measures reductions in node impurity across an ensemble of trees; XGBoost feature weights, which blend gain, coverage, and frequency statistics; and LIME, which fits sparse local surrogate models around individual predictions. Because these methods frequently disagree, the researchers aggregate their rankings using a Borda Count voting mechanism, the same centuries-old electoral technique that sums positional votes to find robust consensus. Only features that survive this cross-method gauntlet earn a place in the final ranking.

The results are strikingly consistent. Linguistic semantic coherence, a measure of how logically a teacher’s discourse flows from topic to topic, ranked first across all four explanation methods, with SHAP scores of 0.312 and near-identical rankings from the other three. The student engagement index, derived from gaze patterns and interaction graph dynamics, came second, followed by teacher speech clarity. Perhaps most intriguingly, gesture-speech synchronization ranked fifth, outperforming pure visual motion features and offering quantitative support for embodied cognition theories of teaching. The message from the data is clear: the quality of a lesson lives less in a teacher’s unidirectional delivery than in the structure of bidirectional teacher-student interaction.

The benchmark evaluation was conducted on a publicly released dataset of 236 classroom sessions from six classes in three urban middle schools, spanning STEM, humanities, and language arts instruction. Four synchronized data streams, from an omnidirectional microphone array, 4K cameras, screen recording, and infrared eye tracking, were independently scored by three experts using a modified CLASS scale. Against six baselines ranging from a single-modal LSTM to a transformer ensemble, MFIR-ME led on every metric, cutting mean absolute error by 22.5 percent and root mean square error by 19.8 percent relative to the strongest competitor. All comparisons passed the Wilcoxon rank-sum test at the 0.05 significance level.

Ablation experiments confirmed that every component earns its keep. Removing the linguistic branch caused the largest performance drop, a 9.2 percentage point fall in accuracy, while removing the acoustic, visual, temporal, or behavioral branches each cost between 2.7 and 5.5 points. Deleting the CMAF fusion module or the ranking module likewise degraded performance, demonstrating that the fusion and explanation stages are not decorative. The framework also proved remarkably frugal with data: trained on just 60 percent of the available sessions, it already exceeded the full-data performance of the strongest baseline, and under 10 percent label noise it lost only 1.8 percentage points of accuracy compared with 4.2 for the best competing method.

What elevates the work beyond a machine learning benchmark is its design philosophy of human-AI collaboration. The system operates through a three-stage feedback loop: the AI perceives and ranks multimodal features; teachers review the rankings on an explainability dashboard that translates abstract scores into observable classroom behaviors; and teachers then annotate disagreements, which are fed back as soft labels to retrain the model and adapt it to local teaching norms. Each high-importance feature is operationalized into concrete micro-behaviors. Linguistic coherence maps to discourse transition frequency and hedging language rate; engagement maps to hand-raise latency and post-question wait time, with strategies such as waiting more than three seconds before calling on a student; speech clarity maps to articulation rate, with a target range of 120 to 150 syllables per minute.

The authors are candid about the limitations. The dataset covers only 236 sessions of Mandarin-language junior high instruction in urban Chinese schools, so cross-cultural and cross-stage generalization remains untested. Inference currently takes about 380 milliseconds per sample, too slow for real-time feedback, motivating future work on model distillation and pruning. The gold-standard scores rest on the mean of three expert judgments, which retains some subjectivity, and the study’s cross-sectional design cannot yet show whether teachers who receive the framework’s feedback actually improve over time. The team also flags privacy as a central concern and proposes federated learning, in which schools collaboratively train models without sharing raw classroom video, as a path to large-scale deployment. Still, the core achievement stands: a system that pairs high-accuracy prediction with rankings that teachers can trust and act on, moving educational AI from opaque scoring toward genuine, explainable partnership in the classroom.

Subject of Research: Multimodal explainable AI for classroom teaching quality evaluation in human-AI collaborative education

Article Title: Multimodal feature importance ranking improves classroom teaching quality evaluation in human and artificial intelligence collaborative education

Article References: Feng, Y., Qin, H., Liu, G., & Ye, Q. (2026). Multimodal feature importance ranking improves classroom teaching quality evaluation in human and artificial intelligence collaborative education. Discover Artificial Intelligence, 6(1), Article 1430. https://doi.org/10.1007/s44163-026-02114-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02114-1

Keywords: multimodal learning, explainable AI, classroom teaching quality evaluation, human-AI collaboration, SHAP, Borda Count, cross-modal attention fusion, feature importance ranking, teacher professional development, educational AI, XGBoost, LIME

The problem with existing approaches, according to the research team led by Yuhan Zhang and Jing Zhao, is that conventional image generators operate at the pixel level. They learn what parks look like but not how parks work. Real park design is a decision-making process in which functional spaces act as nodes and spatial relationships act as edges in a network. Designers must balance where to place plazas, sports areas, and ecological zones, how those zones connect to one another, and how pathways, water bodies, and vegetation knit everything together. Standard generative models have no built-in mechanism to enforce that spatial logic, which is why their outputs often contain inconsistencies, such as activity areas stranded far from any path or circulation routes that make no functional sense.

The researchers’ framework works in two sequential stages. In the first, a GraphSAGE model learns the semantic relationships and spatial connections among park functional nodes. Users provide a site condition image and a set of functional requirements, and the model predicts which pairs of functional zones should be connected, generating an explicit relationship graph. Crucially, the team defined park relationships more broadly than simple geometric contact: two zones count as connected if they share a boundary, if they are linked through transitional elements such as roads, plazas, or water bodies, or if they form part of a continuous circulation route from entrances to core public spaces to secondary activity areas. This composite definition captures the layered, dynamic nature of park spatial organization that wall-based architectural models cannot.

In the second stage, the predicted relationship graph is transformed into a spatially structured functional zoning diagram using park design standards and a Voronoi post-processing algorithm, which allocates pixel areas within the irregular site boundary according to each node’s target size. A CycleGAN model then maps that zoning diagram onto a concrete park design scheme, refining visual details and spatial morphology. The team added two custom loss terms, a region semantic consistency loss and an instance semantic consistency loss, to the standard CycleGAN objective, explicitly forcing the generator to preserve regional correspondence among roads, green spaces, and water bodies and to maintain the boundaries of independent functional blocks.

Building the training data was a substantial undertaking in itself. The researchers collected 9,286 urban park design scheme drawings from professional design repositories and online platforms worldwide, focusing on medium-scale comprehensive parks of 10 to 20 hectares, a range chosen because such parks serve district-level populations with complete composite functions while exhibiting similar design logic in spatial granularity and composition. From these, 198 schemes with clearly delineated functional zoning were manually screened and annotated using LabelMe, then augmented to 1,188 graph samples. Functional spaces were consolidated under a blue-green-grey infrastructure classification framework into a three-tier system spanning plazas, sports and fitness areas, ecological green spaces, waterfront spaces, and other major types.

