AIOS Flashcards
Try to recall the answer before clicking to reveal. Grouped by lecture.
Lecture 1 — Intro & IOS / Elliott CDR
Four bases of an open society?
(1) Openness to diversity of knowledge; (2) openness to emancipatory movements & individual rights; (3) constitutional democracy & rule of law; (4) contestable markets & open borders.
Definition of \"institutions\"?
The building blocks of society — the rules of the game: written rules + associated organisations, and unwritten rules & networks.
Formal vs. informal institution — one AI-relevant example of each?
Formal: the EU AI Act. Informal: user expectations of how a chatbot should behave / journalistic norms about reporting AI-generated content.
Three IOS pillars?
Democracy & Good Governance · Transitions & Wellbeing · Equity & Diversity.
Three lecture types in this course (and what each does)?
Methodological (tools to study AI's impact) · Thematic (examples + mitigation) · Practical (workgroups for the grant).
What does TRUST stand for (Elliott et al. CDR)?
Transparency, Responsibility, Understanding, Stewardship, Truth.
Three central tenets of CDR (Venn diagram)?
Promoting Economic Transparency · Promoting Societal Wellbeing · Reducing Tech Impact on Environment.
What does the intersection of all three CDR tenets represent?
Purpose & TRUST.
The \"Elliott guiding question\"?
"If we permit AI to make life-changing decisions, what are the opportunity costs, data trade-offs, and implications for social, economic, technical, legal, and environmental systems?"
Why is CDR a corporate framework rather than state/individual?
It locates responsibility for AI's societal impact with the firms deploying it, complementing (not replacing) regulators — a governance-by-organisation rather than governance-by-law approach.
How many IOS platforms in total, across the 3 pillars?
15.
Lecture 2 — Modeling individual decision making (Van Maanen) / Palada et al. 2016
What are Newell's four time scales of human action?
Biological band (microseconds to milliseconds) · Cognitive band (100 milliseconds to 10 seconds) · Rational band (minutes to hours) · Social band (days to months). Each higher band is built from the processes of the band below it.
Which time-scale band does Lecture 2 sit in?
The Cognitive band (100 milliseconds to 10 seconds): perception, memory recall, attention, and simple choices.
What are the three levels of integrating cognitive theory into artificial intelligence? (Liefooghe & Van Maanen, 2023)
Anecdotal (general intuitions about cognition) → Computational (specifies what is computed, the inputs and outputs, but not the mechanism) → Algorithmic (specifies the actual cognitive process, for example a memory equation or an evidence-accumulation rule). Each step commits to more mechanism.
Which level of integration gives the best individual-difference predictions, and why?
The algorithmic level. Because it specifies a mechanism with fittable parameters (decay rate, drift rate, threshold), it can capture how each individual user differs. Example from the slides: an algorithmic-level flashcard app (built on the Adaptive Control of Thought-Rational, ACT-R, memory model) produces a larger French-vocabulary test gain (about 7.5) than an anecdotal-level app (about 6.7), and it can also explain each learner's decay rate.
What are the four parameters of the Linear Ballistic Accumulator (LBA) model?
Start point, drawn from a uniform distribution \(U[0, A]\) · Drift rate, drawn from a normal distribution \(N(v, s)\) · Threshold \(b\) · Non-decision time \(t_0\) (time for stimulus encoding plus response production, i.e. everything outside the decision itself).
What is the drift rate in the Linear Ballistic Accumulator (LBA), in plain terms?
The rate at which evidence for a response builds up toward the threshold, i.e. the slope of the accumulator's straight-line rise. It indexes the quality of the evidence, how clearly the stimulus favours one response (its signal-to-noise ratio). A high drift rate means evidence accumulates quickly and reliably, so the decision is fast and accurate; a low drift rate means an ambiguous or degraded stimulus (for example a target hidden behind thick cloud), so accumulation is slow and error-prone. In short, the drift rate is the model's measure of task difficulty / discriminability.
How is a decision produced in the Linear Ballistic Accumulator (LBA)?
Each response option has its own accumulator. Each accumulator starts from a random point in \(U[0, A]\) and rises in a straight line at its own drift rate toward a shared threshold \(b\). The first accumulator to reach the threshold wins, and that option is chosen. Total response time = decision time + non-decision time \(t_0\).
In the simple two-alternative forced-choice (2AFC) task, which manipulation moves the drift rate and which moves the threshold?
Task difficulty (the similarity between target and distractor) moves the drift rate. The speed-versus-accuracy instruction moves the threshold. General rule: the quality of the evidence sets the drift rate; strategic caution sets the threshold.
