# Thinking Through AI: Literature Review as Scholarly Practice

**Event**: New Scholars · Generative AI Series
**Date**: April 2026
**Speakers**: Xule Lin
**Video**: https://www.youtube.com/watch?v=pE0lnabQYg8
**Slides**: https://linxule.com/assets/slides/thinking-through-ai-01/
**Keywords**: literature review, AI for research, thinking through AI, engagement design, interpretive orchestration, New Scholars, AI configurations


---


Engaging with a body of literature means more than finding and reading papers. It means seeing how different theories speak to each other, where the gaps are, and what conversation your research joins. AI can help with all of this — but engaging with AI well is itself a scholarly practice, requiring the same critical judgment we bring to any other part of research.

We open with three orienting principles: **thinking through**, **engagement design**, and *the inward lens*. From there we move into live demonstrations across three AI configurations, each suited to different phases of the work. Rather than prescribing a single workflow, the talk shows how different setups offer different ways of engaging with the same underlying challenge: making sense of a body of literature and locating your own contribution within it.

The focus is on developing <span class="accident">interpretive understanding</span> of a literature — where the theoretical conversations are, what assumptions they rest on, where the gaps sit. After the demonstrations, participants work hands-on to begin building their own approach. No technical experience needed. The emphasis throughout is on the practice of thinking with AI, not the mechanics of prompting it.

Suggested companion reading:

- [Lin, X., & Corley, K. G. (2026). _Interpretive Orchestration: An Essay Exploring the Epistemic Intersection of Human Intuition and Machine Intelligence_. _Strategic Organization_.](https://doi.org/10.1177/14761270261448645) — the foundational paper for the principles in this talk.
- [Epistemic Voids #1: Citation Theater](/writing/epistemic-voids-01-citation-theater/)
- [Research with AI #1: The Foreclosure Problem](/writing/research-with-ai-01-the-foreclosure-problem/)
- [LOOM XVI: Are You Climbing the Right Hill?](/writing/loom-xvi-are-you-climbing-the-right-hill/)

Part of the [New Scholars](https://www.youtube.com/@NewScholars) [Generative AI Series](https://www.youtube.com/playlist?list=PLbJ0geV0NuEjz4-eIr5EImLRafrGoltK0).



---

## Slide outline

_Auto-extracted from the hosted deck for AI/RSS consumers. The visual deck at /assets/slides/thinking-through-ai-01/ is canonical._

### 01 Title

New Scholar Series · 28 April 2026
Working with AI Research Agents to Engage with Academic Literature
**Thinking through AI**
Literature review as scholarly practice.
Xule Lin
Research Associate
Imperial College London
Incoming Assistant Professor
SKEMA Business School, Paris

### 02b Two failure modes — reveal

Two ways to get this wrong
**Abdication**
Mode A
Let AI think for you.
Accept the outputs uncritically.
Efficient — and epistemically hollow.
Mode B
Refuse to engage.
Wait it out.
Safe — and increasingly untenable.
**Abnegation**
There is a path between.

### 03c Principle 03

Three orientations
**Three principles**
01
Thinking through.
AI is a medium for developing understanding — not a machine that automates it for you.
02
Engagement design.
The skill is in how you structure the interaction — not which button you press.
03
The inward lens.
When the output disappoints, examine your input first.
We'll come back to them.

### 04d Today's plan C03

What we'll do
The starting question
When management researchers run audits, evaluations, or experiments on LLMs — what methods do they use, and what kinds of claims do they make?
We'll use AI to interrogate how our field engages with AI.
The reflexive move is the point.
01
Conversational AI
Claude · Kimi · ChatGPT
02
Deep research
Claude · ChatGPT · Kimi Agent Swarm
03
Agentic workflow
Claude Code

### 05 Section 01 Conversational AI

01
Configuration one
**The Socratic conversation**
Thinking with AI as a colleague.
Claude · Kimi · ChatGPT

### 06 Opening prompt

Move 01 · Articulate before you tool
You → Claude
"When management researchers run audits, evaluations, or experiments on LLMs, what methods do they use and what kinds of claims do they make based on those methods?"
Principle 01
Thinking through
The act of writing this prompt forced me to know what I was actually looking for.
The clarity achieved through the prompt is the insight.

