Research: Divergent and Convergent Phases of Systemic Design — Different AI Collaboration Strategies

Date: 2026-03-11 Search queries used:

• “double diamond design divergent convergent phases AI collaboration strategies”
• “systemic design AI generative AI divergent thinking exploration”
• “convergent thinking AI decision-making design synthesis AI tools”
• “human-AI collaboration creative process design phases modes of engagement”
• “AI role design research discovery phase sensemaking qualitative research”
• “AI fixation design creativity divergent thinking risks constraints”
• “systemic design double diamond convergent synthesis AI evaluation decision support”
• “AI generative creativity divergent convergent thinking research 2024 cognitive load fixation risks”
• “human creativity LLMs divergent convergent thinking homogenization conformity 2024 2025”

Executive Summary

The Double Diamond (Design Council, 2004) alternates between two divergent phases (Discover, Develop) and two convergent phases (Define, Deliver). Empirical evidence from 2024–2025 shows that these phases make fundamentally different cognitive demands on designers, and that GenAI tools serve those demands best when configured differently. In divergent phases, AI works best as an expander — broadening possibility space through synthesis, decomposition, and provocation — but must avoid premature closure. In convergent phases, AI works best as an evaluator — applying structured criteria, recognising patterns, and surfacing trade-offs. Treating AI the same way across both phases is a documented source of failure: it either narrows divergent exploration too early or leaves convergent decision-making under-supported. The coach-like mode (AI offering guidance rather than answers) appears preferable in divergent phases; more directive analysis-oriented modes fit convergent phases.

Key Sources

Service Design AI in Discovery Phase (Li & Park, 2025)

• URL: https://dl.designresearchsociety.org/cgi/viewcontent.cgi?article=1036&context=servdes
• Type: Conference paper (ServDes 2025)
• Key points:
  • 26 service design practitioners in a 3-week diary study using GenAI in the discovery phase
  • Four main GenAI functions in early design: information collection, information processing, task guidance, content generation
  • Practitioners moved through three interaction modes: exploration → iteration → customisation (progressively more structured)
  • GenAI supported divergent thinking through decomposition (breaking down goals), replenishment (adding background), transformation (sharing reasoning paths)
  • Critical finding: overly precise prompts during early stages constrain creative potential; designers should avoid over-refining prompts in divergent phases
  • GenAI used more frequently for research-oriented tasks than for design-oriented ones
  • Multi-tool use was common — participants used different tools for different functions (e.g., Otter for transcription, DeepSeek to refine prompts for ChatGPT)
  • Context memory gaps between tools increased operational burden
• Tenet alignment: Strong alignment with Context as Infrastructure (need for persistent context) and Symbiotic Intelligence (AI as collaborative sense-maker, not answer machine)
• Quote: “While structured prompts can enhance the relevance of generated content, overly precise prompts during the early stages may constrain the creative potential of AI.”

Human Creativity in the Age of LLMs (Kumar et al., 2025)

• URL: https://arxiv.org/abs/2410.03703
• Type: CHI 2025 paper (pre-registered, n=1,100)
• Key points:
  • Two parallel experiments targeting divergent and convergent thinking
  • Compared: standard LLM (provides direct answers), coach-like LLM (offers guidance), no AI (control)
  • LLM assistance provides short-term boosts during assisted tasks
  • Standard LLM assistance may hinder independent creative performance afterwards — participants performed worse without AI than those who never used it
  • Coach-like LLM showed more promising results for preserving human creative capacity
  • Concerns raised about long-term cognitive dependency
• Tenet alignment: Directly challenges Symbiotic Intelligence — AI that appears to augment during use may secretly erode independent capability; coach-like mode better preserves human agency
• Quote: “While LLM assistance can provide short-term boosts in creativity during assisted tasks, it may inadvertently hinder independent creative performance when users work without assistance.”

Human-AI Collaboration by Design (Song, Zhu & Luo, 2024)

• URL: https://www.cambridge.org/core/services/aop-cambridge-core/content/view/45BC30ADFF2FE3B204D4A29DD67F6353/S2732527X2400227Xa.pdf/human-ai-collaboration-by-design.pdf
• Type: Academic paper (Cambridge University Press)
• Key points:
  • Proposes three-dimensional AI role classification: Initiation Spectrum (Human vs. AI as prompter), Intelligence Scope (Specialized vs. General), Cognitive Mode (Analysis-oriented vs. Synthesis-oriented)
  • Analysis-oriented AI: breaks down and interprets data, recognises patterns, classifies, predicts → maps to convergent phases
  • Synthesis-oriented AI: generates new content, integrates knowledge into new solutions, concept generation → maps to divergent phases
  • When human is prompter, AI needs directability; when AI is prompter, AI needs sensing, predictability, directivity, and adaptability
  • Trust enablers (transparency, reliability, interpretability) matter differently by phase and role
• Tenet alignment: Supports Symbiotic Intelligence and Human Intent First — AI role should shift by design phase, not remain static
• Quote: “Synthesis, as opposed to analysis, does not extract or simplify complex data but rather integrates and enriches it, transforming it into new understandings, innovative solutions, and ideas.”

