Lecture 7 — Trust in AI (Grimmelikhuijsen + Liefooghe)

Paper: Grimmelikhuijsen, S., & Meijer, A. (2022). Legitimacy of algorithmic decision-making: Six threats and the need for a calibrated institutional response. Perspectives on Public Management and Governance, 5(3), 232–242.

Type: Thematic (orange). Where L5 looked at natural language processing (NLP) as an investigative tool used by institutions, this lecture flips the lens: when an institution uses an algorithm to make a decision about a citizen (parole, welfare, fraud detection, immigration), what does it take for that decision to be legitimate?

⚠️ Note: the slide deck for this lecture is not in the course folder, so this summary is built from the obligatory paper, the lecture description in the student manual, and general knowledge of the public-administration / trust-in-government literature. Treat it as scaffold: verify specifics against the paper's text and any slides shared on Teams.

Site numbering: this page is labelled "L7" here, but the student manual numbers Trust in AI as Lecture 6 (obligatory reading 6, Grimmelikhuijsen & Meijer) and Medical AI / Digital Twins as Lecture 7 (reading 7, Wang). Do not be thrown if the exam refers to a different lecture number; the content pairing (Trust with Grimmelikhuijsen & Meijer) is what matters.

Must-know core → Minimum to Pass

Trust = competence/benevolence/integrity (trust literature, not this paper) · aversion (Dietvorst) vs appreciation (Logg) · input/throughput/output legitimacy · six threats (two each under input, throughput, output: democratic control, responsiveness; procedural fairness, checks and balances; ineffectiveness, undesirable outcomes) + calibrated response · toeslagenaffaire (external case; cabinet resigned Jan 2021).

Lecture in one paragraph

Public organisations increasingly use algorithmic decision-making (ADM) for high-stakes decisions (welfare allocation, predictive policing, child-protection risk scoring, immigration triage, tax-fraud detection). These systems can be more efficient and more consistent than human bureaucrats, but they can also fail in ways that erode the legitimacy of the state itself — the Dutch toeslagenaffaire (childcare-benefits fraud-detection scandal) is the canonical case in this course. Grimmelikhuijsen & Meijer's paper isolates six threats ADM poses to institutional legitimacy and argues for a calibrated institutional response matched to the type and depth of each threat — rather than blanket pro- or anti-AI policies.

The lecture also covers the cognitive-science side (Liefooghe): what trust actually is, psychologically, and how humans learn to trust (or distrust) AI expertise — algorithmic appreciation vs. algorithmic aversion.

Key concept 1 — Trust: a psychological definition

Source note: this section is the lecture's psychological strand (Liefooghe) drawing on the wider trust literature, not on Grimmelikhuijsen & Meijer. The obligatory paper does not use the competence/benevolence/integrity model or the aversion/appreciation findings; do not attribute them to it.

Trust in an actor (human or AI) is conventionally decomposed as (Mayer, Davis & Schoorman 1995):

Competence / ability: can it do the task?
Benevolence: does it act in my interest?
Integrity: does it follow norms / values I endorse?

For AI specifically, this classic trust model is adapted: people decide whether to rely on an algorithmic recommendation by combining a competence judgement (accuracy, robustness) with an integrity judgement (transparency, explainability, governance).

Two empirically well-documented attitudes (also from the trust literature, not the obligatory paper):

Algorithm aversion (Dietvorst et al. 2015): after seeing an algorithm make even one mistake, people prefer human judgement, even when the algorithm is on average better.
Algorithm appreciation (Logg, Minson & Moore 2019): in novel tasks with no prior exposure, people often prefer algorithmic advice over an equivalent human's.

Whether you get aversion or appreciation depends on visibility of errors, domain familiarity, framing, and perceived delegation of agency.

