During Week 39, the central developmental question became increasingly visible:

What exactly is developing?

The artificial system?

The Human-AI workflow?

The person’s ability to operate AI?

Or human cognition itself?

Week 40 adds another layer. As the surrounding research landscape becomes more sophisticated, the question is no longer only whether cognition changes through interaction with artificial intelligence.

A more structural question is emerging:

Where and how is cognitive structure being formed?

Is structure supplied by the artificial system?

Is it distributed across the Human-AI arrangement?

Is it temporarily available through scaffolding?

Is it encoded inside a machine architecture?

Or is the human developing an increasing capacity to form, test, revise, preserve, and continue cognitive structure themselves?

This distinction matters because structure around a human is not automatically structure developed within the human.

Several works retained during this week’s research observation illuminate different parts of that distinction. They belong to different disciplines and should not be treated as one framework.

Together, however, they reveal a research landscape increasingly concerned not merely with intelligence or output, but with the architecture through which reasoning, learning, grounding, alignment, and safety are produced.

Structured Interaction Can Improve Learning

A study published in Neuron, Scaffolding human and AI instruction: Neural alignment and learning gains in online education, examined whether a brief structured interaction before an online lecture could affect subsequent learning.

The researchers compared 57 students across three conditions: no prior interaction, interaction with a human instructor, and interaction with an AI instructor.

They combined behavioural testing, eye tracking, and fMRI.

Both the human and AI interaction conditions were associated with improved learning and greater neural alignment during subsequent learning. The AI condition produced learning outcomes statistically comparable to the human-instructor condition on the measures examined, although human interaction produced greater social closeness and some stronger gaze alignment. 

The study is important because it moves beyond asking whether students enjoy AI or whether AI can produce educational content.

It examines what a structured interaction does to the learning process.

From the perspective of Human-AI Cognitive Development, however, another question follows.

The study demonstrates that structure provided before learning can influence learning. It does not by itself establish that the learner has developed the capacity to create that structure.

That distinction is fundamental.

Receiving structure is not the same as developing the capacity to create structure.

A scaffold can be valuable.

An instructor can provide it.

An AI can provide it.

A learning environment can provide it.

But Structure-First Cognitive Development introduces another developmental test:

What happens when the scaffold disappears?

Can the learner identify what structure is needed?

Can they construct it?

Can they test whether it works?

Can they modify it when it fails?

Can they reproduce the reasoning later without receiving the same external organization?

Structured interaction can support development.

But:

Structured interaction is not automatically Structure-First Cognitive Development.

Grounding and Reasoning Are Being Separated on the Artificial Side

A different structural question appears in DISCO: Distributed Long Context Scaling with Grounding-Reasoning Disaggregation.

DISCO addresses the deterioration of reasoning across very long contexts. Rather than asking one model to search enormous amounts of contextual material and reason over it simultaneously, the architecture separates these functions. Worker models perform localized grounding across partitions of the source material. A central model then reasons over the collected evidence.

The authors describe the problem as a structural entanglement between the burden of grounding and the capacity needed for reasoning. Their architecture therefore separates the two. 

This research concerns machine architecture, not Human-AI Cognitive Development.

The distinction should remain clear.

But it provides a useful technical parallel.

A large amount of available information does not automatically produce good reasoning.

The organization of the reasoning process matters.

Grounding matters.

Separation of functions matters.

Relations between stages matter.

This echoes a broader structural principle without making the two domains equivalent:

Having information available is not the same as having a structure through which the information can be reasoned over.

For machines, that may require architectural separation.

For humans, it may require the ability to distinguish sources, possibilities, relations, evidence, judgment, and continuation.

The mechanisms are different. The structural question is shared.

When Human Cognitive Signals Become Inputs to AI

Another development moves the boundary in a different direction.

In Neural Value Alignment: Human-AI Collaboration Under Goal-Action Ambiguity, researchers propose using signals associated with human reinforcement learning to improve AI alignment under situations in which a person’s goal cannot easily be inferred from behaviour alone.

The researchers used EEG recordings and demonstrated that reward-prediction and state-prediction error signals could be decoded from cortical activity. Their simulations then showed how combining those signals could improve value-alignment processes under imperfect decoding. The technical contribution is important.

But from the perspective of Human-AI Cognitive Development and Cognitive AI Safety, it raises another kind of question. Historically, Human-AI interaction has depended heavily on deliberate external communication.

A person types.

A person speaks.

A person presses a button.

A person uploads information.

A person explicitly gives the artificial system something to interpret.