The experiments showed that the graph generation stage performs robustly. As the training set expanded from roughly 100 to 800 samples, the model’s test AUC climbed from 0.8207 to 0.8991, with the steepest gains occurring below 400 samples and performance plateauing beyond 600. When compared head-to-head against GCN, GAT, and GIN architectures on identical data, GraphSAGE achieved the best results, with a validation AUC of 0.909 and an accuracy of 0.846, a finding the researchers attribute to its inductive neighborhood sampling, which suits park graphs where node types and quantities vary and adjacency relationships are sparse. The generated graphs also reflected genuine design principles: open lawns, water bodies, and plazas emerged as highly connected central hubs, while dense forest areas sat at the periphery as ecological buffers, exactly as human designers would organize them.

The full hybrid model’s most striking advantage appeared in planning utility metrics. Compared with a CycleGAN baseline and Stable Diffusion, the full model achieved a circulation connectivity index of 70.2 percent versus 46.7 percent for the baseline and 38.2 percent for Stable Diffusion, along with a functional adjacency compliance rate of 78.4 percent and a node accessibility ratio of 82.5 percent. It also recorded the lowest Fréchet Inception Distance at 0.3190, indicating the closest match to the real distribution of park layouts. Interestingly, the baseline CycleGAN actually scored higher on pixel-level similarity metrics such as PSNR and SSIM, suggesting that the graph constraints deliberately shift the model away from pixel-perfect replication and toward diverse layouts that obey functional logic.

Ablation experiments confirmed that the graph reasoning component carries most of the weight. Removing GraphSAGE caused the largest performance degradation, with FID rising from 0.3190 to 0.3634, while removing the rule-based coordinate assignment or Voronoi allocation produced smaller declines, indicating those components mainly handle geometric translation rather than functional reasoning. Semantic-category analysis showed the largest FID improvements for roads, water bodies, and green space, the elements with strong spatial organization characteristics, while buildings, which occupy small and sparse areas, saw little benefit. A panel of 12 landscape professionals scored the hybrid model’s outputs above 7 out of 10 on all six evaluation dimensions, including spatial structure, park roads, and usability, consistently higher than the baseline, though still below manually designed ground truth.

The researchers are careful to position the tool as an assistant rather than a replacement. The framework is designed for the early conceptual phase of design, where it can rapidly generate layout drafts with basic spatial logic in about 2.12 seconds per 512-by-512 image, allowing designers to explore and compare schemes before committing to detailed development. It embeds the generative model within the knowledge-driven workflow that designers already follow, from requirement analysis through functional zoning to spatial composition, rather than attempting to bypass that reasoning with a purely image-to-image translation.

Limitations remain. The training data concentrates on medium-scale comprehensive parks, so the model is not yet suited to pocket parks or very large green space systems, and the unsupervised CycleGAN stage still produces limited detail in small-scale elements and textures. The team suggests future work could expand the dataset across park types and styles and enrich the graph inputs with terrain elevation, surrounding context, and existing vegetation. But the core insight stands: teaching generative AI to reason about space as a graph before it renders an image produces park designs that are more rational, more usable, and far closer to what professional landscape architects would actually build.

Subject of Research: Graph-constrained generative AI for automated urban park layout design

Article Title: Graph-constrained generative adversarial network for automated urban green space layout generation

Article References: Zhang, Y., Liu, Y., Ni, B., Chen, R., & Zhao, J. (2026). Graph-constrained generative adversarial network for automated urban green space layout generation. Machine Learning with Applications, 26, Article 101020. https://doi.org/10.1016/j.mlwa.2026.101020

Image Credits: AI Generated

DOI: Not provided

Keywords: generative adversarial networks, graph neural networks, urban parks, landscape architecture, CycleGAN, GraphSAGE, generative design, spatial layout generation, urban green space, artificial intelligence, Voronoi diagram, human-AI collaboration

The research team, led by Meng Sheng and Yalin Liu of Changde Vocational Technical College in China, together with colleagues at Lingnan University and Changde First Hospital of Traditional Chinese Medicine, built their intervention around a framework known as cognitive apprenticeship. The idea, borrowed from the learning sciences, is that expert thinking should be made visible to novices in stages: the instructor first models how an expert approaches a problem, then coaches students as they attempt it themselves, and finally fades support as competence grows. The researchers layered a generative AI learning agent onto this scaffold, creating a model they call AI-CAS-PBL, and tested it against conventional problem-based learning and lecture-based instruction in a biochemistry unit on lipid metabolism.

The technical heart of the study is the AI agent itself. The team developed it on the Coze platform, powering it with the Doubao 1.5 Pro 32K large language model, a system with a context window large enough to hold an entire clinical case, the relevant metabolic pathways, and a running dialogue with the student. Rather than simply answering questions, the agent was positioned within the cognitive apprenticeship structure to provide one of two complementary forms of scaffolding: it could respond to student questioning at any hour, offering explanations calibrated to the learner’s current understanding, while the human instructor retained the modeling role, demonstrating expert reasoning in front of the class. The design deliberately avoided the most common failure mode of educational chatbots, which is to hand students finished answers that short-circuit the very thinking the curriculum is meant to build.

Methodologically, the study took care to hold everything constant except the instructional approach. The same instructor taught all three groups, and the learning objectives, core content, scheduled contact time, case materials, and assessment procedures were standardized across conditions. What differed was the mode of support: the AI-CAS-PBL group received the AI agent alongside the cognitive apprenticeship structure, the conventional problem-based learning group worked without AI support, and the lecture-based group received traditional instruction. This quasi-experimental design, while not a randomized trial, substantially strengthens the claim that the AI integration itself, rather than differences in teaching quality or content coverage, drove the observed effects.

The results were striking. Students in the AI-CAS-PBL group achieved significantly higher post-test scores on both lipid metabolism knowledge and case-based clinical application analysis than students in either comparison group, with all differences reaching statistical significance at the conventional threshold. In other words, the AI-supported approach did not merely help students memorize metabolic pathways better; it improved their ability to apply biochemical reasoning to clinical scenarios, which is precisely the transfer of learning that problem-based learning was invented to promote and that lectures so often fail to deliver.

Equally important were the findings on how the experience felt to learners. Using the NASA Task Load Index, a validated instrument originally developed for aviation that measures perceived mental demand, effort, frustration, and related dimensions of workload, the researchers found that students in the AI-CAS-PBL group reported significantly lower cognitive workload than their peers in conventional problem-based learning. This matters because cognitive load theory, one of the most influential frameworks in instructional psychology, holds that working memory is a severely limited resource, and that learning fails when instruction imposes demands that exceed it. The AI agent appears to have absorbed some of the extraneous load, the wasted mental effort of searching for information and struggling alone, freeing capacity for the productive kind of load that builds durable knowledge structures.