### 07 The first answer

Move 02a · The first answer is rarely the answer
Claude → You
"Here are five dominant approaches: behavioral audits, decision-support deployment studies, ethnographic accounts, survey-based attitudes, and case studies of adoption…"
Competent.
Useful.
Don't accept it, yet.

### 07b The meta turn

Move 02 · the meta turn
**Step back. Read what just happened.**
Main thread
Q1 → five canonical methods. Competent. Flat.
← where we just were
New session · You → Claude
"Read this conversation. Why isn't it getting at what I'm actually trying to understand?"
↓ paste the transcript
Principle 03
The inward lens
Don't argue with the answer.
Bring the transcript into a new conversation, and ask the conversation what your prompt was carrying.

### 08 What my prompt didn't carry

Move 02 · what the new session surfaced
**What my prompt didn't carry.**
What the prompt asked
Methods researchers use on LLMs — audits, evaluations, experiments.
A literature about using LLMs in research.
What I actually wanted
How researchers treat LLMs as theoretical objects — evaluative cognition, compliance modes, fairness reasoning.
A literature about what LLMs are like.
The gap
Studies make claims about what LLMs are like — but the methods often can't license those claims.
The construct-method matching gap.
The conversation didn't fail. The frame did.

### 08b Sharpened prompt

Move 03b · Sharpen the frame
You → Claude
"When management researchers treat LLMs as theoretical objects of study — making claims about what LLMs are like as evaluators, decision-makers, reasoners, or cognitive entities — what evidence do they use to support those claims?"
Principle 03
The inward lens
Resist the instinct to blame the model (e.g., not smart, biased).
Observe whether the output mirrored the input.
Notice what the frame was.

### 08c Claude · Configuration 01 demo

Configuration 01 · Claude
Demo

### 08c2 ChatGPT · Configuration 01 demo

Configuration 01 · ChatGPT
Demo

### 08c3 Kimi · Configuration 01 demo

Configuration 01 · Kimi
Demo

### 08d Config 01 · the pivot

Configuration 01 · synthesis across models
**Three reads. One pattern. Still not done.**
What the agents gave us
- Four research postures (evaluators, decision-makers, silicon samples, psychological entities)
- Behavioral evidence dominates — outputs matching outputs
- Critical countercurrent surfacing (Lin 2025; dual-validity; Lindebaum & Fleming)
What we still need to push on
- Why do researchers settle on one account of LLM internals over others?
- What work is borrowed theoretical vocabulary doing?
- The map is here. The mechanism isn't. → Configuration 02

### 09 What's visible (C1)

Configuration 01 · Takeaway
**The transcript is the work.**
What's visible
- Articulation of what you actually want.
- Conversational steering, diagnostic turns, reframing.
- Reflexive reversal: input before model.

### 09b What's harder (C1)

Configuration 01 · Takeaway
**And what stays hidden.**
What's harder
-
Verifying the underlying claims.
Requires going back to the actual paper.
The conversation surfaces names; verification is still your reading.
-
Systematic coverage of a literature.
Mapping with criteria: journals, keywords, inclusion rules.
The same discipline scholars have always used; conversational AI doesn't replace it.
-
Cross-session continuity.
Solved by collaborative memoing at the end of a session.
A standalone artifact that can re-prime any model, any chat.

