Minimum to Pass

This is the high-yield core — the things you cannot afford to walk in without. If a day runs short, study this first; only go deep into the full lecture pages once the core is solid. Each lecture lists its must-know items and a one-line answer skeleton for its most likely question.

How to use this page

Don't read it like a summary. Cover the answer skeleton and reproduce it from memory. Every item here has appeared in, or directly maps to, Van Rooij's in-class mock (see L8 synthesis). The exam rewards naming the paper and giving the mechanism — not volume.

L1 — Intro & IOS (Elliott, CDR)

Must-know:

Open Society — 4 bases: (1) openness to diversity of knowledge; (2) emancipatory movements & individual rights; (3) constitutional democracy & rule of law; (4) contestable markets & open borders.
Institutions = "rules of the game": formal (laws, regulators — EU AI Act, GDPR) + informal (norms, expectations — what users expect of a chatbot).
IOS = 3 pillars / 15 platforms: Democracy & Good Governance · Transitions & Wellbeing · Equity & Diversity.
TRUST (Elliott et al. 2021, CDR): Transparency · Responsibility · Understanding · Stewardship · Truth.
CDR = corporate responsibility (firms, not just regulators); 3 tenets = Economic Transparency · Societal Wellbeing · Reducing Environmental Impact; centre = Purpose & TRUST.

Skeleton — "Apply TRUST to an AI system X": name X → take each letter (Transparency, Responsibility, Understanding, Stewardship, Truth) → say which fails and the mechanism → name the IOS platform that suffers → one critical note (CDR locates duty with the firm, complementing regulators).

L2 — Decision Making (LBA, Palada)

Must-know:

Newell's time scales: Biological (μs–ms) · Cognitive (100ms–10s ← this lecture) · Rational (mins–hrs) · Social (days–months).
Three levels of integration: Anecdotal → Computational → Algorithmic (mechanism; best individual-difference predictions; the course's "better models → better AI" argument).
LBA — 4 parameters: start point U[0,A] · drift rate N(v,s) · threshold b · non-decision time t₀. Each option's accumulator races; fastest wins.
The two key mappings: difficulty → drift rate; speed/accuracy instruction → threshold.
Palada et al. (2016): UAV-surveillance task (ships obscured by cloud). Selective influence: difficulty (cloud opacity) → drift rate; workload (time pressure) → threshold (lowered to respond faster). NOT "both", each factor drives one parameter.

Skeleton — speed–accuracy trade-off: "Focus on speed" → agent lowers threshold b → decides on less-accumulated, noisier evidence → faster but more errors. (Mechanistic, not just "a trade-off exists.")

Watch (slide mock Q2): parameter for time pressure / UAV operators missing cloud-obscured targets = threshold for caution, drift rate for the cloud-difficulty miss.

L3 — Autonomy & Interaction (Rahwan, digital traces)

Must-know:

Rahwan et al. (2019) Machine Behaviour — study AI like ethology studies animals. 3 scales: individual · collective · hybrid human–machine. 4 domains: Democracy · Kinetics · Markets · Society.
Digital traces (Rafaeli 2019): 3 advantages = (1) bigger/non-WEIRD samples; (2) detailed contextual recording; (3) implicit+explicit 'digital dossier'. Main bias = self-selection.
Kosinski et al. (2013): Facebook likes → SVD → regression predicts traits. Accuracies: ethnicity .95, gender .93, gay-male .88. Hinds & Joinson 2019: computer ≈ spouse (~.56 vs .58) with ~300 likes.
Matz et al. (2017): psychological targeting → ~1.4–1.8× conversion (caveat: Eckles 2018 internal-validity critique).
Kramer et al. (2014) "Facebook study": manipulated ~689k feeds → emotional contagion; tiny effect (d≈.001–.02) but huge population → policy-relevant. No consent → ethics backlash.

Skeleton — small effect, big deal: Kramer's d is tiny but the population is enormous → aggregate real-world influence is large, and it proves platforms can causally shift mood at scale ("machines shape humans"). Effect size doesn't settle the policy question.

L4 — Collective Patterns (Hegselmann–Krause, Douven & Hegselmann)

Must-know:

Emergence: macro patterns from micro interactions; "sum > parts." Coleman's bathtub: Macro → (situation) → Actor → (selection) → Action → (aggregation) → new Macro. Methodological individualism.
Five concepts: emergence · path dependence · tipping points · (non-)monotonicity · direction of effect (aligned / accidental / opposed / no individual counterpart).
Hegselmann–Krause bounded confidence: opinions on [0,1]; agents average only over others within ε. Polarization emerges from unbiased, rational agents (clusters drift out of ε).
Douven & Hegselmann (2021): add truth τ + 3 agent types: Free Riders (social only) · Truth Seekers ((1−α)·social + α·τ) · Campaigners (fixed ρ).
Mis- vs disinformation: mis = make public believe a falsehood; dis = impede belief in a truth. A subtle disinformer (ρ near τ) beats a bold one.

Skeleton — tipping point (slide mock Q3): small parameter change → drastic, possibly discontinuous output change. Example + why (e.g. opinion-dynamics phase transition; climate feedback loops).

Skeleton, free riders vs truth-seekers (slide mock Q3b): free riders don't reduce how many truth-seekers reach \( \tau \), but can worsen their accuracy (sum of squared errors, SSE) by being dragged toward campaigners and pulling truth-seekers off truth. Douven & Hegselmann show this generally; their headline demonstration (Example 3.2) uses a subtle campaign (\( \rho \) near \( \tau \)), where the effect is statistically significant but small (\( \eta^2 \approx 0.004 \)).