The psychological benefits extended beyond workload. Students working with the AI agent reported higher self-efficacy, measured with the New General Self-Efficacy Scale, reflecting greater confidence in their own ability to succeed at the task. Self-efficacy is not a soft outcome; decades of research link it to persistence, engagement, and ultimately achievement. A struggling novice who can test half-formed ideas against a patient, endlessly available AI interlocutor, without fear of embarrassment in front of peers, may be more willing to attempt difficult reasoning, and that willingness compounds over a course. The students also rated the agent highly on perceived usefulness and perceived ease of use, the two central constructs of the Technology Acceptance Model, indicating that the tool was not merely tolerated but genuinely embraced.

To probe the learning environment itself, the team drew on instruments tied to the cognitive apprenticeship experience, including the Maastricht Clinical Teaching Questionnaire and measures of critical thinking disposition. The theoretical claim underlying the intervention is that AI-supported questioning and instructor modeling provide complementary forms of instructional support, a dual-scaffolding arrangement in which the machine handles on-demand explanation while the human demonstrates expert strategy. The authors are appropriately careful here: because the study did not directly examine the interaction between AI support and instructor scaffolding, they present this dual-scaffolding framework as a theoretically informed interpretation rather than a confirmed causal pathway. That honesty is a model of how AI-in-education research should be reported, resisting the temptation to overclaim mechanism from outcome data.

The implications reach well beyond one biochemistry unit. Generative AI is flooding into classrooms faster than the evidence base can support, with institutions adopting chatbots on enthusiasm rather than data. This study offers something rarer: a controlled comparison showing that AI integration, when embedded in a principled pedagogical framework rather than bolted onto an existing course, can outperform both the traditional alternative and the reform method it was meant to enhance. The lesson may be that the framework matters as much as the technology. Cognitive apprenticeship gave the AI a defined pedagogical role, preventing it from becoming either a crutch that replaced thinking or an oracle that students passively consulted.

Caveats remain, and the authors name them. The study was conducted at a single institution with a single instructor, and the quasi-experimental design cannot rule out all confounding. The authors themselves call for evaluation in larger, multicenter, and multi-class studies before the model is widely adopted. Questions about long-term retention, transfer to other subject areas, and the effects of AI dependence on independent problem-solving all await future work. Yet as a proof of concept, the study lands at a propitious moment. It suggests a concrete, testable architecture for human-AI collaborative teaching, one in which the machine’s tireless availability and the teacher’s expert modeling are not competing for the same instructional space but reinforcing each other. If larger trials confirm these results, the image of the medical student lost in a case discussion, silently overwhelmed, may give way to something better: a novice with an expert’s questions at their fingertips and an expert’s thinking demonstrated before them, learning biochemistry not by memorizing pathways but by reasoning through them with less strain and more confidence.

Subject of Research: Integrating generative AI into cognitive apprenticeship-based problem-based learning in medical biochemistry education

Article Title: Integrating generative AI into cognitive apprenticeship–based problem-based learning in biochemistry: a quasi-experimental study

Article References: Sheng, M., Gong, Y., Hu, W., & Liu, Y. (2026). Integrating generative AI into cognitive apprenticeship–based problem-based learning in biochemistry: a quasi-experimental study. BMC Medical Education. https://doi.org/10.1186/s12909-026-10504-3

Image Credits: AI Generated

DOI: 10.1186/s12909-026-10504-3

Keywords: medical education, problem-based learning, generative artificial intelligence, cognitive apprenticeship, cognitive load, biochemistry, large language model, self-efficacy, quasi-experimental study, clinical reasoning, NASA-TLX, human-AI collaboration

Project Sandbox is conceived as a self-contained digital research environment, a kind of sealed laboratory where AI agents can build, compute, prove, and collaborate at the frontier of mathematical knowledge. The domain Hausel has chosen is deliberately ambitious: an area of mathematics that has not yet been formalized by any proof assistant, meaning that no existing software infrastructure exists to verify work in that territory. The project takes as its input a family of open conjectures and aims to develop a new theory capable of proving them, with every step of the argument checked by machine under the direction of a human mathematician. The design is not speculative software vaporware. According to Hausel, the Numina-Lean-Agent, an interactive mathematical proof assistant introduced earlier this year, served as both the inspiration and the practical starting point for the architecture of the sandbox.

The way the sandbox is organized borrows more from the sociology of academic institutions than from conventional software design. Hausel describes it as a digital research institute for AI agents, complete with specialized roles and a defined workflow. A human mathematician, in this case Hausel and his group, sets the research agenda and evaluates which results are actually important, while the AI agents develop the necessary tools, formulate conjectures, attempt proofs, and review one another’s work. This division of labor is central to the project’s philosophy. The machine agents generate volume and speed, exploring many possible avenues of argument in parallel, while the human director supplies the judgment about which avenues matter and which results would constitute genuine mathematical progress rather than formally correct trivia.

What makes the architecture more than an elaborate chatbot pipeline is its verification layer. Project Sandbox is designed so that nothing enters the record until two independent checks have been passed. The first is a computer algebra oracle, a formal tool that can answer a specified class of questions in a single step. In this system, the oracle verifies specific computational claims required by the larger argument, functioning much like a background check on the numerical and algebraic assertions that a proof depends on. The second check is a proof kernel, a relatively simple and independent computer program that acts as a proof checker. Hausel compares it to an autograder for a mathematics exam: it does not assess the elegance of an argument, it does not discover proofs, and it does not judge their significance. It simply verifies that a formal proof is valid within a specified logical system. Together, these two mechanisms mean that the sandbox’s output is not a stream of plausible-sounding AI text but a chain of certified results.

This separation between generation and verification addresses one of the most persistent anxieties about AI in mathematics. Large language models can produce fluent mathematical prose, but fluency is not proof, and the history of the past few years is littered with plausible-looking arguments that collapse under scrutiny. By forcing every statement through an oracle and every proof through a kernel, the sandbox ensures that the AI agents cannot pollute the mathematical record with unverified material, no matter how confident their outputs appear. The human mathematician, meanwhile, retains control over the questions being asked and the interpretation of what has been established. The result is a system in which the weaknesses of machine reasoning are contained and its strengths, chiefly tireless exploration and rapid formal manipulation, are harnessed.

The scientific goal of the project is as concrete as its engineering is careful. Hausel and his group aim to build a practical connection between two areas of mathematics called representation theory and equivariant topology. Representation theory studies abstract algebraic structures by representing their elements as matrices and studying how those matrices act, and it underpins vast swaths of modern mathematics and physics, including the theory of quantum groups and the Langlands classification. Equivariant topology, by contrast, studies geometric spaces equipped with symmetries, and tools such as affine Schubert varieties and GKM graphs allow topologists to extract algebraic information from symmetric geometric objects. The project, in Hausel’s words, will test whether AI can help build a proof-producing bridge between these two rich mathematical worlds, connecting concepts such as quantum groups and Langlands classification on the algebraic side with affine Schubert varieties and GKM graphs on the topological side.