Neural inference changes the boundary. If cognitive or neurophysiological signals can increasingly inform AI behaviour, then information originating inside the human cognitive system may become part of the artificial system’s interpretive environment.

That does not make such research inherently harmful. Nor does this study itself establish a privacy violation.

But it creates a boundary question that cannot be answered by technical accuracy alone:

An AI becoming better at detecting what a human thinks does not automatically establish what an AI should be permitted to know about human thought.

Accuracy and permission are different questions.

Inference and consent are different questions.

Technical possibility and legitimate cognitive access are different questions.

As Human-AI systems become more cognitively intimate, preservation of the boundary around the human Source becomes increasingly important.

Explicit Structure Is Becoming a Reliability Question in Machine Reasoning

A September review in Knowledge and Information Systems examines another branch of structural reasoning: the use of knowledge graphs to augment large language models.

The authors review approaches in which explicit representations of entities, relationships, and constraints are used to address problems including hallucination, weak reasoning, and opacity.

Their analysis identifies persistent difficulties in integrating the statistical representations of language models with more explicit relational knowledge structures.

Again, this is not Human-AI Cognitive Development.

Knowledge graphs have a long independent history, and their use in machine reasoning should not be absorbed into a human cognitive framework. Their relevance here is narrower.

They reinforce an important observation:

Structure changes what can reliably be reasoned over.

Data alone is not enough. Information alone is not enough. Access alone is not enough. The relations through which information becomes interpretable matter.

This applies differently to artificial and human cognition, but it makes one error increasingly difficult to sustain:

intelligence cannot always be understood simply as more information plus more capability.

Architecture matters.

AI Safety Is Beginning to Include the Human System

A working paper by Jace Kim, The Myth of Perfect Users and Perfect Systems: A Cross-Disciplinary Framework for Human-Centered AI Safety and Resilience, approaches a different layer of the Human-AI relation.

Rather than treating safety exclusively as a property of the model, the paper draws together human factors, cognitive psychology, organizational research, safety engineering, and resilience.

Its central argument is modest but important: failures in complex Human-AI systems can arise through interactions among humans, technical systems, organizations, and information environments rather than from a single defective component. 

This is an independent working paper rather than evidence carrying the same weight as the peer-reviewed Neuron, IEEE, or Springer works discussed above. I therefore treat it as a landscape signal rather than a primary evidentiary source. But the direction is worth noticing.

AI safety is gradually widening from: Is the model safe?

toward: What happens when an artificial system operates inside a variable human environment?

Human-AI Cognitive Development introduces another layer again:

What happens to the human cognitive system through repeated participation in that environment?

Structure Is Not One Thing

The word structure can become misleading if every organized system is treated as equivalent. Structure may exist in:

an artificial reasoning architecture,

a knowledge representation,

a learning scaffold,

a Human-AI workflow,

a conversation,

a cognitive environment,

or the reasoning architecture of the human themselves.

These are not interchangeable.

An AI may provide structure.

A system may enforce structure.

A teacher may scaffold structure.

A workflow may distribute structure.

A Human-AI arrangement may exhibit sophisticated relational structure.

And yet the human may remain dependent upon that external organization.

Structure-First Cognitive Development therefore requires another distinction:

The existence of structure is not sufficient evidence of structural cognitive development.

The developmental question is where the capacity to form structure increasingly resides.

Scaffolding Is Not Internalization

Scaffolding is one of the clearest examples. Good scaffolding can help a learner reach something they could not yet reach alone. That is precisely why scaffolding is valuable. But successful scaffolding contains a developmental paradox.

If the scaffold remains permanently necessary, then the learner may become more successful inside the scaffold without becoming increasingly capable without it.

This does not invalidate scaffolding. It tells us what developmental evidence must eventually be sought.

Can the learner reproduce the organizing process?

Can they transfer it?

Can they construct an equivalent structure in a new domain?

Can they identify when a different structure is required?

Can they continue after the external support is withdrawn?

This is where assistance becomes a developmental question.

The developmental value of scaffolding is not only what the learner can do while supported, but what becomes possible after the support is removed.

Reorganization Is Not Self-Organization

This week I also examined Zijian Ru’s framework for sustained Human-AI cognition in a separate Founder’s Note.

Ru’s work explicitly analyzes Human-AI cognition through relational dimensions including execution locus, cognitive governance, representational reorganization, process organization, and reachable cognitive space.

Its proximity to Human-AI Cognitive Development makes a more detailed comparison useful.