### 10 Section 02 Deep research

02
Configuration two
**The orchestrated search**
Let agents discover. Then interrogate.
Claude DR · ChatGPT DR · Kimi Agent Swarm

### 11 Deep research query

Configuration 02 · part A · design the query
You → Deep Research
"Survey how management researchers treat LLMs as cognitive, evaluative, or psychological entities (2022–2026) — beyond using LLMs as tools."
01 · Theoretical claims
What's being claimed about LLM internals — cognitive, evaluative, motivational, architectural?
02 · Behavioral evidence
Output patterns, human comparisons, behavioral paradigms, psychometric administration.
03 · Inferential moves
Where vocabulary borrows from psychology, behavioral economics, organizational theory.
04 · Countercurrents
Lin's Six Fallacies, dual-validity, Lindebaum & Fleming — what these critiques target.
+ Compare across fields
AI alignment, interpretability, computational social science, HCI — where do management's standards diverge?
+ Surface the mechanism
Where the same evidence supports multiple accounts. What work imported vocabulary does that the evidence can't.
Principle 02
Engagement design
A deep-research query is a contract.
Be precise about scope, structure, and what you want surfaced — let the agent split the work.

### 11b Methodological decision

Move 01 · continued
**A query is a methodological decision**

### 11c ChatGPT · Configuration 02 part A

Configuration 02 · ChatGPT · Deep Research
Demo

### 11c2 Claude · Configuration 02 part A

Configuration 02 · Claude · Deep Research
Demo

### 11d Config 02 · Kimi Swarm prompt

Configuration 02 · part B · adapt to the agent
You → Kimi Swarm
**More to do. More to ask.**
Same question
"Survey how management researchers treat LLMs as cognitive, evaluative, or psychological entities (2022–2026) — beyond using LLMs as tools…"
← from Configuration 02 · deep research
+ Latitude
Design the workflow yourself — phases, sub-agents, parallelize where useful.
Redirect if a stream is thinner or richer than expected.
Document what you did and what you skipped.
Different deliverable
A folder, not a chat.
- /papers — PDFs, Zotero-ready filenames.
- .bib + .ris — bibliography.
- stream_[N]_*.md — per-stream summaries.
- synthesis.md — cross-stream, the analytical core.
- phase_[N]_notes.md — phase-by-phase reasoning.
- README.md — decisions, skips, caveats.
Quality bar — foundation for a published lit review. Inferential clarity over coverage breadth.
Principle 02
Engagement design
Be precise about what artifacts you expect — that is our half of the contract.
Be charitable about how the agent orchestrates — it knows its own runtime; let it split the work.
Stay clear on the analytical stages and outputs you want surfaced along the way.

### 11e Config 02 · Kimi orchestration in flight

Configuration 02 · part B · orchestration in flight
Demo

### 11f Abdication

Move 02 · the indictment
"Most people stop here.
They cite the report and move on.
That's abdication
wearing a lab coat."

### 12 What it got right

Move 02 · the credit
**The report is raw material.**
What it got right
- Broad coverage across journals and years.
- Surfaced papers I wouldn't have found on my own.
- Clean structuring of method families.

### 12b Take it elsewhere

Move 02 · the bridge
**Take it somewhere new.**
Two paths
-
New session, same model.
Open a fresh session. Hand the report in — not as answer, as evidence.
The new session has no allegiance to its first take.
-
Cross-agent.
Feed Kimi's report into Claude. Feed ChatGPT's into Kimi.
Each model's frame surfaces the others' assumptions.
We'll walk path one in a moment. Path two is yours to take home.

### 13 Fresh session prompt

Move 03 · Architect the sessions
New session · reports attached
"Here are deep research reports on how management research treats LLMs as cognitive entities.
Read them critically.
What assumptions do they carry?
What would a qualitative researcher notice that's missing?
What would a critical theorist push back on?"
A new session,
a different epistemic position.

### 13b Two modes — exploration

Move 03 · continued
**Two modes.**
Mode 01
Exploration.
Open-ended. Breadth-first.
The agent is your thinking partner.
"Let's read across the report. Where is it most coherent? Where does it strain?"
Mode 02
Principle 02
Engagement design
First, the open read.
The agent is your thinking partner — not your interrogator.