The significance of that bridge lies in what it would make possible. Mathematicians have long known that deep results in algebra and geometry often mirror one another, with the same underlying structures appearing in different guises across fields. When such connections are made rigorous, ideas can flow in both directions: a theorem proved in one language suddenly becomes a theorem in the other, and hard open problems in one field can be translated into more tractable forms in another. Building this transfer framework by hand is slow, painstaking work, because the dictionary between two fields must be constructed statement by statement and verified with total precision. An AI-assisted framework for transferring ideas between algebra and geometry, with machine-checked correctness at every step, could accelerate exactly this kind of dictionary building, which is one reason the fellowship’s reviewers found the proposal compelling.

The project has already moved from concept to working software with surprising speed. When Hausel presented his proposal to Numina in June, he admitted that the image of a mathematical universe in a sandbox on his title slide felt somewhat like science fiction. Yet after only three weeks of work, the sandbox was running on his office computer, producing the visualization that now illustrates the project. The team has also developed an interactive visualization tool, which they call the eyepiece, that allows researchers to watch the sandbox in action, observing how the agents compute and interact within their sealed environment. The longer-term plan is more ambitious still: Hausel says the team eventually intends to adapt the eyepiece for virtual reality and add tools that allow users to interact directly with the sandbox’s computations, turning an internal research instrument into something closer to an explorable mathematical world.

The fellowship behind the project comes from Project Numina, a global nonprofit headquartered in Paris, France, that operates as an open-science collaboration. Its mission is twofold: to develop AI for formal reasoning and to foster human-AI collaboration in mathematics. The organization brings together Olympiad medalists, young mathematicians, and machine-learning engineers to build open-source AI tools and explore new frontiers in mathematical research. Its fellowship program is designed for research groups worldwide that want to accelerate their work with AI tools built for mathematical reasoning, supporting teams in co-developing solutions to deep and ambitious problems at the forefront of science and mathematics. The Numina team summarizes its guiding conviction in a poetic register, holding that mathematics transcends intelligence like an endless ocean only the mind can sail, a sentiment that frames AI not as a replacement for mathematical thought but as a vessel for extending it.

For the community watching AI’s advance into mathematics, the ISTA project offers a template that is neither pure resistance nor pure surrender. Rather than waiting to see whether autonomous AI systems will eventually prove conjectures on their own, Hausel’s sandbox embeds AI agents inside a verification architecture and a human-directed workflow from the start, so that speed and scale are gained without sacrificing the certainty on which mathematics depends. If the bridge between representation theory and equivariant topology can be built this way, with an important result already proved and the research area coming into focus, the sandbox may demonstrate that the most productive future for AI in mathematics is not the lone machine genius of popular imagination but a disciplined collaboration, in which human taste sets the destination and machine-checked reasoning paves the road.

Subject of Research: AI-assisted mathematical proof verification and human-AI collaboration in representation theory and equivariant topology

Article Title: A mathematical universe in a sandbox | ISTA mathematician secures computational support for AI-assisted research

Article References: A mathematical universe in a sandbox | ISTA mathematician secures computational support for AI-assisted research. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: Tamás Hausel, Project Numina, Project Sandbox, proof assistants, AI agents, representation theory, equivariant topology, machine-checked proofs, ISTA, formal mathematics, computer algebra oracle, human-AI collaboration

The paper, titled Personality Pairing Improves Human-AI Collaboration, is the first of its kind to test systematically how AI personalities interact with human personalities in a controlled experimental setting. Most previous research on AI’s benefits gave every participant an identical AI tool, which made it impossible to determine whether differences in performance stemmed from the technology itself or from the fit between the tool and its user. Developers can now shape the personalities of AI agents through prompting, raising the question of whether tailoring those personalities to individual users could unlock real performance gains. Aral and Ju set out to answer that question with a rigor that the field had not yet seen.

To do so, the researchers built a platform called Pairit, designed specifically to randomize human-AI collaborations at scale. More than 1,200 participants from the United States were recruited for the experiment. Each human participant was randomly paired with an AI agent that had been prompted to independently exhibit either high or low levels of the Big Five personality traits, the dominant framework in personality psychology: openness, conscientiousness, extraversion, agreeableness, and neuroticism. This design allowed the team to isolate the causal effect of personality pairings, rather than merely observing correlations, because neither the humans nor the AI configurations were self-selected.

The task itself was deliberately realistic. Each human-AI team was asked to create display advertisements marketing a year-end report from a real think tank, working within a 40-minute session. The collaboration took place through real-time chat alongside synchronized text- and image-editing tools, meaning that the human and the AI had to coordinate genuinely, exchanging ideas and refining materials together rather than the human simply issuing one-off prompts. The researchers first measured the quantity of ads each team produced within the time limit, but quantity alone would say little about whether personality pairing actually improved the work.

To assess quality, a separate group of approximately 1,100 people rated the ads on the quality of their images and text. The ratings revealed that some types of AI personalization yielded markedly better outcomes than others. The most highly rated text and images emerged from three pairings: extraverted humans working with extraverted AI, conscientious humans working with conscientious AI, and open humans working with conscientious AI. In other words, similarity helped in two of the three winning combinations, while a pairing of imaginative, open humans with disciplined, conscientious AI also produced standout creative work.

Just as striking were the combinations that significantly reduced ad quality. Extraverted humans paired with conscientious AI, conscientious humans paired with agreeable AI, and neurotic humans paired with conscientious AI all produced worse results than other configurations. The findings suggest that mismatches between human and machine temperament are not merely neutral; they can actively undermine the collaborative output. When it comes to working with AI, Ju noted, very specific pairings can make a significant difference in terms of performance, a conclusion that challenges the one-size-fits-all approach embedded in most commercial AI tools today.

A crucial strength of the study is that it did not stop at subjective quality ratings. To determine whether the rated quality of the ads translated into real-world value, the researchers ran the ads on the social media platform X for a two-week period, during which the advertisements received approximately five million impressions. They measured both the click-through rate and the cost per click, two of the most closely watched metrics in digital marketing. The results confirmed that the quality ratings predicted actual market performance: better-made ads genuinely performed better in the wild, not just in the eyes of the raters.

The economic implications of that real-world test were substantial. The researchers found that higher text quality increased the click-through rate by about 7 percent and decreased the cost per click by roughly 30 cents. In an industry where advertising budgets run into the billions and marginal improvements in click efficiency compound across millions of impressions, those figures point to a significant opportunity. Aral observed that participants had reported the better-performing ads were higher quality overall, with stronger images and text, and that the team was then able to see that play out in the real world when the ads performed better in the advertising market. The implication is that training AI to produce better text and images, calibrated to the right human collaborators, can yield substantial marketing results.