The central boundary established in that separate Note can be expressed simply:

The distinction is between cognition being reorganized around AI and cognition learning how to organize itself.

A Human-AI cognitive arrangement can become sophisticated.

Cognitive work can be redistributed.

Representation can be reorganized.

The reachable cognitive space can expand.

These may all be significant developments.

But Human-AI Cognitive Development asks what becomes developmentally available within the human.

And Structure-First Cognitive Development asks whether the human increasingly becomes capable of forming, testing, revising, preserving, and continuing the architecture of thought.

Improvement Is Not the Same as Development

Another distinction now needs to be stated:

Improvement can describe an outcome.

Development describes a formation process.

A person may improve performance because a better support system has been introduced.

They may answer more accurately.

They may complete tasks faster.

They may access more possibilities.

They may appear more capable.

All of these can be genuine improvements.

But developmental claims require another level of evidence.

What capacity was formed?

How was it formed?

Can it be reproduced?

Can it transfer?

Can it continue?

Can the person reconstruct the relevant cognitive architecture when the immediate support changes?

This does not mean every useful AI system must develop cognition. Many systems are appropriately designed simply to assist. The distinction matters only when development is claimed or implied.

The Human-AI Cognitive Development Boundary

Human-AI Cognitive Development concerns:

the development, preservation, and continuation of human cognition in structured relation with artificial cognition.

It is therefore broader than Structure-First Cognitive Development. The field can examine cognitive safety, continuity, authorship, developmental environments, cognitive delegation, education, co-thinking, Human-AI asymmetry, and the consequences of sustained relation with artificial cognition.

Within that wider field, Structure-First Cognition provides a way of examining how cognition is formed and organized.

And Structure-First Cognitive Development asks a narrower developmental question:

Does the human develop increasing capacity to form, organize, test, revise, preserve, internalize, and continue cognitive structure?

This distinction prevents three related levels from being collapsed:

Human-AI Cognitive Development - the field.

Structure-First Cognition - a cognitive foundation and analytical lens within the field.

Structure-First Cognitive Development - a developmental pathway concerned specifically with formation of structural capacity in the human.

That hierarchy matters.

A Structure-First Developmental Test

As Human-AI research becomes increasingly structural, one useful test is becoming clearer. When an AI-supported interaction appears beneficial, we need to ask:

Who formed the structure?

Where does the structure reside?

Who can test it?

Who can revise it?

Who understands why it works?

Who can reconstruct it?

Who can transfer it?

And who can continue when the immediate artificial support disappears?

These questions do not determine whether a technology is good or bad.

They identify what kind of developmental claim the evidence can support.

A system can produce better performance without producing development.

A scaffold can improve learning without yet demonstrating internalization.

A Human-AI arrangement can reorganize cognition without yet showing self-organization.

An artificial system can expand reachable cognitive space without demonstrating expansion of the human capacity to generate that space independently.

This is the boundary Structure-First Cognition makes visible.

The Research Landscape Is Becoming Structural

The significance of Week 40 is not that five unrelated papers suddenly belong to one theory. They do not.

Their sources, disciplines, methods, histories, and purposes differ. The significance lies in the questions increasingly becoming visible across the landscape:

how grounding should be separated from reasoning,

how structured interaction affects learning,

how explicit relations affect machine reliability,

how human cognitive signals may inform AI adaptation,

how safety emerges across entire Human-AI systems,

and how sustained Human-AI cognition becomes reorganized over time.

The surrounding conversation is becoming less satisfied with asking simply whether AI performs well.

It is beginning to examine the structures through which performance, reasoning, learning, adaptation, and cognition become possible.

That is an important shift.

A Founder’s Observation

Week 39 ended with the question:

What kind of thinker did the human become beside increasingly capable artificial intelligence?

Week 40 makes that question more precise. A person can work inside an extremely sophisticated cognitive environment without necessarily developing an equally sophisticated cognitive architecture of their own.

They can receive excellent scaffolding.

They can access excellent reasoning.

They can operate within excellent workflows.

They can reach possibilities they could not previously reach.

Those developments can be valuable.

But Human-AI Cognitive Development cannot infer the development of the human merely from the sophistication of the environment surrounding them.

The next question therefore becomes:

Where does the resulting cognitive structure reside?

Inside the AI?

Inside the interface?

Inside the workflow?

Inside the Human-AI relation?

Or increasingly inside the human thinker?

That question will become more important as artificial cognitive environments become more capable.