What is the mechanistic explanation for the speed-accuracy trade-off?
A "focus on speed" instruction makes the person lower the threshold \(b\). The decision is then triggered on less-accumulated, noisier evidence, so responses are faster but more error-prone. The drift rate (evidence quality) is unchanged, only the amount of evidence required changes.
What is the Adaptive Control of Thought-Rational (ACT-R) base-level activation equation, and what drives it?
The activation \(B_i\) of memory item \(i\) is summed over its \(n\) prior uses, where \(t_j\) is the time elapsed since use \(j\) and \(d\) is the decay rate. So activation rises with frequency (more past uses means more terms in the sum) and with recency (a smaller \(t_j\) makes its term larger).
What is need probability in the Anderson & Schooler (1991) rational model of memory?
The probability that an item will be needed again in the near future. The model keeps an item while retrieving it is worthwhile and forgets it when \(p(\text{Activation}) \times \text{Gain} < \text{Cost}\).
Palada et al. (2016): what was the task, and what two factors were manipulated?
Task: a simulated unmanned aerial vehicle (UAV) surveillance task, judging whether ship stimuli were targets or non-targets. Two factors were manipulated: classification difficulty, via cloud opacity (three levels of cloud obscuring the ships), and time pressure / workload, via the number of ships present at once (four levels). (Air traffic control is an analogous applied domain often used to motivate the work, but it was not the actual task.)
Palada et al. (2016): what was the key selective-influence result, and why does it matter?
The best-fitting models (chosen by the Akaike Information Criterion, AIC) showed clean selective influence: difficulty affected only the drift rate (harder, cloud-obscured discriminations slowed evidence accumulation) and workload affected only the threshold (under time pressure, people lowered the threshold to respond faster). Workload did not raise the accumulation rate, there was no "super-capacity" effect. This matters because it shows standard evidence-accumulation models transfer cleanly from simple laboratory tasks to complex, applied settings. (Note: this corrects the common shorthand that "workload lowers both", each factor maps to one parameter.)
Palada et al. (2016): what happened to response time and error rate as workload increased?
Mean correct response time (RT) dropped markedly (over 0.5 seconds faster at very-high versus low workload), while the error rate rose only slightly (about 3%). This is the pay-off of lowering the threshold under time pressure: much faster responses, which minimise missed targets, at only a small accuracy cost.
Palada et al. (2016): how does the model support individual user-modelling?
Fitting the model to each person yields three individual-difference axes: cautiousness (the threshold), processing efficiency (the drift rate) and execution time (the non-decision time \(t_0\)). These classify someone as a good, medium, or poor decision-maker. (The three-axis classification figure shown in the lecture is from Katsimpokis et al., 2020, not from Palada.)
Which Institutions for Open Societies (IOS) platforms does Lecture 2 most directly connect to?
Future of Work (the Transitions & Wellbeing pillar): modelling individual workers in order to allocate, train, and protect them. In/Equality (the Equity & Diversity pillar): re-examining algorithmic decisions in hiring and finance once you know what human deciders' trade-offs actually look like.
Lecture 3 — Autonomy / interaction with AI (Hortensius) / Rahwan Machine Behaviour
Rahwan et al. (2019) — working definition of AI agent?
"Complex and simple algorithms used to make decisions."
Three reasons machine behaviour needs to be studied as its own field?
(1) Ubiquity of AI; (2) complexity/opacity (often closed code/data); (3) beneficial AND detrimental effects.
Rahwan's three scales of machine behaviour?
Individual machine · Collective machine · Hybrid human–machine.
Three modes of hybrid human–machine behaviour?
Machines shape humans · Humans shape machines (engineering) · Human–machine co-behaviour (e.g., Tay).
Rahwan's four domains where machine behaviour matters?
Democracy · Kinetics · Markets · Society.
Three vertices of the interdisciplinary triangle?
Engineering of AI · Scientific study of behaviour · Study of impact of technology — machine behaviour bridges them.
What are digital traces (Rafaeli et al. 2019)?
Records/logs of behaviour (Facebook likes, tweets, browsing, cookies). Contextual data: when, where, how long.
Three advantages of digital traces for psych research?
(1) Bigger / different samples — beyond WEIRD; (2) detailed contextual measurement of behaviour-in-the-wild; (3) "digital dossier" reduces experimental demand bias.
Main bias of digital-trace data?
Self-selection — which platform's users you observe shapes your conclusions.
Kosinski et al. (2013) — input, sample, method?
Facebook likes (binary 1/0) of 58,466 US users (myPersonality app); SVD to 100 components → regression with 10-fold CV.