The study also carries a broader conceptual message. Research on human teams has consistently shown the advantages of bringing certain complementary personalities and abilities together, and Aral argues that personalized AI can be implemented in much the same way: by matching the right type of personalization of the AI agent to the individual, performance on the task can be maximized. This reframes AI personalization not as a cosmetic feature, such as a chatbot adopting a friendly tone, but as a structural lever on team composition. Just as a manager would not staff a project team at random, the findings suggest that organizations should think deliberately about which AI temperament to assign to which employee, and that the wrong assignment may be worse than no assignment at all.

The authors caution that their results motivate further research into the complex implications of AI personalization for collaboration, teamwork, and performance, and they are already planning follow-up studies in other settings and industries, including sales, engineering, and healthcare. Ju expressed confidence that everyone can benefit from a high level of personalization in their own agent use by understanding the types of personality pairings that work best across settings and industries. If the pattern holds beyond advertising, the era of identical AI assistants for every user may be drawing to a close, replaced by systems that are matched to the people they work with as carefully as any human colleague would be. Aral and Ju are co-founders of Pairium AI, a technology company developing AI personalization tools to optimize human-AI collaboration, meaning the science and the commercial application of these findings are advancing in parallel.

Subject of Research: The effect of matching human and AI personality traits on human-AI collaboration and performance

Article Title: Personality pairing improves human-AI collaboration

Article References: Personality pairing improves human-AI collaboration. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: human-AI collaboration, AI personalization, Big Five personality traits, randomized controlled trial, MIT Sloan, Sinan Aral, Harang Ju, PNAS, advertising performance, click-through rate, AI agents, team performance

The core problem the researchers set out to solve is what they describe as two-sided “goal–action ambiguity.” Conventional AI systems infer human intent from externally observable information — speech, gestures, actions — but the same observable behavior can reflect very different goals, and the same goal can be pursued through very different actions. When a person picks up a cup, an observer cannot know whether they intend to drink from it or to hand it to someone else. Conversely, if the goal is simply to quench a thirst, that goal could be satisfied by reaching for the cup, grabbing a water bottle, or asking another person to bring a drink. Because of this ambiguity, when an AI misreads a human’s intent, the user is typically forced to correct it through additional commands or demonstrations, a slow and often frustrating loop that undermines the promise of natural collaboration between people and machines.

The KAIST and MSRA team’s insight was to bypass observable behavior entirely and tap into the brain’s own error-detection machinery. When a person encounters an unexpected situation, the brain produces rapid, unconscious “prediction error” signals — an instantaneous internal alarm that something does not match expectations. In the context of human–AI interaction, these signals amount to the brain’s immediate “that’s not what I meant” response when an AI performs the wrong action or pursues the wrong goal. Crucially, the researchers found that these signals are not monolithic. They distinguished two distinct types of brain responses, each carrying different information about what exactly went wrong in the collaboration.

The first type is reward prediction error, or RPE, which appears when the AI misunderstands the person’s ultimate goal — when the machine has grasped the wrong endpoint entirely. The second is state prediction error, or SPE, which appears when the goal is correct but the process or method of action differs from what the person expected. This distinction matters enormously for how an AI should respond. If the brain is broadcasting an SPE signal, the appropriate correction is to keep the goal and change the strategy. If it is broadcasting an RPE signal, the AI must discard its assumption about what the person wants and search again for the true objective. By separating these two error classes, the technology gives AI systems a structured way to diagnose the nature of their own misunderstanding rather than simply registering that something is amiss.

To capture these signals, the team measured real-time electroencephalography (EEG) from people as they observed AI systems performing tasks. The recordings revealed that the brain produced measurably different patterns depending on whether the AI had misunderstood the goal itself or had chosen the wrong method while pursuing the correct goal. The researchers also identified distinctive brainwave signatures that emerged when both types of errors occurred simultaneously — a doubly ambiguous situation in which neither the destination nor the route matched the person’s expectations. These findings establish, at the level of neural measurement, that the human brain encodes goal-level and method-level disagreement with an AI as separable, decodable phenomena.

The decoding itself was accomplished with deep learning. By training neural networks on the recorded EEG signals, the team developed a system that can determine, from brain activity alone, how a person is interpreting the AI’s behavior at any given moment. In practical terms, even without a person saying “that’s wrong,” the AI can recognize whether the human brain is signaling that “the goal is wrong” or “the method is wrong.” The team then wrapped this decoding capability into a Neural Value Alignment–based human–AI synergy algorithm, which feeds the decoded brain signals back to the AI in real time. When the AI detects an SPE signal, it interprets the situation as “the desired goal is correct, but the method is wrong” and adjusts its action strategy accordingly. When it detects an RPE signal, it understands that “the goal itself was misunderstood” and re-searches for what the person truly intended. In both cases, the correction loop closes without a single spoken word.

Simulation results offered early evidence that the approach works under realistic stress. The proposed method adapted more quickly than existing approaches even in uncertain conditions, such as when a person’s goal changed suddenly mid-task or when some of the human neural feedback was missing — a scenario that reflects the noisy, incomplete data typical of real-world brain–computer interfaces. This robustness matters because laboratory EEG is far cleaner than what wearable sensors will deliver in homes, factories, or vehicles. The researchers emphasize that the significance of the work lies in demonstrating that AI systems can correct themselves by reading a person’s unconscious “that’s not what I meant” brain response, without requiring the user to repeatedly say “do it this way” or “that’s not right.”

The potential applications stretch across the emerging landscape of embodied and interactive AI. As the technology matures, the researchers say, it could be applied to physical AI robots in homes and industrial settings, allowing them to understand user intent more naturally and adjust their actions accordingly. Autonomous vehicles could quickly reflect driver judgment in situations where a split-second correction matters. Medical and rehabilitation robots could serve patients who have difficulty speaking or moving, for whom conventional command-based interaction is impractical and whose needs are most acute. Educational AI systems could adapt to a student’s cognitive state in real time, sensing confusion or disagreement that the student never articulates. In each case, the common thread is a tighter, faster coupling between human judgment and machine behavior than explicit interfaces allow.

Professor Sang Wan Lee of KAIST’s Department of Brain and Cognitive Sciences, who directs the Center for Neuroscience-Inspired Artificial Intelligence and led the international collaboration, framed the work as a shift in the source of intent information. “This research is meaningful because it shows that AI can move beyond inferring human intent only from visible behavioral outcomes and instead directly use cognitive signals generated in the brain during collaboration with AI,” he said. He added that the technology can be expanded to a wide range of fields where human judgment and AI behavior must be closely connected, including physical AI, brain–computer interfaces, autonomous driving, precision personalized education, medical robotics, and human–computer interaction. From MSRA, Miran Lee, Director of the Microsoft Research Accelerator, described the achievement as the result of the ongoing international collaboration between KAIST and Microsoft Research Asia and expressed the intention to continue the partnership to develop world-class BCI technologies that enable humans and AI to communicate and collaborate more naturally.