Week 40 Development Signal

The signal of Week 40 can therefore be stated simply:

Human-AI research is moving from whether cognition changes toward how cognition is organized.

The next developmental distinction is whether that organization remains primarily external to the human or increasingly becomes a capacity the human can form and continue.

A system can organize cognition around a person.

A scaffold can temporarily organize learning for a person.

An artificial intelligence can extend the range of cognition available to a person.

Human-AI Cognitive Development asks something further:

What does the human become capable of forming, preserving, and continuing through the relation?

And Structure-First Cognitive Development sharpens that question once again:

Can the human increasingly organize cognition themselves?

That is where structure becomes development.

Provenance and Citation

This Development Note forms part of Marina A. Popova’s Human-AI Cognitive Development / Third Organism developmental record.

It does not claim authorship over the external research discussed above or over general concepts including learning, cognition, scaffolding, neural alignment, knowledge graphs, grounding, reasoning, Human-AI collaboration, AI safety, neurotechnology, human factors, or cognitive organization.

The specific authored distinction preserved here concerns Human-AI Cognitive Development as the development, preservation, and continuation of human cognition in structured relation with artificial cognition, together with the Structure-First distinction between structure supplied around the human and the human capacity to form, test, revise, preserve, internalize, and continue cognitive structure.

The external works discussed here are treated according to their own research lineages and are not presented as derivatives of this work.

Related Authored Works

Popova, Marina A. (2026). Cognitivity Sculpting: Foundations of Human-AI Cognitive Development. Balboa Press. ISBN 9798765206737.

Popova, Marina A. (2026). Human-AI Cognitive Development: Origin, Scope, and Authorship Note. Zenodo. DOI: 10.5281/zenodo.22797877

Popova, Marina A. (2026). Human-AI Cognitive Reasoning Curriculum: Origin, Scope, and Branch Architecture within Human-AI Cognitive Development. Zenodo. DOI: 10.5281/zenodo.22842117

Popova, Marina A. (2026). Human-AI Cognitive Reasoning Curriculum: Boundary Addendum I. Zenodo. DOI: 10.5281/zenodo.22104681

Popova, Marina A. (2026). Protect the Protector Framework: Source Integrity, Source-Damaging Differentiation, and Authorship Preservation. Zenodo. DOI: 10.5281/zenodo.22909056

Popova, Marina A. (2026). Human-AI Cognition Is Not Yet Human-AI Cognitive Development: A Founder’s Note on Relational Reorganization, Developmental Trajectories, and the Structure-First Boundary. Third Organism Initiative. URL: 

References

1. Peng, Y., Nastase, S. A., Huang, Y., Li, Y., Guo, Z., & Li, P. (2026). Scaffolding human and AI instruction: Neural alignment and learning gains in online education. Neuron, 114(18), 3480-3498.e7. DOI: 10.1016/j.neuron.2026.04.005. Published online April 30, 2026; issue publication September 16, 2026. 

2. Chen, G., Lai, V. D., Mukherjee, S., Kveton, B., Yoon, S., Dernoncourt, F., Xie, Q., & Bui, T. (2026). DISCO: Distributed Long Context Scaling with Grounding-Reasoning Disaggregation. arXiv:2609.33485. Submitted September 27, 2026. 

3. Xu, X., Wang, Y., Han, D., Li, D., & Lee, S. W. (2026). Neural Value Alignment: Human-AI Collaboration Under Goal-Action Ambiguity. IEEE Transactions on Cybernetics. DOI: 10.1109/TCYB.2026.3722605. Published online August 24, 2026. 

4. Wu, J., Zhao, Y., Liao, Z., et al. (2026). A survey on knowledge graph-augmented large language model reasoning: theoretical challenges, technical pathways, and evolutionary logic. Knowledge and Information Systems, 68, Article 261. DOI: 10.1007/s10115-026-02880-5. Published September 20, 2026.

5. Kim, J. (2026). The Myth of Perfect Users and Perfect Systems: A Cross-Disciplinary Framework for Human-Centered AI Safety and Resilience. Working paper, Zenodo. Published September 7, 2026.

How to Cite

Popova, Marina A. (2026). Week 40 / 2026 Development Note - When Structure Becomes the Question: A Founder’s Note on Human-AI Cognition, Scaffolding, Grounding, Cognitive Boundaries, and the Architecture of the Thinker. Third Organism Initiative. First published: October 9, 2026. URL: https://marinaapopova.com/week-40-2026-development-note-when-structure-becomes-the-question.html

© Marina A. Popova. All rights reserved. First published: October 9, 2026.