The answer arrived in seconds. Did my understanding?

Metacognition is thinking about my own thinking. I use it to tell knowing from recognising, and to decide when to act, check, or ask.

My research asks when a model should say it does not know. This guide asks the same question of me.

Two levels of thinking. Below, a task becomes an answer while an AI draft feeds answers in. Above, a monitor reads how sure I am and a controller decides whether to act, check, or ask. Meta level Object level How sure am I? act check ask What do I do next? monitor control AI draft task thinking answer
Monitoring reads the work below. Control changes what happens next.
Start with the two layers

Metacognition has two parts

Knowledge is what I know about how I think. Regulation is what I do with that knowledge while I work.

Knowledge

What I know, which strategies I have, and when each one fits.

I remember a formula better after deriving it once.

Regulation

Planning, monitoring, and checking my own work while it happens.

I have read this twice. Can I say it without looking?

Working with AI makes this harder

My shortcuts misread it

Habits built on people mispredict how an AI behaves.

It struggles with new problems

Novel, badly defined problems are where it is weakest and I most need to check.

It gives no natural feedback

A colleague frowns when I am wrong. An assistant keeps going.

It cannot do this part for me

The monitoring has to happen in my own head.

The 4 difficulties come from CSIRO’s research on skills for collaborative intelligence. The two parts follow the edtechdev AIED wiki.

Fluent text feels like understanding

Smooth reading gives me a feeling of knowing. An AI writes smoothly whether or not I have understood anything.

Could they quote the essay they had just written?Each dot is one writer. A filled dot could not quote their own essay.
With an LLM15 of 18
With a search engine2 of 18
Brain only2 of 18

Kosmyna et al. (2025), first session. This is a preprint and has not been peer reviewed.

Try it on yourself

Sunlight contains every visible colour. Air molecules scatter short wavelengths far more than long ones. Blue light gets spread across the whole sky. Violet scatters even more than blue. We see blue anyway, because sunlight carries less violet and our eyes favour blue.

Reflective thought starts with being puzzled and needs time with judgement suspended. An instant, confident answer can skip both steps (Singh et al., 2025).

Confidence is a prediction I can score

“I’m 80% sure” is a claim about my own accuracy. In 2 experiments, 5 rounds of prediction with feedback were enough to improve it.

Claim 1 of 5

Sunlight takes about 8 minutes to reach Earth.

Your average confidence and hit rate appear here after each claim.

Ngai and Gilbert (2026) found that 5 practice rounds pairing a prediction with feedback improved calibration. Predictions without feedback did not.

I weigh my confidence against the machine’s

A confident tone raises trust even when the answer is wrong. So I need to know whether its confidence tracks its accuracy.

✓ right · ✗ wrong · filled means I accepted it · dashed means I checked it myself

Wrong answers accepted
1
Answers I checked
6
Right answers accepted
13

Illustrative data: 20 answers, 14 of them right, from 2 made-up assistants with the same accuracy.

How often students accepted wrong ChatGPT advice342 undergraduates on reasoning and evaluation tasks.
Open ChatGPT support62.4%
Same support with a brief reflection prompt39.7%

Ren (2026). Reflection made students more selective, and they kept taking the advice that was right.

Lee et al. surveyed 319 knowledge workers (CHI 2025). Those more confident in generative AI reported less critical thinking. Doyeon Lee and colleagues argue in PNAS Nexus that assistants should report more than confidence. They should say how well that confidence has tracked accuracy before. Pairs who share their confidence can decide better than either person alone (Bahrami et al., Science 2010).

Put the pause where bias gets in

Bias enters twice: in how I ask, and in how I accept the answer. A small, deliberate pause at each point gives my judgement time to arrive.

  1. I frame a question
  2. Pause 1Is my prompt leading?
  3. The AI answers
  4. Pause 2What would make this wrong?
  5. I decide

Turn a leading prompt into a fair one

Press a prompt to rewrite it.

