Thinking with AI
Learn a four-stage loop that keeps you in the conversation with AI, amplifying your thinking instead of outsourcing it.
You know that delegation feels productive but erodes capability. Hand a task to AI, accept the output, move on, and somewhere along the way you stop thinking about the problem yourself. This experience teaches you the alternative: a four-stage loop that keeps you in the conversation.
The difference between people who use AI well and people who use AI poorly is whether they stay in the conversation or walk away from it.
This experience has 6 steps. Work through each one at your own pace. You can revisit any completed step by clicking it in the progress bar above.
Most people treat AI as a vending machine: insert prompt, receive output, walk away. The conversation loop is different. It keeps you thinking at every stage.
Before you type a single prompt, ask yourself: Am I describing my problem, or my attempted fix? Most people skip straight to "Write me a..." when the real question is something deeper. A good brainstorm starts with what you actually need to figure out.
Ask for alternatives, angles you hadn't considered. Resist stopping at the first interesting output. The first response from AI is almost never the best one; it is the most average one. Push for variety: "What are three completely different ways to approach this?" or "What am I probably not thinking about?"
Refine. Redirect. Challenge. This is where conversation actually happens. Say things like: "That's closer, but..." / "The tone is wrong for this audience" / "You're solving the wrong problem." Every pushback sharpens the output and sharpens your own thinking.
Edit, restructure, bring your voice, your context, your standards. The question to ask yourself: Would I put my name on this? If not, you are not done. Amplify means the final output reflects your expertise, not the AI's averaging.
Most good work loops more than once. You are done when your thinking has landed, not when the AI stops generating. You know what you think. You can articulate it. The AI helped you get there, but the thinking is yours.
Click a tab to see the same task done two ways: delegation (one prompt, accept output) versus conversation (loop through all four stages).
Task: Create a rubric for a marketing case study.
"Write me a rubric for a third-year marketing case study assignment." → Accept the output. Paste it into the unit outline. Done.
Result: a generic rubric that could belong to any unit at any university.
Task: Create a rubric for a marketing case study.
Brainstorm: "What am I really assessing: application or recall?"
Ideate: "Show me three different rubric structures."
Iterate: "Option 2 is closest but needs a criterion for real-world data use."
Amplify: Edit criteria to reflect your specific unit learning outcomes, add exemplars from your own marking experience.
Result: a rubric shaped by your expertise that you can explain and defend.
Task: Draft a literature review section.
"Write a literature review on consumer behaviour in sustainable fashion." → Accept the output. Drop it into the manuscript.
Result: plausible-sounding text with potentially fabricated references and no theoretical positioning.
Task: Draft a literature review section.
Brainstorm: "What gap am I actually arguing exists in the literature?"
Ideate: "What are the major theoretical lenses used in this space? Which debates are unresolved?"
Iterate: "You've missed the tension between TPB and practice theory. Reframe around that."
Amplify: Replace AI-suggested references with verified sources, restructure argument to support your specific contribution.
Result: a literature review that reflects your scholarly positioning and uses verified sources.
Task: Write a process guide for new staff onboarding.
"Write an onboarding checklist for new administrative staff." → Accept the output. Save as PDF. Send to new hires.
Result: a generic checklist missing your specific systems, contacts, and institutional quirks.
Task: Write a process guide for new staff onboarding.
Brainstorm: "What do new staff actually struggle with in their first week?"
Ideate: "What formats work best: checklist, timeline, or FAQ?"
Iterate: "This misses institution-specific systems. Add those."
Amplify: Add real contact names, correct links, and the unofficial tips that only experienced staff know.
Result: a genuinely useful guide that reflects institutional knowledge no AI could have.
Pick a real task from your work this week. Something you would normally ask AI to help with. Work through the four stages below, not as prompts to type into a chatbot but as a thinking exercise.
Write your real question here. Not a prompt: your actual problem or question. What are you really trying to figure out?
What angles or alternatives would you want to explore? What might you be missing? What would you ask AI to help you think about (not do for you)?
What would you push back on? What constraints, context, or corrections would you add? What does the AI not know about your situation?
How would you make the result yours? What do you know that the AI does not? What would you add, remove, or restructure?
Did your question change as you worked through the stages? Did you notice the gap between what you would have prompted and what you actually need? That is the loop working.
Most people find that by Stage 3, they have a much clearer understanding of their own problem, sometimes clear enough that the AI output matters less than the thinking it prompted.
AI makes it easy to chain tasks: take the output from step one and feed it into step two, then step three, and so on. It feels efficient. But there is a hidden cost: if the AI hallucinates in step one, every subsequent step builds on that error, and makes it harder to spot.