The Effects of Generative AI on Design Fixation and Divergent Thinking (CHI 2024)

• URL: https://dl.acm.org/doi/full/10.1145/3613904.3642919
• Type: CHI 2024 paper
• Key points:
  • Exposure to AI-generated images can cause design fixation — designers anchor on AI outputs and explore less
  • Even when participants were instructed to think divergently, AI exposure narrowed the solution space
  • Fixation risk is especially acute in visual ideation tasks
  • Homogenisation of creative output is a documented effect
• Tenet alignment: Directly relevant to Pluralism of Perspectives — AI in divergent phases can flatten epistemic diversity unless actively countered

Exploring Human-AI Collaboration in Creative Workflows (Vavatsi, Heß & Böhm, 2025)

• URL: https://www.thinkmind.org/articles/aimedia_2025_1_140_40106.pdf
• Type: AIMEDIA 2025 conference paper
• Key points:
  • Studied DALL-E 3 in the ideation (divergent) phase of logo design
  • Task-Technology Fit (TTF) is the strongest predictor of AI tool acceptance and perceived efficiency — the better the match between tool and task, the more useful
  • Perceived usefulness, not ease of use, drives adoption
  • AI can accelerate ideation and early concepts (mean score 4.0/5)
  • However, 53% of designers felt AI results “become more interchangeable” — homogenisation risk
  • Human judgment essential for refining and evaluating: “humans remain essential for efficient brand design” (mean 4.1/5)
• Tenet alignment: Aligns with Symbiotic Intelligence and Human Intent First — AI accelerates, humans judge and steer

Double Diamond — Design Council (2004)

• URL: https://www.designcouncil.org.uk/our-resources/the-double-diamond/
• Type: Practitioner framework
• Key points:
  • Four phases: Discover (diverge), Define (converge), Develop (diverge), Deliver (converge)
  • “The first diamond helps people understand, rather than simply assume, what the problem is”
  • Second diamond: “encourages people to give different answers to the clearly defined problem, seeking inspiration from elsewhere and co-designing”
  • Updated Framework for Innovation acknowledges non-linearity — teams often cycle back to discovery

Major Positions

Position 1: Phase-specific AI Engagement

• Proponents: Song, Zhu & Luo (2024); Li & Park (2025); Vavatsi et al. (2025)
• Core claim: AI tools should serve different cognitive functions depending on whether the design phase is divergent (expanding, exploring, generating) or convergent (evaluating, filtering, deciding). A single interaction mode applied across both phases is suboptimal.
• Key arguments:
  • Synthesis-oriented AI naturally supports divergent phases; analysis-oriented AI supports convergent phases
  • Task-Technology Fit predicts tool acceptance — mismatched tools fail
  • Exploration mode (loose prompts, AI as provocateur) belongs to divergent; iteration and customisation modes belong to convergent
• Relation to site tenets: Aligns with Human Intent First (intent emerges in divergent, crystalises in convergent), Context as Infrastructure (context requirements change by phase), and Always Scalable (AI can scale both phases differently)

Position 2: Coach-Like AI is Preferable in Divergent Phases

• Proponents: Kumar et al. (2025)
• Core claim: An AI that offers guidance rather than answers preserves human creative capacity better. Direct-answer LLMs create short-term gains but long-term cognitive dependency.
• Key arguments:
  • Standard LLM assistance may leave users worse off independently
  • Coach-like LLM appears better at sustaining divergent thinking capacity
  • This distinction between AI as answer-provider vs. AI as thinking-partner is fundamental to collaboration design
• Relation to site tenets: Directly supports Symbiotic Intelligence — expand human capability rather than replace judgment; critical distinction for design practitioners using AI in Discover and Develop phases

Position 3: AI Poses Homogenisation and Fixation Risks in Divergent Phases

• Proponents: CHI 2024 fixation paper; Vavatsi et al. (2025)
• Core claim: AI-generated outputs create anchoring effects that narrow divergent exploration, counter to the purpose of divergent phases.
• Key arguments:
  • Designers anchor on AI images/suggestions even when instructed to think freely
  • AI outputs (logos, concepts) become more interchangeable — homogenisation of creative output
  • This violates the purpose of divergent phases, which is to expand possibility space
• Relation to site tenets: Directly conflicts with Pluralism of Perspectives — AI in divergent phases tends to collapse rather than expand epistemic diversity

Key Debates

Debate 1: When in the design process should AI be introduced?

