Reasoning scaffolds: An infrastructure for human subjective reasoning

22/10/2025 ☼ not-knowing ☼ uncertainty ☼ reasoning ☼ critical thinking ☼ learning ☼ education ☼ AI

tl;dr: Every important decision requires subjective reasoning about objective facts—deciding what matters and why. Yet we have almost no explicit frameworks for this work, which is a source of civilisational fragility in an age of AI and accelerating political and social fragmentation. Reasoning scaffolds fill this gap: they’re an explicit structure to make subjective reasoning visible and learnable. When tested with university students, they were able to use it to generate dramatically better arguments in 75% less time. But we still lack a robust theoretical foundation for reasoning scaffolds—understanding what scaffolds share across domains, how to annotate the artifacts of subjective reasoning to show their structure, and how to make this machine-legible for human and AI training. A research programme addressing these questions would create infrastructure for both better human reasoning and AI systems that support human judgment rather than displace it.

Every important individual or collective decision humans make—about business, policy, technology, ethics, coordination—requires subjective reasoning about objective facts. The fact that a new apartment building has been built in a neighbourhood isn’t subjective. Whether this is good or bad, desirable or undesirable, requires judgment that is necessarily subjective because it varies by perspective. One person could argue it’s unambiguously good because more supply means less expensive and more affordable housing. Another could say it’s bad because more supply means property values will fall, hurting long-term neighbourhood loyalists. A third could say it would be good, but only if transit infrastructure is expanded to accommodate the neighbourhood’s increased population. The list goes on. Each subjective position about the fact of the new apartment building requires constructing an argument, understanding the audience, and marshalling appropriately persuasive evidence.

This work of human subjective reasoning (HSR) isn’t optional. It drives business decisionmaking, technological and business model innovation, Supreme Court decisions, AI governance choices, climate policy, and all forms of political coordination. Doing it well and transparently is critical for civilisational persistence, because coordination breaks down when we can’t agree about subjective things.

Yet we have almost no explicit frameworks for it. We lack the theoretical and practical infrastructure for making subjective reasoning visible, teachable, and auditable. This gap isn’t just an academic curiosity—it’s a civilisational fragility.

I’ve been thinking about this problem from different angles for two decades. My research on uncertainty focused on how teams and organisations make decisions when the requirements of formal rationality don’t apply and they can’t rely on pure calculations of risk. My work on developing a theoretical framework for meaning-making explores how humans decide what matters and why—and highlights how meaning-making about the subjective value of things is the bright line between what humans must do and what machines can do. During my recent Future of Life Foundation fellowship on AI for human reasoning, these threads came together into something concrete: what I call reasoning scaffolds.

What are reasoning scaffolds?

A reasoning scaffold is the explicit articulation of the structure of a piece of human subjective reasoning. At minimum, a reasoning scaffold must identify the component pieces of a reasoning process, how those components are sequenced, and—crucially—which components involve subjective judgment versus objective fact.

Let’s return to that housing example. All arguments are statements of subjective positions, not statements of objective facts. Making an argument thus requires subjective reasoning. A strong argument about whether more housing is good requires several distinct elements: the subjective position you’re taking (good/bad/conditional), the audience you are speaking to (whether the audience currently disagrees with, is neutral to, or supportive of your position), the kinds of evidence that would convince that specific audience, the rhetorical framing that resonates with them, and the existing arguments you’re responding to or building on. These elements aren’t arbitrarily sequenced. There’s a logic and structure to effective reasoning scaffolds for subjective reasoning.

During the fellowship, I developed and tested one particular reasoning scaffold, designed for university students. I implemented it as an LLM-supported web application and tested it with undergraduates.

Users of the webapp took a quarter of the time needed to develop a robust and well-articulated provisional argument, with dramatic quality improvements in the arguments as assessed by third parties. Users went from vague proposals to sharp, well-evidenced positions rapidly, often in just 10-20 minutes (compared to 2 hours or more when I’ve facilitated pen-and-paper workshops using a similar reasoning scaffold).

The scaffold worked by providing structured prompts that guided users through iterative stages of subjective reasoning by functioning as a Socratic mirror. The crucial design choice turned out to be making explicit at every step what the user must decide (subjective judgments about value and meaning) versus what the system could support (information organisation, systematic prompting). Unlike standard AI chat interfaces that obscure this boundary between meaning-making and other work, the scaffold was designed to make it visible and unavoidable.

Why this matters

Here’s what became clear from that work: there are multiple reasoning scaffolds, not one universal scaffold. A scaffold for convincing opponents differs from one for helping neutral parties form views. A scaffold for Cartesian scientific reasoning differs from one for theological argument or legal reasoning or political negotiation.

The prototype application I built has been tested by undergraduates writing term papers or proposing customised majors. However, the same reasoning scaffold has also been used by multilateral agencies proposing economic development strategies, startup founders articulating new business models, and public utility strategy teams developing plans for long-term infrastructure investment. The insight here is that reasoning scaffolds likely have an underlying general pattern—a meta-structure that generalises across contexts.

Understanding this meta-structure is essential for three reasons.

First, for human reasoning improvement. When we make scaffolds explicit, humans can learn to reason more effectively. They can see where subjective judgment enters their arguments, identify appropriate evidence for specific audiences, and construct more rigorous positions. Instead of treating good reasoning as mysterious talent, we can identify its components and teach them systematically.

Second, for better, more robust coordination. Groups cannot coordinate voluntarily without shared subjective reasoning about collective goals. If a group doesn’t share reasoning about why a goal matters, the only alternative is coercion—which is fragile and inefficient. Making scaffolds explicit enables better deliberation by showing where disagreements are genuine (different values) versus where they stem from poor argumentation or unstated assumptions. This matters for everything from climate policy to AI governance to democratic stability.

