Lecture 8 — Synthesis & Mock Exam (Van Rooij)
This is the "putting everything together" lecture and includes the in-class mock exam Van Rooij used. This is the closest thing to a real exam preview you will get — pay attention to question style, not just content.
Exam logistics (from the slides)
- Date / time: Fri 2026-05-29, 11:00–13:00 (re-confirmed)
- Location: EDUC-BETA (Rupert D for extra-time students)
- Laptops are provided — you do not bring your own
- Format: 9 essay questions; 2 hours total; ~13 min/question
- (Mock had 5 questions covering L1–L6 + papers) — the real exam scales that up to 9 questions but in the same style
The unifying scaling Van Rooij draws across the course
The lectures aren't a grab-bag — they ascend through levels of system description:
| Level | Lecture(s) | Methodological tools |
|---|---|---|
| Human cognition | L2 (Van Maanen) | Computational modelling, LBA, behavioural prediction |
| Human/AI psychology | L3 (Hortensius), L7 (Trust) | Interaction with AI systems, trust, ethics, digital traces |
| Human/AI networks | L4 (Klein), L5 (Van der Vegt) | Agent-based models, NLP on groups, emergent patterns |
| Human/AI society | L6 (Medical AI / DTs), L1 (CDR/IOS) | Institutions, governments, societal groupings, infrastructure |
The exam will reward answers that locate a phenomenon at the right level and link it to neighbouring levels.
Potential grant-proposal topics per lecture (also exam-relevant)
Van Rooij gives one or two illustrative project ideas per lecture. They are useful as example-fodder for essay answers — pick one of these if asked to illustrate.
| Lecture | Potential research question (per Van Rooij) | IOS link |
|---|---|---|
| L2 — Van Maanen | "Using AI-enhanced image processing, which factors decrease decision time and/or increase accuracy for radiologists examining medical scans? Can we predict which cues work for which doctors?" | Future of Work |
| L3 — Hortensius | "Investigating bias in the YouTube algorithm." | Futures of Democracy, Behaviour and Institutions |
| L4 — Klein | "How might the US presidential election have changed with altered campaigning compositions and stances? How does disinformation alter voting behaviour?" | Future of Democracy, Open Cities, Future of citizen-based initiatives |
| L5 — Van der Vegt | "Predicting riots around football matches with trained AI" (LSTM on event timing). "Predicting violent events from language use on far-right forums" (NLP). | Security in Open Societies |
| L7 — Trust | (Trust + algorithmic legitimacy in public services) | Democracy & Good Governance |
| L6 — Medical AI / DTs | "How do changes in medical AI use alter openness, fairness, security or democratic nature of our future society? How can DTs shape evolution of institutions (link with L4 modelling)?" | Transitions & Wellbeing; Open Cities |
THE MOCK EXAM (as run in class) — with model answers
Reproduced verbatim from Van Rooij's slides, then commented.
Question 1 (L1 — Intro)
"During the introductory lecture, the IOS platforms defined by Utrecht University were discussed. Name two of the fifteen IOS platforms. Describe the aim of each topic you named in one sentence each."
Style: Definitional listing + one-sentence elaboration. Easy marks if you know the 15-platform diagram from L1.
Strategy: pick two platforms you can describe crisply. Suggested safe picks: Open Cities (interdisciplinary research on how cities can be open, accessible and resilient) and Behaviour and Institutions (how behaviour and institutional rules shape each other, including with AI).
Question 2 (L2 — Modelling individual decision making)
"What parameter of the Linear Ballistic Accumulator Model is typically associated with time pressure?"
Model answer: Response caution / threshold / boundary / b.
"According to Palada et al. (2016), which diffusion-model parameter explains why Unmanned Aerial Vehicle operators miss more target ships when these are obscured by clouds than when they are not? Explain your answer."
Model answer: Drift rate / evidence accumulation. Clouds make it more difficult to extract all information, increasing difficulty, decreasing the rate.
Style: Mechanical recall of the methodological lecture and the paper. Knowing which LBA parameter does what is essential.
Question 3 (L4 — Emergence of collective patterns)
"What is a tipping point of an emergent system? Explain and illustrate with an example."
Model answer: Small changes in one parameter have drastic effects on the output. Example part: pick a clear case (climate change tipping; sudden emergence of trust/revolutions; phase transitions in opinion dynamics).
"According to Douven and Hegselmann (2021), how can the presence of free riders impact the performance of truth seekers? Explain this finding."
Model answer: Free riders don't reduce the number of truth-seekers reaching truth, but can have a slightly negative impact on truth-seekers' performance (measured by mean-squared error from truth) — only if campaigners are subtle. Explanation: free riders, when influenced by campaigners, can prevent truth-seekers from reaching truth (dragging them away).
Style: Concept definition + paper-specific empirical claim. Memorise the counter-intuitive D&H findings.
Question 4 (L5 — Attitude & Linguistic Models)
"We talked about using AI to predict events based on human group behaviour. Explain one of the key problems associated with using AI to predict rare events, like terrorist attacks."