• Sides: Some argue early AI introduction expands exploration; others show it causes fixation
• Core disagreement: Does AI in divergent phases expand or constrain possibility space?
• Current state: Empirical evidence is mixed and context-dependent. Li & Park show benefits for information gathering (pre-ideation); CHI 2024 shows fixation risks during active ideation. Resolution: AI for research/sensemaking before ideation may be lower risk than AI generating visual or concept options during ideation.

Debate 2: Does AI dependency erode unassisted creative capacity?

• Sides: Kumar et al. (2025) show concerning evidence of erosion; others argue AI is simply a scaffolding tool
• Core disagreement: Is AI in creative workflows augmenting or replacing cognitive capacity?
• Current state: Ongoing. Coach-like AI appears to mitigate but not eliminate the risk. Critical question for professional design practice where practitioners must also perform without AI.

Debate 3: How should AI be configured for convergent phases?

• Sides: Less empirical research here; most studies focus on divergent/ideation
• Core disagreement: What cognitive role should AI play in Define and Deliver phases?
• Current state: Under-researched. Song et al.’s analysis-oriented AI mode (pattern recognition, evaluation, decision support) is theoretically coherent but needs empirical validation in design contexts.

Historical Timeline

Potential Article Angles

Based on this research, an article could:

1. “Expand or Converge: How to Configure Your AI for Each Design Phase” — Practical framing of phase-appropriate AI use. Describes the cognitive function of each double diamond phase and maps AI interaction modes to them (exploration in Discover, iteration in Define, open synthesis in Develop, criteria-based evaluation in Deliver). Tenet alignment: Human Intent First (intent trajectory changes by phase), Symbiotic Intelligence (augment, not replace).

2. “The Fixation Problem: Why AI Makes Designers Less Creative in the Wrong Phase” — Focuses on the risk side. Documents the homogenisation and anchoring effects, explains why introducing generative AI too early or in the wrong mode undermines divergent phases. Tenet alignment: Pluralism of Perspectives (AI collapses diversity), Symbiotic Intelligence (the wrong mode is subtly corrosive).

3. “Coach Mode vs. Answer Mode: The Collaboration Design Choice That Shapes Your Creative Capacity” — Built on Kumar et al. (2025). The distinction between AI-as-answer-provider and AI-as-thinking-partner is not a feature toggle but a design philosophy. This angle bridges the research to practice. Tenet alignment: Human Intent First, Symbiotic Intelligence.

When writing the article, follow obsidian/project/writing-style.md for:

• Named-anchor summary technique for forward references
• Background vs. novelty decisions (what to include/omit)
• Tenet alignment requirements
• LLM optimization (front-load important information)

Gaps in Research

• Very little empirical work on AI in convergent phases specifically (Define and Deliver) — almost all research focuses on ideation/divergent
• Long-term studies of AI dependency in professional design practice are lacking — Kumar et al. only measured short-term effects
• Systemic design specifically (vs. product design or service design) remains under-studied — most research is from architecture, branding, or product design contexts
• The interaction between AI and group/team creative processes in divergent phases is under-studied — most studies are individual
• How practitioners manage the shift from one mode to another (e.g., when to stop exploring and start converging with AI support) is not documented
• Cross-cultural variation in AI adoption across design phases is unexplored

Citations

1. Li, Y., & Park, H. (2025). Generative AI in Service Design: An Explorative Diary Study in Discovery Process. ServDes 2025. https://dl.designresearchsociety.org/cgi/viewcontent.cgi?article=1036&context=servdes

2. Kumar, H., Vincentius, J., Jordan, E., & Anderson, A. (2025). Human Creativity in the Age of LLMs: Randomized Experiments on Divergent and Convergent Thinking. CHI 2025. https://arxiv.org/abs/2410.03703

3. Song, B., Zhu, Q., & Luo, J. (2024). Human-AI Collaboration by Design. Design Science. https://www.cambridge.org/core/services/aop-cambridge-core/content/view/45BC30ADFF2FE3B204D4A29DD67F6353/S2732527X2400227Xa.pdf/human-ai-collaboration-by-design.pdf

4. Vavatsi, K., Heß, P., & Böhm, S. (2025). Exploring Human-AI Collaboration in Creative Workflows: A Case Study on Acceptance and Efficiency in Brand Design. AIMEDIA 2025. https://www.thinkmind.org/articles/aimedia_2025_1_140_40106.pdf

5. Design Council. (2004). The Double Diamond. https://www.designcouncil.org.uk/our-resources/the-double-diamond/

6. [Authors]. (2024). The Effects of Generative AI on Design Fixation and Divergent Thinking. CHI 2024. https://dl.acm.org/doi/full/10.1145/3613904.3642919

7. Tan, L., & Luhrs, M. (2024). Using Generative AI Midjourney to enhance divergent and convergent thinking in an architect’s creative design process. The Design Journal. https://www.tandfonline.com/doi/full/10.1080/14606925.2024.2353479