L5 — Linguistic Models (van der Vegt, "proceed with caution")

Must-know:

NLP = text → numbers → predict outcomes. Two methods: supervised ML (Google Perspective API: toxicity, severe toxicity, identity attack, insult, profanity, threat) vs dictionary-based (transparent, interpretable, but context/sarcasm-blind).
Study 1 (slides, a separate van der Vegt study, not the 2023 commentary): ~1.9M tweets at Dutch party leaders. Result: female + ethnic-minority politicians received the most threatening tweets; no significant gender difference on "threats."
The caution (slides; the 2023 commentary's own content is the four cautions + VISOR-P): Perspective's identity-attack measure mis-classifies misogynistic content, under-counting gendered abuse, so the "no gender difference" finding is partly a measurement artefact. Opaque, proprietary, biased tools mis-diagnose who is under attack.
Study 2 (Baele 2024): dictionary NLP, ~11.7M incel posts across 33 platforms — interpretable tracking of violent language over time.
CTAP-25: human structured-judgement tool (25 indicators → Low/Med/High Level of Concern). NLP should augment, not replace it.

Skeleton — "proceed with caution": name where Perspective failed (gendered slurs scored ~0.06–0.11 as not identity attacks) → mechanism (biased proprietary model) → consequence (under-protects women & minorities, over-polices peaceful speech) → conclude: augment human SPJ, don't replace.

L6 — Medical AI & Digital Twins (Van Rooij + Bontje, Wang)

Must-know:

Classification (supervised) = predict known labels (patient vs control); Stratification (unsupervised) = find hidden subgroups. Examples: ADHD-from-fMRI (Gaussian Process, ~77%) · COVID-from-demographics (logistic) · ASD stratification (normative modelling + spectral clustering).
Risks of medical classification: wrongful biological determinism (77% ≠ "brain causes ADHD") · inaccurate individual predictions · malignant use (insurers).
Digital twin = virtual, bi-directional real-time model of a city (reflects sensor data and feeds decisions back).
Bontje SWOT: S scenario testing · W limited real-time data / model uncertainty · O national network + reusable modules · T privacy + overreliance.
Wang et al. (2023) federated edge learning (reconstructed; paper paywalled, unverified): train a shared model across local devices without moving raw data. Likely rationale: bandwidth, latency, privacy, resilience. Would mitigate the privacy threat, not overreliance. Confirm against the paper.

Skeleton, federated edge & TRUST: if it works as described, federation keeps raw data local (typically only model updates leave), operationalising Stewardship and helping privacy (Bontje's Threat); but overreliance/Truth needs auditing standards, which the architecture alone doesn't provide.

L7 — Trust in AI (Grimmelikhuijsen & Meijer — paper-only)

Must-know:

Trust (psychological): Competence/ability · Benevolence · Integrity. Algorithm aversion (Dietvorst 2015: one visible error → revert to humans) vs algorithm appreciation (Logg 2019: novel tasks → prefer algorithm).
Input / Throughput / Output legitimacy (who decided / how the process works / are outcomes good).
Six threats (G&M 2022) = two per legitimacy type (the paper's own headings, Tables 2-4): Input — (1) erosion of democratic control; (2) limited responsiveness. Throughput — (3) fails procedural fairness (privacy, legal translation); (4) insufficient checks and balances (opacity, accountability). Output — (5) ineffective/inefficient; (6) undesirable outcomes (bias, lost human contact, value erosion). The shorthand deskilling/opacity/bias/privacy/accountability/value-erosion are the concrete failures filed under these six.
Calibrated response: each threat gets its own strategy (the paper's pairings): democratic control → oversight of purchase/monitoring; responsiveness → civic participation; procedural fairness → Data Protection Impact Assessments (DPIAs) + legal checks; checks and balances → explainable AI (XAI) + clearly recorded responsibility; ineffectiveness → upfront expertise + public-value evaluation; undesirable outcomes → bias minimisation + external evaluation + human contact. Ostrom's "getting the institutions right": no single magic bullet.
Toeslagenaffaire (external illustrative case, not in the paper): ethnic-proxy fraud model → wrongful repayments, bankruptcies, children removed → Rutte III cabinet resigned Jan 2021. Illustrates several of the six threats.

Skeleton, six threats on the toeslagenaffaire: list the paper's six (input: democratic control, responsiveness; throughput: procedural fairness, checks and balances; output: ineffectiveness, undesirable outcomes) → map the case's concrete failures onto them (undesirable outcomes = ethnic-proxy bias; checks and balances = opacity + "computer says no" accountability; procedural fairness = privacy) → name the calibrated reform for the most acute one. Flag that the case is external to the paper.

L8 — Synthesis (the exam map)

Must-know:

Four-level scaling: Cognition (L2) → Psychology (L3, L7) → Networks (L4, L5) → Society (L6, L1). Answers that locate a phenomenon at the right level and link to neighbours score best.
The mock taught five answer shapes: recall+gloss (L1 platforms/TRUST) · model-parameter+mechanism (LBA) · define+illustrate (tipping point) · open-ended critical 4-move (class imbalance) · qualifier-watching MCQ.
Open-ended 4-move (slide mock Q4, rare-event prediction): name the phenomenon (class imbalance / base-rate fallacy) → mechanism (1-in-a-million base rate → false-positive flood even at 99% accuracy) → connect a paper (van der Vegt 2023) → normative conclusion (prioritisation aid, not autonomous decider).
Integrative move: pick an IOS platform → name lectures touching it → paper per lecture → unifying method (simulate before deploy, audit during deploy).

Skeleton — integrative question: take one platform (e.g. Open Cities): L1 CDR · L4 ABM pedestrian flows · L5 NLP safety · L6 Bontje digital twins → unifying move = make policy consequences visible in advance, then audit.