Three highest-accuracy predictions from Kosinski et al.?
Ethnicity (Caucasian/African-American) AUC 0.95 · Gender 0.93 · Gay (male) 0.88.
Hinds & Joinson (2019) key finding?
With ~300 Likes the algorithm matches a spouse's personality-prediction accuracy (~0.56), beating friends, family, cohabitants.
Matz et al. (2017) — psychological targeting result and key caveat?
Trait-congruent ads ~1.4–1.8× higher conversion. Caveat: Eckles et al. 2018 letter — field targeting studies face internal-validity threats.
Kramer et al. (2014) \"Facebook study\" — what was manipulated and what was found?
Newsfeed valence (more/fewer positive vs. negative posts) for ~689k users. Users' own posts shifted in the manipulated direction. Effect tiny (d≈0.001–0.02) but population enormous.
Why does the small effect in Kramer matter despite being tiny?
At population scale (billions of users × continuous feed) the cumulative behavioural and democratic implications are large; also: ethics — no informed consent.
Cambridge Analytica connects which loop?
Digital traces (Facebook likes) → psychographic prediction → psychological targeting → political behaviour (Brexit / 2016 election). The full Rahwan "Democracy" hybrid loop.
Lecture 4 — Emergence of collective patterns (Klein) / Douven-Hegselmann
Define emergence in one sentence.
Complex social patterns arise from individual interactions and may have properties that do not immediately follow from individuals or their properties — "the sum is more than its parts."
Coleman's bathtub — four moves?
(1) Macro → Situation → Actor (downward); (2) Actor → Selection → Action; (3) Action → Aggregation → Updated Macro; (4) methodological individualism: every macro explanation refers to individual agents.
Five analytic concepts for ABMs?
Emergence · Path dependencies · Tipping points · Non-monotonicity · Direction of effect.
Four \"directions of effect\" between individual motives and emergent pattern?
(1) Aligned (possibly stronger than individual motive — Schelling segregation); (2) Accidental; (3) Opposed (Adam Smith's invisible hand); (4) No individual counterpart (trends, flocking).
Hegselmann–Krause model — what is ε and how does it work?
Each agent's confidence interval. The agent only listens to others whose opinion lies within ε of its own; new opinion = average over those neighbours.
HK update rule (gloss)?
pos_new^i = mean of {pos_old^j : |pos_old^i − pos_old^j| < ε}.
What does the base HK model show about polarization?
Polarization can emerge even when every agent is unbiased and rational — clusters drift outside one another's ε and stop talking.
Three agent types in Douven & Hegselmann (2021)?
Free Riders (social only) · Truth Seekers (pos_new = (1−α)·pos_social + α·τ) · Campaigners (fixed position ρ).
Misinformation vs. disinformation?
Misinformation: aim to make public believe a falsehood. Disinformation: aim to impede/distract from believing a truth. Misinformation logically implies disinformation, not vice versa.
Two counter-intuitive D&H findings without truth-seekers?
(1) More extreme campaign positions hurt the campaigner (isolation outside ε); (2) more/stronger campaigners can impede their own campaign.
What flips when you add truth-seekers?
A subtler campaign (ρ close to τ) outperforms a bold one, and adding more campaigners now helps the campaign.
Main limitations of the model (why it's qualitative not quantitative)?
No real network structure · no substantive arguments exchanged · same ε for everyone · double-counting · heavily idealised.
Why is ABM the \"AI tool\" for L4 phenomena?
Tipping points and path dependencies make analytic solutions intractable and intuitions unreliable; ABMs simulate the micro→macro link explicitly.
Lecture 5 — Attitude & Linguistic Models (Van der Vegt) / threat assessment paper
What is NLP, in one sentence?
Large-scale text analysis: quantify human language (text → numbers) and use linguistic features in AI models to predict outcomes.
Two NLP methods covered in L5?
Supervised machine learning (e.g., Google Perspective API) · Dictionary-based NLP (curated word lists).
Six Perspective API measures?
Toxicity · Severe toxicity · Identity attack · Insult · Profanity · Threat. (Each 0–1, where 1 = 100% of people would agree.)
Study 1 — data and N?
1,909,844 tweets @mentioning all Dutch party leaders (n=22) in 2022, via Twitter Academic API.
Study 1 — main results?
Male politicians get higher toxicity/insults/profanity scores; no gender difference for threats; significant gender × ethnic-minority interactions — female ethnic-minority politicians receive the most threatening tweets.
Van der Vegt et al. (2023) — the methodological warning?
