Cognitive Offloading
Interpretation, synthesis and judgement can increasingly be delegated to AI, reducing the immediate cognitive effort required from the user.
Second Thought
Better thinkers, not faster answers.
Second Thought is a spatial Human–AI reasoning workspace designed to keep people actively involved in thinking with generative AI.
It makes context, contribution and judgement visible — giving users more control over how AI participates in their reasoning.
Product Development Process
AI can make capability development a black box.
As generative AI becomes embedded in design work, polished outputs are becoming easier to produce. But capabilities such as questioning, interpretation and contextual judgement develop through repeated practice — especially when people have to work through uncertainty themselves.
My research suggested that AI could disrupt this development through both individual cognitive processes and educational processes.
Key Behaviours
Premature Acceptance
Fluent and coherent AI responses can create confidence before the user has fully understood, questioned or verified the output.
Delegated Uncertainty
Instead of working through what is unknown, users can ask AI to define the direction, interpretation or next step for them.
Reduced deliberate practice
When AI repeatedly performs interpretation and judgement on the user's behalf, fewer opportunities remain to practise the questioning, reframing and deliberation through which independent professional judgement develops.
Reduced visibility for meaningful critique
When polished outputs obscure how reasoning developed, educators have less visibility into what students understand, where uncertainty remains and where meaningful critique or scaffolding is needed.
A weakened capability-development mechanism
AI may be changing not only what student and junior service designers produce, but also how they develop the judgement, patience, confidence and reflective capabilities required to practise independently.
Where should design intervene?
How might we help students use generative AI in ways that keep their thinking visible and critical reflection active — so AI strengthens rather than bypasses the development of professional judgement?
Intervention Strategies
Re-engage people in the thinking process
Keep interpretation, questioning and judgement active during Human–AI interaction.
Reconnect AI-assisted work with critique
Make reasoning visible enough to support dialogue, scaffolding and meaningful intervention.
Design Principles
Externalise / spatialise thinking
External representations can become persistent, manipulable objects to think with.
Create purposeful friction
Small interruptions or moments of slowness can create space for reflection.
Ask, Don’t Tell: AI provokes thinking
AI can question, explain, discuss and scaffold rather than simply complete the task
Keep judgement human
Delegation, verification and final judgement remain human-led when using AI.
Make thinking discussable
Externalised thinking can become material for critique and reflective dialogue.
Low-risk experimentation
Safe place to experiment with AI to help students understand conflict effects
Two intervention hypotheses. One direction to pursue.
Second Thought was not the starting concept.
Shared AI Learning Space
Bring AI-assisted learning into a shared environment where tutors can scaffold, discuss and intervene.
Visible Reasoning
Externalise critical moments of Human–AI reasoning so students can recognise, question and articulate how their thinking develops.
Multi-stakeholder Validation
- Students
- Learning & Teaching
- Service Design Tutor
Selected: Visible Reasoning
- Student relevance ↑
- Problem alignment ↑
- Adoption risk ↓
Stronger alignment with student needs and the core capability problem, with fewer concerns around monitoring, control and implementation complexity.
I built early to let evidence change the product.
First Prototype
Visual Reasoning Workflow
Can spatial interaction make Human–AI reasoning easier to follow?
Reflective Checkpoints
Can selective friction bring people back into interpretation and judgement?
Testing
11 participants · 2 iterative rounds
- Task-based Testing
- Structured Evaluation
- Follow-up Interviews
Quantitative Evidence
What Changed My Judgement
Interaction ≠ Critical Engagement
Interaction mechanics alone did not guarantee that users were genuinely questioning AI output.
Selecting and annotating something does not necessarily mean the student has questioned it.
AI Behaviour Matters
“Spoon-feeding” and overly agreeable AI could still encourage premature acceptance.
…sometimes AI feels like it's spoonfeeding people.
Friction Must Be Selective
Reflection created value at some moments, but became unnecessary interruption at others.
When I'm trying out this product and haven't seen any results yet, asking me to enter information right away might put me off a bit.
Evidence → Product Redesign
Justification→Provenance
Default AI Behaviour→Interaction Strategies + Independent Stance
Fixed Reflective Checkpoints→Graduated Cognitive Friction
Evidence changed three major product decisions.
How Second Thought Works
Spatial Reasoning & Context
Branch and connect ideas while carrying relevant context into each line of inquiry.
Interaction Strategies & Independent Stance
Choose how AI participates while reducing automatic agreement.
Cognitive Friction
Vary how strongly users are asked to re-enter the thinking process.
Human Contribution & Provenance
Turn AI output into evidence, interpretation, insight and decisions that remain traceable.
Provenance shows contribution and influence — not truth.
External Expert Validation
The deployed prototype was independently tried by a J.P. Morgan Service Design Vice President, followed by structured ratings and open feedback.
CORE MECHANISM EVALUATION
- Reasoning Workspace & Visualisation
- 7/ 7
- Friction Modes
- 6/ 7
- Thinking & Reflection Capture
- 7/ 7
- Provenance & Evidence Awareness
- 7/ 7
OTHER EVALUATION DIMENSIONS
- Agency & AI Control
- 7 / 7
- Collaboration Transparency
- 7 / 7
- Reasoning Articulation
- 6 / 7
- Critical Engagement
- 6 / 7
- Concept Clarity
- 6 / 7
- Information & Interaction Clarity
- 5 / 7
- Ease of Use
- 6 / 7
- Overall Product Value
- 5 / 7
7-point evaluation following independent product use by one external expert.
Expert Voice
“You’re powering your thinking with AI, but at the same time, it remains yours.”
What resonated
Visible reasoning and retained ownership became distinctive once experienced.
What remained unclear
The landing experience initially felt too similar to a standard AI chatbot.
What real projects would need
Richer document and multi-source context.
Scope Decision
I deliberately deferred multi-source RAG, a composable Node Library, adaptive supervision and complex-canvas scaling to first validate the core Human–AI collaboration model.
Scope was treated as a product decision — not a list of unfinished features.
What’s Next
Clarify the Value
Make the difference from a standard AI chat immediately visible.
Bring Real Project Context In
Support richer documents and multi-source inputs while preserving traceability.
Make Human–AI Workflows Composable
Explore reusable cognitive and AI-interaction capabilities through a Node Library.