AI-F06 — Human Judgment, Verification, and Critical Thinking with AI

Course Code: AI-F06  |  School: School of Artificial Intelligence  |  Cluster: Level 1 — Foundations

Level: Introductory  |  Duration: 4 weeks · 12–16 learning hours  |  Language: English  |  Certificate: Certificate of Completion (non-degree)  |  Format: Self-paced with AI support under human supervision

Overview

Access to a capable generative system changes the cost of producing plausible text to almost nothing. It does not change the cost of being right. This course is about the gap between those two facts, and about the working habits that keep a professional on the correct side of it.

The material draws on the literature of judgement and error rather than on the literature of machine learning. Learners study how expert judgement actually degrades, why fluent explanations are disproportionately persuasive, and how structured verification protocols outperform unaided confidence. They then build a personal verification protocol and test it on real tasks under time pressure, which is the condition under which verification is normally abandoned.

The course takes a specific position: the risk introduced by generative tools in professional work is not mainly fabrication, which is detectable, but automation bias, which is not. A wrong answer that a competent person would have caught is dangerous precisely because the person stopped looking.

Learning outcomes

On completion, a successful learner will be able to:

1. Distinguish claims that require verification from claims that do not, using consequence and reversibility rather than intuition.
2. Apply a documented verification protocol to AI-assisted output, including source tracing, independent confirmation and disconfirmation search.
3. Identify automation bias, anchoring and confirmation bias in your own working record, with evidence rather than self-report.
4. Calibrate and express confidence in a way that can later be scored against outcomes.
5. Design a disclosure statement that lets a reader judge how much of a document was machine-assisted and how it was checked.
6. Decide, with reasons, when a task should not be delegated to a generative system at all.

Who this course is for

Any professional whose output carries consequence for others: lecturers, researchers, analysts, policy staff, editors, clinicians in non-clinical decision roles, and managers who sign off on work they did not produce.

Prerequisites

None. AI-F01 or AI-F03 is helpful but not assumed. Learners should be willing to submit examples of their own work for critique.

Syllabus

Module 1 — What judgement is, and how it fails

Focus. Two-system accounts of reasoning and their limits, the conditions under which expert intuition is reliable, and the distinction between bias and noise as sources of error.

Lessons. 1.1 Fast, slow, and the caricature of both. 1.2 When intuition is trustworthy: the Kahneman–Klein conditions. 1.3 Bias versus noise. 1.4 Error you can see and error you cannot.

Core reading. Daniel Kahneman & Gary Klein, “Conditions for Intuitive Expertise: A Failure to Disagree”, American Psychologist 64, no. 6 (2009): 515–526. Daniel Kahneman, Olivier Sibony & Cass R. Sunstein, Noise: A Flaw in Human Judgment (New York: Little, Brown Spark, 2021), part I.

Deliverable. Error log: three past judgements of your own, classified by whether the error was bias, noise, or absent.

Module 2 — Automation bias and the fluency trap

Focus. Why plausible prose suppresses scrutiny, how automation bias operates in professional settings, and the specific failure of accepting an answer because it is well written.

Lessons. 2.1 Automation bias in practice. 2.2 Fluency as false evidence. 2.3 Anchoring on a first draft. 2.4 The disappearance of the second opinion.

Core reading. Richards J. Heuer Jr., Psychology of Intelligence Analysis (Washington, DC: Center for the Study of Intelligence, CIA, 1999), chapters 2–3 and 8. Ziwei Ji et al., “Survey of Hallucination in Natural Language Generation”, ACM Computing Surveys 55, no. 12 (2023).

Deliverable. Blind exercise report: comparison of your judgements on machine-drafted and human-drafted versions of the same brief.

Module 3 — Verification as a protocol, not an attitude

Focus. Source tracing to primary material, independent confirmation, deliberate disconfirmation search, and the discipline of recording what was checked. Structured analytic techniques adapted to AI-assisted work.

Lessons. 3.1 Tracing a claim to its primary source. 3.2 Independent confirmation and what counts as independent. 3.3 Searching for the disconfirming case. 3.4 Analysis of competing hypotheses, applied to machine output.

Core reading. Heuer, Psychology of Intelligence Analysis, chapter 8. Philip E. Tetlock & Dan Gardner, Superforecasting: The Art and Science of Prediction (New York: Crown, 2015), chapters 4–7.

Deliverable. Written verification protocol for your own role, with a worked example applied to a real AI-assisted task.

Module 4 — Calibration, disclosure and refusal

Focus. Expressing confidence so that it can be scored, disclosing machine assistance so that a reader can discount appropriately, and deciding which tasks to withhold from delegation entirely.

Lessons. 4.1 Calibrated language and numerical confidence. 4.2 Scoring your own record over time. 4.3 Writing a disclosure a reader can use. 4.4 The refusal decision: irreversibility, accountability, and dignity.

Core reading. Tetlock & Gardner, Superforecasting, chapters 8–10. Kahneman, Sibony & Sunstein, Noise, part V.

Deliverable. Final portfolio: verification protocol, calibrated judgement set with scoring plan, disclosure template, and a written refusal policy for your role.

Assessment

Pass mark 70 per cent. All assessed components must be attempted. Every mark in this course is issued by a human assessor; no assessment outcome is generated automatically.

Rubric criteria

Each assessed artefact is marked against four criteria at four levels (distinction, pass with merit, pass, fail).

1. Honesty of self-examination: does the learner document error against their own interest rather than presenting a flattering account?
2. Protocol rigour: is the verification procedure specific, repeatable, and actually applied rather than described?
3. Calibration: is confidence expressed in a form that could later be scored, and is it proportionate to the evidence?
4. Judgement about limits: does the learner identify tasks that should not be delegated, with defensible reasons?

Reading list

Core. Richards J. Heuer Jr., Psychology of Intelligence Analysis (Washington, DC: Center for the Study of Intelligence, Central Intelligence Agency, 1999). Philip E. Tetlock & Dan Gardner, Superforecasting: The Art and Science of Prediction (New York: Crown, 2015).

Peer-reviewed. Daniel Kahneman & Gary Klein, “Conditions for Intuitive Expertise: A Failure to Disagree”, American Psychologist 64, no. 6 (2009). Ziwei Ji et al., “Survey of Hallucination in Natural Language Generation”, ACM Computing Surveys 55, no. 12 (2023).

Wider reading. Daniel Kahneman, Thinking, Fast and Slow (New York: Farrar, Straus and Giroux, 2011). Daniel Kahneman, Olivier Sibony & Cass R. Sunstein, Noise: A Flaw in Human Judgment (New York: Little, Brown Spark, 2021).

Standards. NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023), MEASURE function.

All items are published works identifiable by author, title and publisher. Learners obtain them through an institutional library or the publisher. The Academy does not distribute copyrighted texts.

Academic integrity and use of AI

Generative tools may be used in producing assessed work under three conditions. Use must be disclosed in a short statement appended to each submission, naming the tool and the task it performed. Any factual or technical claim originating from a generative tool must be verified against a citable source before it enters assessed work, and the verification must be evidenced. The analytical judgement in each artefact must be the learner’s own and must be defensible in a short follow-up. The portfolio is a record of your own reasoning. Submissions in which the verification protocol was itself produced by a generative tool without adaptation to the learner’s actual role will not be credited.

Instructor: pending owner confirmation. Pricing: pending owner approval. Reference list verified against publisher records; any later addition is marked for verification before publication.