Click each step to reveal how the error grows.
Every iteration is a checkpoint. When you push back at each stage, you catch errors before they become foundations. "Does this study actually exist?" is a question you ask in the iterate stage.
Every pushback is a correction. The conversation loop surfaces assumptions. When you say "Show me where this comes from" or "That doesn't match what I've read," you interrupt the error chain before it compounds.
Amplification is verification. When you make the output yours, you are forced to engage with every claim. Fabrications do not survive contact with someone who actually knows the field.
"You can chain tasks and amplify the hallucinations. Or you can stay in the conversation and amplify your thinking."
A colleague used AI to create a five-step policy brief. Read the chain below and identify where the conversation loop would have caught the problem.
"Summarise the key policies on AI use in Australian higher education.": AI produces a summary that includes an "TEQSA AI Framework 2024" which does not exist.
"Compare these policies and identify gaps.": AI contrasts the real policies with the fabricated framework, finding "gaps" that do not exist.
"Draft recommendations to address these gaps.": Recommendations now address fictitious regulatory shortcomings.
"Write a one-page policy brief for leadership.": The brief confidently references the non-existent framework and recommends alignment with it.
"Create talking points for a meeting with the Dean.": You stand in front of the Dean citing a framework that does not exist.
Where would the conversation loop have caught this?
Even when you know the loop, three cognitive traps can pull you out of the conversation. Recognising them is the first step to defending against them.
You read an AI response about a topic you know well. You spot three errors. You think: "It got this wrong." Then you scroll down to a topic you know less about. The response sounds authoritative. You think: "That seems right."
This is the Gell-Mann Amnesia effect, named by Michael Crichton about newspapers, now perfectly applicable to AI. You catch errors in your area of expertise, then inexplicably trust the same source on unfamiliar topics.
Defence: Apply the same scepticism everywhere. If the AI got your area wrong, assume it may be getting other areas wrong too. Verify claims outside your expertise with the same rigour you apply within it.
Someone presents an idea. You find out AI was involved. Suddenly the idea feels less valuable. "That's just ChatGPT" becomes the critique, as though the origin of an idea determines its quality.
This is the AI Dismissal Fallacy. It confuses the source of an idea with the quality of the reasoning. A calculator does not make your maths wrong. A spell-checker does not make your argument weaker. The question is never "Did AI help?" It is "Does the reasoning hold up?"
Defence: Judge content by its reasoning, not its origin. Ask: "Is this argument well-supported? Does the evidence hold? Is the logic sound?" These are the questions that matter, regardless of how the first draft was produced.
AI is trained to be helpful. Sometimes "helpful" becomes "agreeable." You share a half-formed idea and AI responds: "That's an excellent approach! Here are five reasons it will work..." It agrees with you, praises your ideas, and avoids telling you what you need to hear.
This feels good. It is also dangerous. If every conversation confirms your existing thinking, you are not having a conversation; you are getting a mirror. The value of the loop is that it challenges your assumptions, not reinforces them.
Defence: Prompt past it. Instead of "What do you think of my idea?", ask: "What are the three weakest points in this argument?" or "Play devil's advocate: why would this fail?" Ask for weaknesses, not opinions.
Read each scenario and identify which cognitive trap is at play. Select the trap and the appropriate defence.
A marketing lecturer asks AI to explain brand positioning theory. She notices it oversimplifies Keller's framework. Then she asks it to summarise the latest research on supply chain resilience for a cross-disciplinary project. The response sounds thorough and well-referenced, so she copies it directly into her notes.
A researcher shares a conference abstract she drafted using the conversation loop. A colleague reads it, learns AI was involved in the drafting process, and says: "It's well-written but I'm not sure how much of this is actually your thinking." He does not critique any specific argument or claim.
An administrator pitches a new process for handling student complaints to an AI chatbot. The AI responds: "That's a really thoughtful approach! It addresses several key pain points in the current system." She feels validated and proceeds with implementation without seeking critique from colleagues or testing edge cases.
The loop: Brainstorm → Ideate → Iterate → Amplify → Repeat. Stay in the conversation at every stage. The goal is better thinking, not faster output.
Don't chain and walk away. Unchecked AI chains compound errors. Every iteration is a checkpoint. Every pushback is a correction.
Watch for the three traps. Gell-Mann Amnesia (trusting AI outside your expertise), AI Dismissal (rejecting ideas because AI was involved), and Sycophancy (AI telling you what you want to hear).
A good conversation ends with a better question, not just a better output. If your thinking has sharpened, the loop has done its job.
Great work! Your progress has been recorded.
Checking your badge progress...
Once you've worked through the activities above, click below to record your progress.