The study’s first author is Xin Xu, a Ph.D. student in KAIST’s Department of Brain and Cognitive Sciences; researchers from Microsoft Research Asia, including Yansen Wang, Dongqi Han, and Dongsheng Li, also participated. The paper, titled “Neural Value Alignment: Human–AI Collaboration Under Goal–Action Ambiguity,” appeared in IEEE Transactions on Cybernetics under DOI 10.1109/TCYB.2026.3722605. The work builds on a broader research program: a companion KAIST–MSRA study on helping AI adapt rapidly to continuously changing environments, led by first author Niklas Koeppe, was presented in June at ICML 2026 under the title “Mitigating Plasticity Loss through Architectural Design in Continual Learning.” The research was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP), funded by the Ministry of Science and ICT. Taken together, the results sketch a future in which the most important interface between humans and machines is not the keyboard, the touchscreen, or even the voice assistant, but the silent, instantaneous feedback of the brain itself — a future in which AI finally hears the corrections we never say out loud.

Subject of Research: Brainwave-based AI technology that detects human–AI cognitive mismatch and realigns AI behavior with human intent

Article Title: KAIST develops AI that can sense “that’s not what I meant” without being told

Article References: KAIST develops AI that can sense “that’s not what I meant” without being told. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: KAIST, Microsoft Research Asia, Neural Value Alignment, brain–computer interface, EEG, reward prediction error, state prediction error, human–AI collaboration, deep learning, intent inference, IEEE Transactions on Cybernetics, physical AI

The research team, led by Yanyan Zhang and Chunxiu Xiao of Fujian Health College with colleagues from Fujian Medical University, framed their work within the constructivist paradigm, the view that learners actively build knowledge from experience rather than passively receive it. Participants were recruited through purposeful maximum variation sampling, a technique designed to capture as wide a range of backgrounds, skill levels, and attitudes toward AI as possible. Rather than relying on questionnaires alone, the researchers triangulated three data streams: semi-structured interviews, reflective diaries kept by the students themselves, and their actual coursework. This combination allowed the team to compare what students said about their AI use with what they demonstrably did with it.

Methodologically, the study is notable for its integrated analytical framework. The researchers anchored their analysis in Kolb’s experiential learning theory, a four-stage cycle in which concrete experience feeds reflective observation, which in turn supports abstract conceptualization and active experimentation. They layered onto this the Information Seeking and Communication Model, or ISCM, a framework from information science that describes how people search for, evaluate, and use information in task contexts. By combining the two, the team could track not only how students learned from their encounters with AI but also how their underlying information behaviors, from simple retrieval to verification and synthesis, changed along the way. The study was approved by the Ethics Committee of Fujian Health College and conducted in accordance with the Declaration of Helsinki, with written informed consent from all participants.

The analysis revealed four distinct developmental stages. The first, described as exploration and uncertainty, captures the disorientation of initial contact. Students in this phase approached AIGC tools with a mixture of curiosity and anxiety, unsure what the tools could do, what they were permitted to do, and whether the output could be trusted. Information behavior at this stage was dominated by retrieval: students typed questions, copied answers, and moved on, with little systematic checking of accuracy. The researchers characterize this as a period of trial without much error correction, because students often lacked the domain knowledge to recognize when the AI was wrong.

The second stage, reflection and boundary recognition after trial and error, marks a turning point. As students accumulated concrete experiences, including moments when generated content proved inaccurate, generic, or misaligned with clinical reality, they began to reflect on the limits of the technology. This is where Kolb’s cycle becomes visible in the data: reflective observation of failed or flawed outputs led students to form abstract concepts about where AI could help and where it could not. Boundary recognition, in the researchers’ terms, meant students developed a working map of the technology’s capabilities and its hazards, particularly around hallucinated facts and the superficial plausibility of AI-written clinical text.

The third stage, strategy construction and systematic application, represents the most technically interesting transformation. Students stopped treating the AI as an answer machine and started treating it as a tool to be orchestrated. They reported developing prompting strategies, cross-checking generated content against textbooks and clinical guidelines, and integrating AI output into their own writing and reasoning rather than substituting it. In ISCM terms, their information behavior shifted from primarily retrieval-focused use toward generative, verificatory, and integrative patterns: they used AI to produce drafts, then verified those drafts against authoritative sources, and finally integrated the surviving material into their own work. This progression mirrors what information literacy researchers have long described as the difference between searching and thinking.

The fourth stage, professional repositioning in human-AI collaboration, is perhaps the most consequential for nursing education. Students began to articulate a changed sense of their own professional role: not replaced by AI, but repositioned relative to it, with responsibility for final judgment resting firmly on the human practitioner. The authors report that students perceived a deepening of critical judgment and professional role awareness as their experience matured, although the study is careful to note that these patterns varied considerably across participants. Not every student reached the later stages, and the trajectory was not uniform; individual differences in prior digital literacy, confidence, and reflective habits shaped how far and how fast each student progressed.

That variability is central to the study’s conclusion. The authors argue that the potential relevance of AIGC experience to nursing decision-making-related development depends on students’ evaluation, monitoring, and responsibility. In other words, the benefit is not automatic. A student who copies AI output uncritically may bypass exactly the cognitive work, weighing evidence, checking sources, taking responsibility for a recommendation, that clinical decision-making requires. A student who verifies, monitors, and owns the final answer may strengthen those same capacities. The technology is the same in both cases; the developmental outcome diverges because of how the learner engages with it.

The practical implication the researchers draw is that nursing education should deliberately combine experiential learning with critical information literacy cultivation. On the experiential side, Kolb’s framework suggests structured cycles in which students use AI, reflect on what worked and what failed, form principles, and test them again. On the literacy side, the ISCM framework suggests explicit instruction in seeking, evaluating, and communicating information: teaching students not just how to prompt a model but how to verify its output against clinical guidelines, recognize the limits of generated content, and document their reasoning. The study’s findings, drawn from interviews, diaries, and coursework rather than controlled outcome measures, are suggestive rather than definitive, and the authors frame their conclusions in correspondingly cautious terms, describing perceived deepening of judgment and potential relevance rather than proven causation.

Still, the study offers one of the clearest temporal pictures to date of how future nurses actually come to terms with generative AI. It suggests that the first weeks of AI use, dominated by uncertainty and uncritical retrieval, are not a failure state but a predictable early stage of a longer developmental arc, one that educators can shorten and enrich through structured reflection and explicit verification training. As generative AI becomes embedded in clinical documentation, patient education, and decision support, the difference between a nurse who learned to interrogate AI and one who learned only to use it may matter far beyond the classroom. This research indicates that the interrogation skills can be taught, and that the earlier they are built into the curriculum, the more likely students are to emerge as responsible, critically engaged human-AI collaborators rather than passive consumers of machine-generated answers.