Answer first, then ask

  1. ThinkWork the problem alone first.
  2. AnswerWrite down my answer and how sure I am.
  3. AskAsk pointed questions without revealing my answer.
  4. CritiqueGive it my answer and ask where it fails.

The two pauses follow Lim’s DeBiasMe work (2025). The answer-first order is Michael Gerlich’s advice in the APA Monitor. It stops an early AI answer from anchoring my own.

Decide what to hand over

Some tasks only need doing. Others are how I build a skill. Handing those over can cost me the skill before I notice.

Choose for each task. My own choice appears beside yours.

Detection rate on colonoscopies done without AIEndoscopists in Poland, before and after AI arrived in their clinics.
28.4%Before
22.4%3 months after

Budzyń et al. (2025), reported in the APA Monitor. The rate measures the same skill the AI had been helping with.

Sun et al. ran a field experiment with 250 employees (Journal of Applied Psychology, 2025). ChatGPT access raised rated creativity most for those strong in metacognition. Mutlu Cukurova suggests the same sort for each person’s tasks. Some only need completing, and some build essential learning.

A 5-minute loop around each AI task

Attention can be trained. Clear goals, quick feedback, and a task just beyond my current skill keep the practice honest.

Before · 1 minute

Predict

Write my answer or plan, and a number for how sure I am.

During · 2 minutes

Check

Explain one step back without looking. Trace one claim to its source.

After · 2 minutes

Score

Compare my prediction with the result. Note one place my confidence was off.

When I lead a team

  • Direct attention with purpose. Say why we use AI on this task and which skills we want to keep.
  • Model conscious use. Show my own answer-first habit, including the times the AI was right and I was wrong.
  • Make room for reflection. Ask what people noticed about their own thinking. Keep “I don’t know” safe to say.

The 3 moves come from Hyper Island. The practice conditions follow A Human Edge’s account of flow.

My rule: Before I ask, I write my answer and a number for how sure I am. Afterwards I check one claim and score the guess.

Sources

Guides and articles

  1. Hyper Island (2026). Metacognition: the essential AI leadership skill for 2026.
  2. CSIRO Collaborative Intelligence (IEEE CAI 2024). The importance of metacognitive thinking in an AI-enabled workforce.
  3. Singh, Taneja, Guan and Ghosh (CHI 2025 Tools for Thought workshop). Protecting human cognition in the age of AI.
  4. A Human Edge. Human capability in the AI era.
  5. Mason, Sidra, Reeson and Paris (2023). Collaborating with artificial intelligence? Use your metacognitive skills. Times Higher Education.
  6. Lee, Pruitt, Zhou, Du and Odegaard (PNAS Nexus 2025). Metacognitive sensitivity: the key to calibrating trust and optimal decision making with AI.
  7. Abrams (2026). How AI is reshaping human skills and thinking. Monitor on Psychology 57(5).
  8. Lim (AIREASONING-2025 workshop). DeBiasMe: de-biasing human-AI interactions with metacognitive AIED interventions.
  9. edtechdev AIED wiki. Metacognition.

Studies cited

  1. Kosmyna et al. (2025). Your brain on ChatGPT: accumulation of cognitive debt when using an AI assistant for essay writing task. Preprint.
  2. Ngai and Gilbert (Cognitive Research: Principles and Implications 2026). Metacognitive training facilitates optimal cognitive offloading.
  3. Ren (Frontiers in Psychology 2026). College students’ metacognitive awareness of generative-AI reliance.
  4. Budzyń et al. (Lancet Gastroenterology & Hepatology 2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study.
  5. Sun, Li, Foo, Zhou and Lu (Journal of Applied Psychology 2025). How and for whom using generative AI affects creativity: a field experiment.
  6. Lee et al. (CHI 2025). The impact of generative AI on critical thinking.
  7. Bahrami et al. (Science 2010). Optimally interacting minds.