Google Perspective API's "identity-attack" measure under-detects misogynistic content even though gender is in its definition. Using it uncritically under-counts abuse against women → mis-prioritised protection.
Two strengths and two weaknesses of dictionary-based NLP?
Strengths: transparent, interpretable. Weaknesses: labour-intensive to curate; blind to context/irony/sarcasm.
Study 2 (Baele, Brace & Ging 2024) — input and method?
172-word expert-curated dictionary (violent verbs, weapons nouns, dehumanising nouns) applied to 11.7M posts from 33 incel-related platforms.
What is threat assessment?
Estimating the risk of violence (plus seriousness and likelihood) by teams of police, mental-health pros and investigative psychologists, using structured professional-judgement tools.
What is CTAP-25?
A 25-indicator communications threat-assessment checklist (the slides give only the acronym "CTAP-25") that yields a Level of Concern (Low/Medium/High). Indicators include threats, weapons references, end-of-tether language, homicidal ideation, divine-mission belief, "gut reaction," etc.
Three applications of AI in close protection?
OSINT dashboards (collect/summarise) · Sentiment analysis tracking general attitude over time · Prioritisation models that rank incoming messages by predicted call-for-violence score.
Two civil-liberty risks of AI-augmented OSINT?
(1) Proprietary biased models systematically under- or over-flag particular groups (Perspective example); (2) chilling effect on legitimate political speech if false-positive rate is non-trivial.
Which IOS pillars does L5 most directly connect to?
Security in Open Societies (Democracy & Good Governance) and Equity & Diversity (algorithmic justice of moderation tools).
Lecture 7 — Trust in AI / Grimmelikhuijsen-Meijer legitimacy
Three psychological components of trust (in an actor)?
Competence/ability · Benevolence · Integrity.
Algorithm aversion vs. algorithm appreciation?
Aversion (Dietvorst 2015): after seeing an algorithm err, people prefer humans even when the algorithm is on average better. Appreciation (Logg, Minson & Moore 2019): on novel unfamiliar tasks, people prefer algorithmic advice over equivalent human advice. Moderators: visibility of errors, domain familiarity, framing.
Three categories of legitimacy in public-admin theory?
Input (who decided, who was represented) · Throughput (process: transparency, accountability, contestability) · Output (outcome quality: effective, fair).
Grimmelikhuijsen & Meijer's six threats to ADM legitimacy?
Two under each legitimacy type (the paper's own headings, Tables 2-4). Input: (1) erosion of democratic control over algorithmic decision-making; (2) limited responsiveness. Throughput: (3) does not meet standards of procedural fairness (privacy, legal translation); (4) insufficient checks and balances (opacity, blurred accountability). Output: (5) ineffective and inefficient; (6) leads to undesirable outcomes (bias, lost human contact, eroded public values). (The shorthand "deskilling / opacity / bias / privacy / accountability / value-erosion" names the concrete failures, not the paper's top-level six.)
What does \"calibrated institutional response\" mean?
Each threat requires a different institutional mitigation; the paper's own pairings: democratic control → oversight of algorithm 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 fix addresses all six.
The Dutch toeslagenaffaire — what was it and what did it bring down?
The childcare-benefits fraud-detection scandal. The Belastingdienst's algorithmic risk model used ethnic-proxy variables (dual nationality, postal code) → systematic over-flagging of immigrant families → mass wrongful debt collection. Caused the resignation of the entire Rutte III cabinet in January 2021.
Which legitimacy threats showed up most in toeslagenaffaire?
All six, but especially: bias (ethnic proxies), opacity (families couldn't see why flagged), accountability (no one was responsible), and public-value erosion ("hard line" optimisation crushed fairness).
How does L7 connect back to Elliott's TRUST framework (L1)?
Grimmelikhuijsen & Meijer's threats are essentially failures of one or more TRUST letters — opacity = failed Transparency; accountability gap = failed Responsibility; bias = failed Truth; data sprawl = failed Stewardship.
Lecture 6 — Medical AI / Digital Twins (Van Rooij + Bontje) / Wang et al. 2023
Classification vs. stratification in medical ML?
Classification = supervised, predict a-priori group labels (patient/control); identifies which features are most predictive. Stratification = unsupervised, identify hidden subgroups within a population (data-driven phenotyping).
Topic 1 example (Van Rooij) — task, method, performance?
ADHD vs. controls from fMRI inhibition-task activation. Gaussian Process Classifier. Acc 77%, Sens 75%, Spec 80%, ROC AUC ~0.82.
Three risks of classifying psychiatric patients from neural data?
Determinism / wrongful interpretation; inaccurate predictions; malignant use (e.g. insurers denying coverage).