Third, for better AI alignment and safety. This is where it gets interesting. Text that encodes or represents the artifact of a subjective reasoning process has a latent structure that current training doesn’t capture. If we can annotate reasoning artifacts—Supreme Court decisions, policy documents, legislation—to make their reasoning scaffolds explicit, we create fundamentally new training data. We can more clearly distinguish factual claims from subjective positions, identify audience-appropriate argumentation, and show the architecture of rhetoric and persuasion.

This could enable several things that current AI systems struggle with:

Better human-AI reasoning interfaces. Instead of the blank prompt box that obscures what AI can and cannot do, scaffold-aware tools can explicitly support humans in organising information while keeping subjective judgment clearly in human hands. My fellowship prototype demonstrated this works—users maintained agency over value judgments while getting systematic support for information processing.

More inspectable AI systems. If we train AI on scaffold-annotated data, we can potentially query what reasoning structures it has learned and how it’s applying them. This makes AI reasoning more interpretable—not by trying to peek inside the black box, but by training systems on data that explicitly encodes reasoning structure.

Richer alignment targets. Current AI alignment work often treats “human values” as an undifferentiated blob to be learned from preference data. But human subjective reasoning has structure. It involves positions, audiences, evidence standards, and sequencing. Making this structure explicit gives us clearer targets for what we want AI systems to learn and support.

A reasoning scaffolds research programme

The fellowship work produced one scaffold that works well in specific contexts. But we lack theoretical foundations that would make it possible to generalise scaffolds across contexts. We don’t understand what scaffolds share across domains, how to annotate existing artifacts to show their scaffolding, or how to make scaffolds machine-legible for training and deployment.

So a research programme on reasoning scaffolds would tackle three connected questions.

First, understanding meta-structure. Survey reasoning scaffolds across domains—scientific research, policy analysis, legal reasoning, political argumentation. The goal isn’t comprehensive coverage but sufficient examination to identify patterns. What elements appear across all or most contexts? What varies by domain and what remains constant? Is there a pattern language that can describe different scaffolds using shared vocabulary? How context-dependent are scaffolds?

This would involve analysing existing artifacts (court decisions, policy documents, research papers) for their implicit scaffolding, plus interviewing practitioners about how they actually construct arguments in their domains. The output would be a theoretical framework—potentially a pattern language—suitable for later formalisation.

Second, developing annotation methods. Experiment with ways to annotate artifacts of subjective reasoning to make scaffolding explicit. A Supreme Court decision currently reads as continuous prose, but underneath lies a scaffold: factual claims, value judgments, audience considerations, rhetorical strategies, precedent citations, and logical sequencing. Making this visible through annotation serves multiple purposes.

For learning: students could see how effective arguments are actually constructed. For auditing: analysts could evaluate whether reasoning is sound or spot where subjective judgment masquerades as fact. For machines: annotated artifacts become training data that teaches recognition of reasoning structures rather than just surface patterns.

The goal wouldn’t be a finished annotation standard but proof-of-concept methods demonstrating feasibility and surfacing key design choices.

Third, preliminary machine-learning experiments. Take a small language model and annotate a subset of its training data—say, Supreme Court decisions—to explicitly mark reasoning scaffolds. Distinguish factual claims from subjective positions, identify audience considerations, show argumentative sequencing. Fine-tune the model on this enriched data.

Then test: Can it better recognise reasoning scaffolds in new artifacts? Can it provide better support to humans constructing arguments by making scaffold elements explicit? Does it maintain the boundary between subjective human judgment and machine support?

This is exploratory work, not ready for production at scale. But it would demonstrate whether scaffold annotation is technically feasible, whether annotated training data changes model behaviour productively, and whether machine-legible scaffolds open new possibilities for human-AI reasoning collaboration.

Why now?

We’re at an inflection point. LLMs are increasingly capable of producing text that looks like human reasoning. But they don’t actually do subjective reasoning or meaning-making. They can’t make genuine value judgments or decide what should matter—and they shouldn’t be allowed to. As these systems become more integrated into decision-making, we’ll become increasingly prone to mistaking this superficial fluency for actual subjective reasoning capacity.

Meanwhile, our political and cultural discourse is literally falling apart in real-time. Coordination is getting harder. The gap between our technological capabilities and our collective decision-making capacity is widening. Many catastrophe scenarios involve cascading failures of coordination and judgment—groups unable to reason together effectively about what matters, even when everyone individually wants to avoid disaster.

Reasoning scaffolds provide infrastructure to address both problems. They make subjective reasoning explicit, teachable, and auditable for humans. They create new possibilities for training AI systems that genuinely support human reasoning rather than displacing it. They enable political and policy discourse that’s more rigorous and effective.

This isn’t a complete solution to AI alignment or coordination problems. But it’s crucial infrastructure we currently lack. The theoretical work needs to be done. The sooner we have a way to talk about the meta-structure of human subjective reasoning, the sooner we can build tools and systems that make it more legible—both to ourselves and to our machines.

---

For the last few years, I’ve been wrestling with the practical challenges of meaning-making in our increasingly AI-saturated world, developing frameworks for how humans can work effectively alongside these powerful tools while preserving the meaning-making work that is the irreplaceably human part of the reasoning we do. I’ve published this as a short series of essays on meaning-making as a valuable but overlooked lens for understanding and using AI tools.

I’ve also been working on turning discomfort into something productive. idk is the first of these tools for productive discomfort.

And I’ve spent the last 15 years investigating how organisations can succeed in uncertain times. The Uncertainty Mindset is my book about how to design organisations that thrive in uncertainty and can clearly distinguish it from risk.