What is Anthropic's '4D' framework for Artificial Intelligence (AI) fluency,…

Research Question

What is Anthropic's "4D" framework for Artificial Intelligence (AI) fluency, what do each of the four Ds, Delegation, Description, Discernment, and Diligence, mean in practice, and how does this framework compare to other published frameworks for taxonomising or compartmentalising AI terminology and concepts, both in scope and in practical design guidance for teams building or governing AI systems?

Findings

Executive Summary

Anthropic's 4D framework is a human-AI fluency model rather than a full taxonomy of AI systems, because it organises the user's work into deciding, specifying, evaluating, and taking responsibility instead of classifying activities, system attributes, harms, or capability levels.

Its four components are clear in Anthropic's official materials: Delegation decides the human-AI split, Description specifies the output, process, and interaction style, Discernment evaluates the result and the reasoning behind it, and Diligence governs responsible choice, disclosure, and ownership.

Compared with NIST, OECD, the collaborative harms taxonomy, Anthropic's workflows-versus-agents framing, and DeepMind's AGI levels, 4D is narrower in analytical coverage but stronger as a day-to-day operating heuristic for teams learning how to work with AI.

Teams building or governing AI agents should therefore pair 4D with a structural taxonomy such as NIST or OECD and, when architecture decisions matter, with Anthropic's workflows-versus-agents distinction.

Key Findings

1. Anthropic defines the 4D framework as four interconnected competencies necessary for AI interactions to remain effective, efficient, ethical, and safe, which makes it a fluency model for human practice rather than a classification scheme for AI systems themselves.
2. Delegation in Anthropic's materials means setting goals and deciding whether, when, and how to engage with AI, and the framework breaks that work into problem awareness, platform awareness, and task delegation across automation, augmentation, and agency modes.
3. Description is the framework's specification layer because Anthropic divides it into product, process, and performance description that respectively define the desired output, the system's method, and the behaviour expected during collaboration.
4. Discernment and Diligence make the framework explicitly evaluative and accountability-oriented, since Anthropic asks users to assess product, process, and performance while also selecting systems carefully, disclosing AI use honestly, and taking responsibility for deployed outputs.
5. NIST's AI Use Taxonomy is broader and more operationally neutral than 4D because it classifies 16 human-AI activity types independent of technique or domain, whereas 4D focuses on competencies a person should apply in any interaction with AI.
6. The OECD framework and the collaborative harms taxonomy both cover governance surfaces that 4D leaves largely implicit, including stakeholders, economic context, data inputs, model properties, task outputs, and harms categories, so they are better suited to policy, registry, and risk mapping work.
7. Anthropic's separate workflows-versus-agents guidance complements 4D by supplying an architecture choice model that the fluency framework does not provide, which means teams can use 4D to structure human practice and use workflows-versus-agents to structure system design.
8. DeepMind's AGI framework shows that 4D sits on a different taxonomy axis from capability and autonomy taxonomies, because it explains collaboration quality while other frameworks explain what systems do, what risks they present, or how capable they are.

Assumptions

• DeepMind's AGI levels are included as a valid comparator because the question asks about frameworks that compartmentalise AI terminology and concepts broadly, not only about narrow agent-operating models.

Analysis

The strongest evidence is around Anthropic's own definitions, because the course page, framework summary, and terminology sheet align on the names and practical meaning of the four Ds.

The main analytical move is therefore not recovering what 4D says, but determining what kind of framework it is relative to other schemes.

On that comparison, 4D resembles a user operating model more than a taxonomy in the NIST or OECD sense, because it tells people how to structure collaboration rather than how to catalogue system features or risk surfaces.

That distinction also aligns with the repository's earlier concept-first taxonomy, which classifies prompts, memory, controls, and tools as system concepts rather than as human competencies, making the two frameworks complementary rather than contradictory.

For practitioners, the trade-off is straightforward: 4D is easier to teach and apply in day-to-day work, while NIST, OECD, harms taxonomies, and architecture taxonomies are better for system inventory, formal evaluation, policy review, and design governance.

Risks, Gaps, and Uncertainties

• Anthropic's public evidence base currently exposes the 4D framework through course assets and downloadable teaching materials rather than through a single standalone technical paper, so the official definitions are clear but the public explanatory depth is thinner than in the NIST and OECD publications.
• The comparison set mixes frameworks designed for different classification objects, which means some differences are purpose differences rather than rival claims about the same object.
• The DeepMind comparator is informative but less directly relevant to day-to-day agent-building teams than NIST, OECD, or Anthropic's own workflows-versus-agents guidance.

Open Questions

• Will Anthropic publish a fuller public paper or transcript that explains the pedagogical rationale behind 4D beyond course assets and summaries?
• Are there other training-oriented AI fluency frameworks from major model providers that are comparable to 4D on pedagogy rather than on governance or architecture?
• How should organisations map 4D-style user competencies onto formal assurance or audit controls without losing the practical simplicity that makes the framework useful?

sources

• [x] Anthropic (n.d.) AI Fluency
• [x] Anthropic (2025) The AI Fluency Framework
• [x] Anthropic (2025) AI Fluency key terminology cheat sheet
• [x] Anthropic (2024) Building effective agents
• [x] Theofanos et al. (2024) AI Use Taxonomy: A Human-Centered Approach
• [x] OECD (2022) OECD Framework for Classifying AI Systems
• [x] OECD (2022) OECD Framework for Classifying AI Systems, two-page overview
• [x] Benbouzid et al. (2024) A Collaborative, Human-Centred Taxonomy of AI, Algorithmic, and Automation Harms
• [x] Morris et al. (2025) Levels of AGI for Operationalizing Progress on the Path to AGI
• [x] Mitchell (2026) AI concept classification taxonomy: prompts, instructions, memory, failure modes, controls, and problem domains
• [x] Mitchell (2026) AI coding harnesses: agent execution model, memory, and context management across commercial and OSS tools
• [x] Mitchell (2026) Agent evaluation framework: cross-repo pattern analysis, commonality detection, and regression identification
• [x] Mitchell (2026) An Integrative Framework for Agent Decision-Making