Topic 2 (COVID-19 demographics) — risks?
Discrimination/inequality, wrongful causal attribution (black-box confounders), biological determinism, "what you put in is what you get out."
Topic 3 (ASD stratification) — methods?
Structural morphometry of 53 brain segments → normative modelling + spectral clustering → data-driven clusters with distinct clinical profiles.
Definition of a digital twin (Bontje)?
A virtual, bi-directional model of the physical city; visualises urban processes in real time; supports planning/management/decision-making.
What does \"bi-directional\" mean for a DT?
The twin both reflects sensor data from the city AND feeds decisions back to physical infrastructure (vehicles, traffic lights).
Dutch-DT current-state assets (shared data, models and visualisations)?
3D city models · Dashboards · Simulations · Fieldlabs / practical pilots (the DMI programme). (The slide lists these four under "shared data, models and visualizations.")
DMI future goal?
Move from isolated pilots to open modular reusable systems — reusable building blocks, shared standards, "Digital Twin as a Service," European Digital Twin Appstore.
Bontje's SWOT — one item per cell?
Strengths: scenario testing. Weaknesses: model uncertainty. Opportunities: national DT network with reusable modules. Threats: privacy risks; overreliance on models.
Which IOS platforms does Bontje connect DTs to?
Open Cities (redevelopment scenarios) · Behaviour & Institutions (embedded cognitive models of pedestrians/cyclists) · Fair Transitions (impact visible across user groups).
Why federated + edge learning for traffic DTs (Wang et al. 2023)?
Reconstructed, the Wang paper is paywalled and unverified. The likely rationale: bandwidth (can't stream all raw sensor data), latency (need millisecond responses), privacy (raw data stays local; in standard federated learning only model updates leave the device), resilience (local model keeps working if the cloud is down). Confirm against the paper before relying on these as exam facts.
How does federated edge learning relate back to Elliott's TRUST?
(Conceptual link; the Wang mechanics are a reconstruction.) If it works as described, it operationalises Stewardship (data minimisation) at city scale, and can support Transparency/Truth where standards mandate auditability. It does not by itself solve overreliance on models.
Lecture 8 — Synthesis (Van Rooij)
Van Rooij's four-level scaling?
Human cognition (L2) · Human/AI psychology (L3, L7) · Human/AI networks (L4, L5) · Human/AI society (L1, L6).
Exam logistics?
Fri 2026-05-29, 11:00–13:00, EDUC-BETA (Rupert D extra time). Laptops provided. 9 essay questions in 2 hours.
The "Mock-exam Qn" cards below refer to Van Rooij's 5-question in-class slide mock (reproduced in lecture_08_synthesis.md), not to the 9-question mock_exam.md practice set. The two numbering schemes are different.
Slide-mock Q2 first answer (LBA + time pressure)?
Threshold (response caution, boundary, b). Time pressure leads people to lower their threshold, giving faster but less accurate responses.
Slide-mock Q2 second answer (Palada UAV operators + clouds)?
Drift rate. Clouds obscure target features → difficult to extract evidence → lower rate of accumulation → more missed targets.
Slide-mock Q3 first answer (tipping point definition)?
Small changes in a parameter produce drastic, often non-continuous changes in the output. Examples: climate, revolutions, polarization phase shifts.
Slide-mock Q3 second answer (D&H free riders → truth seekers)?
Free riders don't reduce the number of truth-seekers reaching truth, but can negatively affect truth-seekers' accuracy (sum of squared errors, SSE, from truth) by dragging them toward campaigners. Douven & Hegselmann show this generally; their headline case (Example 3.2) uses a subtle campaign (\( \rho \) near \( \tau \)), where the effect is statistically significant but small (\( \eta^2 \approx 0.004 \)).
Slide-mock Q4 — one key problem of AI for rare-event prediction (e.g. terror)?
Base-rate / class-imbalance fallacy: even a 99%-accurate model produces overwhelmingly more false positives than true positives at a 1-in-a-million base rate; combine with biased measurement (van der Vegt 2023) and the false positives concentrate on marginalised groups. Tools should aid analysts, not replace them.
Slide-mock Q5 — which digital-traces statements are TRUE?
A (bigger/non-WEIRD samples), C (detailed contextual measurement), D (ethical questions about consent and ownership). False: B (digital dossier is implicit+explicit, not only explicit) and E (digital traces still affected by self-selection / awareness biases).
One-sentence cross-lecture synthesis test?
Pick an IOS platform → list which lectures speak to it → which paper supports each → state the unifying methodological move.