No model answer given in the slides — this is the most open-ended question and likely the kind the real exam will lean on.
Strong answer template: - Class imbalance / base-rate fallacy. Terror attacks are extraordinarily rare. Even a model with 99% accuracy and 99% sensitivity will produce overwhelmingly more false positives than true positives (because the base rate is, say, 1-in-a-million). A model with high AUC can still be useless or harmful in deployment. - Connect to van der Vegt et al. (2023) "proceed with caution": proprietary models like Google Perspective have biased measurement, so the false positives concentrate on already-marginalised groups → algorithmic injustice on top of statistical futility. - Conclude: such tools may be useful as prioritisation aids for human analysts (CTAP-25 plus AI dashboard) but cannot stand alone as decision-makers.
Question 5 (L3 — Interaction with AI systems / Hortensius)
"Digital traces are sometimes seen as the goldmine of data for psychological science. Which of the following statements about digital traces is/are true (multiple possible)?"
| Option | Statement | T/F |
|---|---|---|
| A | Digital traces allow for larger and more varied sample populations, potentially moving beyond traditionally WEIRD samples. | TRUE |
| B | The 'digital dossier' can capture only explicit behaviours of individuals, thus limiting its usefulness in reducing experimental bias. | FALSE (digital dossier captures implicit and explicit) |
| C | Digital traces can provide a detailed recording and measurement of behaviours that include contextual factors. | TRUE |
| D | The use of digital traces can potentially raise ethical questions around privacy and data usage (e.g., consent and ownership). | TRUE |
| E | Digital traces have a stable meaning and are not affected by self-report or awareness biases. | FALSE (they're still affected by self-selection and awareness biases) |
Correct answers: A, C, D.
Style: Multiple-correct multiple-choice. Likely 1-2 of these on the real exam. The discriminating reading is paying attention to qualifier words ("only," "not affected by," "stable meaning") — those are usually wrong.
What the mock teaches us about the real exam
- Question 1-style questions (recall + brief explanation) are gimmes if you have the L1 IOS taxonomy and the 5 TRUST letters memorised.
- Question 2-style questions (model parameters + paper-specific empirical mechanism) require naming the parameter AND giving the mechanistic reason. Half marks for one without the other.
- Question 3-style questions (define-a-concept + illustrate) need a clear textbook definition plus a concrete example (not "for example, in society"). Pick the example you'd give before writing the definition.
- Question 4-style questions (open-ended critical) want you to (i) name a phenomenon (class imbalance), (ii) give the mechanism (1-in-a-million base rate → false-positive flood), (iii) connect to a paper (van der Vegt 2023), (iv) conclude with a normative implication. Four moves per answer.
- Question 5-style MCQs test attention to qualifier words.
Likely real-exam structure (educated guess from a 5→9 expansion of the mock)
- 2 questions on methodological foundations: LBA + Palada (L2), Newell time scales + cognitive integration levels
- 2 questions on collective / society modelling: Hegselmann–Krause + D&H findings (L4), Rahwan framework (L3)
- 2 questions on thematic applications: NLP + van der Vegt caution (L5), CDR + TRUST (L1) or G&M six threats (L7)
- 2 questions on hybrid topics: digital traces + Kosinski/Kramer (L3), Medical AI risks (L6)
- 1 integrative question that asks you to map across multiple lectures (e.g. "trace one IOS platform through three lectures' methodological lenses")
This is my best guess, not the real spec.
Cross-lecture synthesis map (for the integrative question)
The single most useful study habit at this point is being able to do this fast:
Pick any IOS platform → name which lectures speak to it → which paper supports each → what's the unifying methodological move?
Worked example for Open Cities (Equity & Diversity pillar): - L1 (Elliott CDR) — corporate digital responsibility for sustainable, equitable urban tech. - L4 (Klein ABM) — agent-based models of pedestrian flows and crowd patterns are literally what Bontje builds. - L5 (Van der Vegt NLP) — sentiment / threat analysis to make cities safer for all residents, not just the dominant group. - L6 (Bontje digital twins) — DTs let you test redevelopment scenarios before implementation, comparing impact across user groups (Fair Transitions). - Methodological move uniting them: simulate before deploy, then audit during deploy — the IOS argument for AI methods is fundamentally about making the consequences of policy visible in advance.
Do this exercise once for two or three platforms and the integrative question is half-written.
Quick self-test
- State Van Rooij's four-level scaling (cognition → psychology → networks → society) and assign each lecture to a level.
- What does the LBA threshold parameter encode? What does the drift rate encode? Cite the Palada paper finding linking workload to one of these.
- Define a tipping point and give a non-trivial example (i.e., not "climate change" alone — explain why climate has tipping points).
- State the Douven & Hegselmann finding about free riders' effect on truth-seekers — and the condition under which it appears.
- Why is class imbalance especially dangerous for AI tools that predict rare events? Tie to van der Vegt et al. (2023).
- Of the five statements about digital traces in the mock Q5 — recite which two are false and why each is false.