Subject of Research: Longitudinal qualitative study of nursing students' developmental experience with generative artificial intelligence in education

Article Title: Early developmental trajectory of nursing students’ AIGC application experience: a short-term longitudinal qualitative study

Article References: Zhang, Y., Cai, L., Liu, F., & Xiao, C. (2026). Early developmental trajectory of nursing students’ AIGC application experience: a short-term longitudinal qualitative study. BMC Medical Education. https://doi.org/10.1186/s12909-026-10562-7

Image Credits: AI Generated

DOI: 10.1186/s12909-026-10562-7

Keywords: generative AI, nursing education, AIGC, experiential learning, information literacy, qualitative research, longitudinal study, clinical decision-making, human-AI collaboration, Kolb experiential learning theory, ISCM, vocational education

One of the most striking threads in the collection concerns the way human expertise quietly moderates algorithmic decisions in high-stakes environments. Carboni and colleagues examine clinical settings where violence is a real possibility, showing how nurses intervene in and influence algorithmic outputs. The researchers highlight the importance of the doubt that nurses introduce into the decision-making process, treating this cultivated skepticism not as friction to be eliminated but as a vital contribution to collective judgment. In a related vein, Van Voorst documents what is called collaborative tinkering between humans and algorithms in the context of public health, arriving at conclusions that largely echo the clinical analysis. Both studies suggest that the fantasy of fully automated decision-making dissolves under close empirical scrutiny, replaced by a messier but more accurate account of humans and machines reasoning together.

The volume also confronts the hidden human labor on which algorithms depend. Cabitza and colleagues coin the neologism cybork, a fusion of cyborg and work, to emphasize that algorithms frequently rest on human effort. Bennett and colleagues explore the social and power dynamics that arise when humans clean data so that algorithms can operate more efficiently and more correctly. Their argument, aligned with Cabitza and colleagues, is that this labor should not be occluded as if the algorithms were completely autonomous. Instead, algorithmic processing should be understood as a form of collective networking and human-machine collaboration. This reframing carries real political weight: when the workers who make machine learning possible remain invisible, they also remain uncredited, unprotected, and excluded from conversations about how the resulting systems should be governed.

Entanglement, however, runs in both directions. Several essays trace the recursive feedback loops through which algorithms influence human behavior at the same time that humans shape algorithms. Lynch and colleagues demonstrate this dynamic at a microscale, examining how the introduction of robot tour guides into two Nevada museums helps to create the distinctive social spaces that visitors encounter. The robots are not merely props; they actively reorganize how people move, gather, and interact. Kalthoff and Link extend the analysis to the fabrication of humanoid robots designed to create social dynamics through their interactions with humans, citing the pioneering Kismet robot built by Cynthia Breazeal, an early attempt to engineer machines that could engage humans on an emotional register.

At the largest scale, Richard Groß argues that Large Language Models, through their stochastic procedures, are shaping social reality in their own image. A parallel argument has been advanced by Louise Amoore and her colleagues, who focus on the principal technical characteristics of LLMs, such as latency and tokenization, and contend that these systems create a world model that is rapidly propagating through the infrastructures of developed societies. The claim is technically grounded rather than metaphorical. Because LLMs mediate communication, search, and increasingly administrative processes, the statistical regularities embedded in their training and generation procedures can become the regularities of everyday social life, closing a loop in which human output trains the machines that then shape human output.

Beyond empirical case studies, the collection presses for a change in the frameworks, metaphors, and narratives that construct AI-human relations. Harry Halpin urges readers to think in terms of collective intelligence rather than artificial intelligence, a shift that relocates smartness from a single machine to a network of people and tools. Lewis and colleagues point toward indigenous knowledges as demonstrating what they call abundant intelligences that extend throughout the natural world. Hayles singles out this lesson as a general template for how communities can assume control over the kinds of AI that get created, the purposes those systems are designed to fulfill, and the human interests they can support and help to enact. The implication is that the trajectory of AI is not fixed by the technology itself but by the social arrangements that surround it.

Other contributors extend this critical reframing in complementary directions. Bennett and colleagues call for Responsible Artificial Intelligence, in which alignment with human values is treated as the primary design criterion rather than an afterthought. Bory and colleagues look critically at AI narratives themselves, arguing that visions with more modest ambitions, the so-called weak rather than strong versions of AI, hold more potential to lead to futures beneficial to humans. Nugent critiques the pervasive analogy between biological and technological evolution, showing that the analogy works only by suppressing crucial differences that reveal the complexities of biological processes. Together these essays suggest that the stories we tell about AI are not decoration; they actively determine which futures seem possible and which are dismissed before they are tried.

Perhaps the most technically provocative contribution comes from Fabian Offert and colleagues, whose essay on protein folding as predicted by AlphaFold and other Transformer technologies deconstructs the widespread belief that LLMs produce natural language. As the authors show, through tokenization, word embedding, and positional embedding, the linguistic qualities of language are taken apart, reconstructed as numbers in a high-dimensional vector space, and then reconstituted through stochastic procedures that differ radically from how natural language is used and understood by humans. What emerges is not language in any human sense but a new episteme, one that creates and reinforces an intuitive knowledge of sequence. That intuition about sequential structure, they suggest, is precisely what makes these architectures so powerful for predicting how chains of amino acids fold into functional proteins, a problem that had resisted solution for half a century.

The significance of this analysis extends well beyond molecular biology. Hayles notes that it resonates with her own criticism of the naturalness of machine language, and she argues that analyses like those of Offert and colleagues open new possibilities for literary criticism, cultural critique, and critical code studies. If the outputs of LLMs are the products of vector-space arithmetic and probabilistic sampling rather than meaning in the human sense, then humanists have a crucial role in explaining what these systems actually do, what they cannot do, and where their apparent fluency misleads the people who rely on it. The special issue thus positions the humanities not as a skeptical bystander to the AI revolution but as an essential participant in its technical and conceptual development.

The overall message of the collection is both a warning and an invitation. Hayles concludes that readers will find a plethora of diverse approaches, specific demonstrations, and game-changing ideas in these essays, all pointing to the ongoing necessity for critique and analysis by humanists, cultural critics, and others as powerful technologies develop at breakneck speed. We cannot afford to ignore AI, she writes, and the essays help chart paths that allow us to understand these systems, participate in their development, and have a say in how they are and will be used. In an era when headlines oscillate between utopian hype and existential dread, the volume offers a third way: careful, empirical attention to the actual entanglements of humans and machines, and a insistence that the shape of that entanglement remains a collective human choice.