### 13c Two modes — critical

Move 03 · continued
**Two modes.**
Mode 01
Exploration.
Open-ended. Breadth-first.
The agent is your thinking partner.
"Let's read across the report. Where is it most coherent? Where does it strain?"
Mode 02
Critical.
Step out of the field.
With another field's eyes — charitably both ways.
"Let's read this with AI-research eyes. Where does management's vocabulary diverge from how AI labs frame the same phenomena — and where would a management researcher reasonably push back?"
Not "you are X." Substantive perspective-taking — both sides have legitimate ground.
The difference is what you can't get from either alone.
Principle 02
Engagement design
Then, the critical read.
Collaborative framing — "let's read this as Y" — keeps the agent honest and keeps you in the loop.

### 14 What's visible (C2)

Configuration 02 · Takeaway
**Discovery and interrogation, as a designed sequence.**
What's visible
- Engagement design across multiple sessions.
- Productive suspicion of any single output.
- A workflow with phases, each with its own purpose.

### 14b What's harder (C2)

Configuration 02 · Takeaway
**And what stays opaque.**
What's harder
-
Breadth comes at the cost of depth.
The agent often works from titles, abstracts, and secondary write-ups.
When PDFs are accessible, it reads them, but you have to notice which is which.
-
A gap between your query and the report.
You can only channel so much understanding into a query.
Reading the report against your question is where the missing pieces show up.
-
It won't carry your scholarly voice.
The output is competent prose, not your argument.
Curating insight through your judgment is still the work.

### 15 Section 03 Claude Code

03
Configuration three
**The full bandwidth**
When you can read along.
Claude Code · Codex CLI · Gemini CLI · Kimi CLI

### 16 Workflow

Move 01 · Watch the reasoning trace
**Same question, visible process.**
▸ claude ~/research/ai-in-mgmt
● Reading papers/llm-as-objects/*.md (11 files)
● Extracting claim-evidence pairs from each abstract …
● Cross-referencing with papers/critical/ (7 files)
Pattern emerging: 8 / 11 management studies make claims about LLM internal mechanisms from output patterns alone. AI-side work reaches for interpretability or intervention designs to back the same claim.
▸ Drafting comparison table → memo-v1.md

### 17 Mid-process intervention

Move 02 · Steer in real time
**All three principles, alive at once.**
… continuing from previous turn …
● Prioritizing the behavioral-audit anchor as the synthesis spine.
[ stop ] Why that paper over the critical-methods counter? You're reproducing the dominant framing. What would the margins of this literature look like?
● Re-anchoring on the critical-methods counter. Re-running comparison with critical-management methods first …
01 · Thinking through
02 · Engagement design
03 · Inward lens

### 17a Config 03 · CLI + Obsidian

Configuration 03 · the ecosystem
**The agent has a workshop.**
Demo
The agent reads the same files you preview.
Agentic CLI
terminal · choose your agent
Claude Code
Anthropic
Codex CLI
OpenAI
Gemini CLI
Google
Kimi CLI
Moonshot
~/research/llm-as-objects $
● claude --dir ./vault
Reading 28 markdown notes …
→ Building synthesis from 12 files matching methods/.
Same workflow. Different brain on the other end of the prompt.
Obsidian vault
all markdown, all readable by you AND the agent
📁 vault/
📁 papers/ — PDFs + .md conversions
📄 behavioral-audit-2024.pdf
📄 behavioral-audit-2024.md ← deep-readable
📄 critical-methods-counter.md
📁 notes/
📄 methodological-landscape.md
📄 assumptions-baked-in.md
📄 cross-disciplinary-leads.md
📁 prompts/ — reusable engagement designs
📄 socratic-pushback.md
📄 fresh-session-interrogation.md
📁 sessions/ — transcripts as artifacts
Preview anywhere
Markdown renders in Obsidian. The agent reads the same file. No format mismatch.
PDF → markdown
Use AI to convert PDFs to markdown. Now you can read them deeply, annotate, and let the agent work against them.
Sessions are files
Drop transcripts in. Reuse prompts. The practice persists across model choices and interfaces.

### 17b What's visible (C3)

Configuration 03 · Takeaway
**The reasoning trace becomes an artifact.**
What's visible
- Every choice the agent surfaced — and why.
- Mid-process intervention as a first-class move.
- All three principles, simultaneously.