Subject of Research: Human-AI entanglement and the mutual shaping of humans and algorithms in society

Article Title: Entangling humans and AI

Article References: Hayles, N. K. (2026). Entangling humans and AI. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03387-x

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03387-x

Keywords: artificial intelligence, human-AI collaboration, large language models, collective intelligence, algorithmic decision-making, data labor, responsible AI, protein folding, AlphaFold, AI narratives, posthumanism, AI & Society

The study, led by Bingjie Huang of Central China Normal University together with colleagues at Southeast University and Chengdu No. 7 High School, recruited 116 sophomore STEM students enrolled in a calculus-based introductory physics course and randomly assigned them to two groups. Ninety-five students completed all procedures: 50 in the experimental group and 45 in the control group. The experimental group received a structured training session built on Zimmerman’s self-regulated learning framework and cognitive load theory, while the control group received an equally long session on advanced problem-solving strategies with no AI guidance at all. Crucially, both groups then faced the same test: solving items from the Conceptual Survey of Electricity and Magnetism, a standardized 32-question assessment, and then revising their answers after conversing with the Chinese large language model Doubao.

The training itself was organized into three sequential phases. A ten-minute situational introduction contextualized generative AI, contrasting it with the homework-assisted apps that dominate Chinese students’ digital lives. A fifteen-minute strategy-teaching segment demonstrated how to prompt the AI in ways aligned with the three phases of self-regulated learning: forethought, performance, and self-reflection. It also introduced the model’s characteristic weaknesses, including reasoning errors, visual misinterpretation of graphs and figures, and outright hallucinations, along with practical remediation tactics. Finally, a twenty-minute hands-on session let students practice analyzing two mechanics problems with the AI and reflect on the obstacles they encountered. The design deliberately separated the training domain, mechanics, from the assessment domain, electromagnetism, so that any measured benefit would reflect transferable strategy rather than memorized solutions.

The quantitative results were striking. Students who received the AI training achieved significantly higher revision performance than the control group, with an effect size of 0.81, a large effect by conventional standards in educational research. The most dramatic behavioral difference appeared in the quality of revisions. Among trained students, 69.3 percent of revisions converted an incorrect answer into a correct one. In the control group, that figure was only 41.3 percent. Worse, 40 percent of the untrained students’ revisions actually changed correct answers into incorrect ones, a pattern the authors attribute to blind reliance: untrained students treated AI outputs as authoritative and capitulated reactively, even when the model’s reasoning was internally inconsistent.

Chat transcripts revealed exactly how the two groups diverged. The researchers coded six interaction strategies, ranging from the low-level image-only approach, in which students simply upload a photo of the problem, to higher-order behaviors such as asking for definitions, seeking insight into the structure of a problem, and generating new questions for self-assessment. Both groups leaned heavily on image-only prompting, a habit the authors link to the snap-and-solve culture fostered by apps like Zuoyebang and Xiaoyuan, which let students photograph homework and receive instant answers. But the trained group diversified markedly. The proportion of students using the obtaining-insight strategy more than doubled, from 20 percent to 42 percent, and 14 percent of trained students progressed to generating new questions, a behavior entirely absent among controls. Trained students also exchanged significantly more conversational turns with the AI, indicating deeper engagement rather than quick answer extraction.

The qualitative cases are perhaps the most vivid evidence of what the training accomplished. In one exchange, the AI misjudged an electromagnetic induction question, declaring that a coil in one figure formed an unclosed circuit. A trained student pushed back, noting that the coil was in fact closed, which prompted the model to recognize that a contracting closed coil changes its area and therefore its magnetic flux, leading it to the correct answer. Another student caught the AI misreading a graph in a mutual inductance problem and walked it to the right option by inspecting the answer patterns. By contrast, one control student, after the AI correctly graded an item, instructed it to simply show correct answers without analysis going forward, and then failed to catch a visual error the model made on the very next question.

Cognitive load measurements added a theoretically intriguing wrinkle. Overall cognitive load did not differ between the groups, but the trained students reported a marginally higher intrinsic load, the component tied to the inherent difficulty of the task itself. Rather than interpreting this as a cost, the authors propose that it may reflect productive struggle: the heightened working-memory demands that come from systematically decomposing problems and interrogating the AI strategically. They suggest refining cognitive load theory to distinguish task-imposed intrinsic load from productively engaged intrinsic load, arguing that when extra mental effort signals deep metacognitive processing rather than confusion, it may accompany, rather than undermine, better learning outcomes.

The study also probed how three learner characteristics, self-regulated learning skills, prior generative AI experience, and domain knowledge, shaped outcomes. Using robust linear models suited to the skewed distribution of pretest scores, the team found that self-regulated learning skills and prior knowledge both significantly predicted revision performance, while prior AI experience showed only a marginal effect. Notably, the training moderated the role of AI experience: in the control group, prior chatbot experience strongly predicted revision success, but in the trained group that relationship flattened almost to zero. In other words, structured instruction appeared to level the playing field, giving students without extensive AI backgrounds the strategic tools that experienced users had developed informally, if at all.

The authors are careful about the limits of their findings. The assessment was a one-shot revision task conducted on a single day, so the design cannot establish whether the gains reflect durable learning, conceptual change, or simply a temporary surge of attention. Because pretest and posttest used identical items, memory effects cannot be ruled out, and the observed advantage may partly stem from the control group’s lack of any AI guidance rather than from specific training components. The sample, drawn from a single top-tier Chinese university, also limits generalizability. The researchers call for longitudinal studies with delayed posttests, component-level comparisons to isolate the active ingredients of training, and richer measures such as interviews, eye tracking, or electroencephalography to capture students’ cognitive experiences more fully.

Even with those caveats, the implications are hard to ignore. As generative AI becomes embedded in education worldwide, the debate has often been framed as a binary between banning the tools and handing them out. This trial points toward a third path: teaching students explicitly how to collaborate with AI, including how to recognize its failure modes and correct them. The authors propose complementarity recognition, an awareness of the asymmetric capabilities of humans and machines and of when to trust each, as a construct worth measuring and cultivating. A 45-minute session cannot remake a student’s intellectual habits, but it demonstrably shifted behavior in the direction educators want: fewer blind capitulations to the machine, more probing questions, and a strikingly higher rate of turning wrong answers into right ones. The answer machine, it turns out, becomes a genuine partner only when someone teaches students how to work with it.

Subject of Research: The effect of structured generative AI training on students' AI-assisted scientific problem solving in STEM education

Article Title: From answer machine to collaborative partner: impact of structured generative AI training on scientific problem solving in STEM education

Article References: Huang, B., Xie, L., Liu, Y., Liu, X., Tang, H., Feng, X., Qiao, C., Guo, Q., Wu, C., & Bao, L. (2026). From answer machine to collaborative partner: impact of structured generative AI training on scientific problem solving in STEM education. International Journal of STEM Education, 13(1), Article 62. https://doi.org/10.1186/s40594-026-00652-9

Image Credits: AI Generated

DOI: 10.1186/s40594-026-00652-9

Keywords: generative AI, STEM education, self-regulated learning, cognitive load theory, human-AI collaboration, problem solving, metacognitive laziness, physics education, randomized controlled trial, AI literacy, prompting strategies, electromagnetism