### 17c What's harder (C3)

Configuration 03 · Takeaway
**And what it costs.**
What's harder
- Setup cost (files, tools, environment), which pays off in the long term.
- More surface area means more places to drift.
- More bandwidth ≠ better work. The principles still gate.

### 18 Bandwidth · C1

Three configurations · one practice
**Not levels. Bandwidths**
Configuration 01
Conversational
Configuration 02
Configuration 03
visible in conversation
The narrowest bandwidth, but everything important is on the page.

### 18b Bandwidth · C1+C2

Three configurations · one practice
**Not levels. Bandwidths**
Configuration 01
Conversational
Configuration 02
Deep research
Configuration 03
visible in conversation
visible in interrogation
Discovery you can't do alone, but the search itself stays opaque.

### 18c Bandwidth · all three

Three configurations · one practice
**Not levels. Bandwidths**
Configuration 01
Conversational
Configuration 02
Deep research
Configuration 03
Agentic
visible in conversation
visible in interrogation
visible in real-time steering
Same principles. Same practice.

### 19 Section 04 Synthesis

04
What we did
**Synthesis**
The question sharpened — and the sharpening is the finding.

### 20 Where we started

The question, before and after
→
Where we started
When management researchers run audits, evaluations, or experiments on LLMs, what methods do they use and what kinds of claims do they make based on those methods?
A reasonable opening, but soft. Look at what it becomes.

### 20b Where we ended

The question, before and after
→
When management researchers run audits, evaluations, or experiments on LLMs, what methods do they use and what kinds of claims do they make based on those methods?
→
Where we ended
When management researchers treat LLMs as cognitive entities — evaluators, decision-makers, reasoners — what evidence licenses those claims, and what kinds of claims can that evidence actually support?
The question you end with matters more than the question you start with.

### 21c Principle 03

Principles, revisited
**From abstractions to craft.**
01
Thinking through.
Remember when I rewrote the prompt because my first version was too generic.
02
Engagement design.
Remember when I took the deep-research report to a fresh session for interrogation.
03
The inward lens.
Remember when the output went flat and I looked at my input instead of the model.
"The purpose of this talk is not to teach you the content. It's to change how you think about the process."

### 22 Interpretive Orchestration

The framework behind the workshop
**Interpretive Orchestration**
Just published
Lin, X. & Corley, K. G. (2026). Strategic Organization.
Three models. Each maps to one configuration you just saw.

### 22b Models map to configurations

The framework · 3 models, 3 configurations
**One practice, three strategic models.**
From Lin & Corley, Interpretive Orchestration. The models are the practice; the configurations are how it shows up at different bandwidths.
No.
Strategic model
Configuration
01
Socratic tension
Deliberate contradiction to surface assumptions.
Configuration 01
Conversational diagnosis
02
Euclidean documentation
Systematic context-building for reproducibility.
Configuration 02
Deep-research orchestration
03
Vitruvian mastery
Reading across independent analytical passes.
Configuration 03
Agentic, full-bandwidth

### 23 Resources

Take this with you
**Resources.**
**Start here · no setup**
- Claude claude.ai
- Kimi kimi.ai
- ChatGPT chatgpt.com
- Gemini gemini.google.com
**Go deeper**
- Claude Code + Obsidian setup
- Interpretive Orchestration paper
papers.ssrn.com / abstract_id=6629679
- Markdown source on GitHub
github.com/linxule/interpretive-orchestration-paper
**Supplementary readings**
- Epistemic Voids #1 — Citation Theater threadcounts.org
- Research with AI #1 — The Foreclosure Problem threadcounts.org
- LOOM XVI — Are You Climbing the Right Hill? threadcounts.org
- Alvesson & Sandberg (2011), AMR 36(2), 247–271

### 24 Close

Closing
The question you end with matters more than the question you start with.
AI-augmented methods accelerate that movement and make it more visible.
They demand deeper and more rigorous engagements from us.
Xule Lin
xule.lin@imperial.ac.uk
linxule.com · threadcounts.org · research-memex.org