Paper 6 — Grimmelikhuijsen & Meijer (2022): Legitimacy of ADM

The framing question

Public administration uses ADM more and more. When does that strengthen and when does that threaten democratic legitimacy? The paper draws on input / throughput / output legitimacy (a standard taxonomy in public-admin theory):

Input legitimacy — Who decided to deploy this algorithm? Were affected citizens represented?
Throughput legitimacy — How does the decision process work? Is it transparent, accountable, contestable?
Output legitimacy — Does it produce good outcomes (effective, fair)?

The six threats (the core of the paper)

The paper organises its six threats as two under each of the three legitimacy types, and matches each with a mitigation strategy: the central argument is that each threat needs a different institutional response. The threat labels below are the paper's own (Tables 2 to 4 and Figure 1, p.240). The "concrete manifestations" column maps the familiar shorthand failures (opacity, bias, privacy, accountability, deskilling, value erosion) onto the paper's actual six, so reproduce the headings below, not the shorthand, on a paper-specific question.

Legitimacy	#	Threat (paper's wording)	Concrete manifestations
Input	1	Erosion of democratic control over algorithmic decision-making	Algorithms bought from third parties or continuously self-updating drift "under the radar" of elected oversight.
Input	2	Limited responsiveness of algorithmic decision-making	Affected citizens cannot feed their preferences into how the system is designed or used.
Throughput	3	Algorithmic decision-making does not meet standards of procedural fairness	Privacy infringement from wide data integration; poor translation of legal requirements into code.
Throughput	4	Insufficient checks and balances	Opacity / black-box decisions (citizens and oversight bodies cannot see why) and blurred responsibility ("computer says no").
Output	5	Algorithmic decision-making is ineffective and inefficient	Large upfront expertise and technology costs; loss of frontline expertise (deskilling) can degrade real-world performance.
Output	6	Algorithmic decision-making leads to undesirable outcomes	Bias and unequal treatment of protected groups; loss of human contact; erosion of public values such as mercy and individualised judgement.

The argument: calibrated institutional response

Blanket positions ("ban all algorithms" or "let civil servants choose") are inadequate: the paper pairs each threat with its own strategy (Tables 2 to 4) and invokes Ostrom's "getting the institutions right" (p.240) to argue there is no single fix. The paper's own pairings:

Erosion of democratic control → strengthen democratic control over the purchase and monitoring of algorithms, especially third-party systems.
Limited responsiveness → build in civic participation and client councils so affected citizens shape design and use.
Procedural fairness (privacy + legal translation) → conduct Data Protection Impact Assessments (DPIAs) and systematically check specific legal requirements.
Insufficient checks and balances (opacity + accountability) → require transparency and explainable AI (XAI) for consequential decisions, and clearly assign and record responsibilities.
Ineffective / inefficient → invest upfront in expertise, and build evaluation methods that reward public value, not only technical efficiency.
Undesirable outcomes → minimise bias, use independent external evaluation, and preserve a right to human contact.

The paper's contribution is the catalogue and the map from threat-type to response-type, not a single magic bullet. (The popular shorthand list, deskilling/opacity/bias/privacy/accountability/value-erosion, names the concrete failures the paper discusses; the six headings above are its actual top-level taxonomy.)

The Dutch toeslagenaffaire: the case that's almost certainly in the exam

(Childcare-benefits fraud-detection scandal, 2010s to 2021)

Source note: Grimmelikhuijsen & Meijer do not discuss the toeslagenaffaire; the paper's own Dutch examples are SyRI and CoronaMelder. The case below is accurate external course context used to illustrate the six threats, not a claim from the paper.

The Belastingdienst (Dutch tax authority) used an algorithmic risk model to flag potentially fraudulent childcare-benefit claims.
Variables used included proxies for ethnicity (dual nationality, postal code) → systematic over-flagging of immigrant and dual-national families.
Flagged families were forced to repay tens of thousands of euros, often wrongly. Many were driven to bankruptcy. Children were taken into care.
The scandal led to the resignation of the entire Rutte III cabinet in January 2021.

Reading the concrete failures against this case (each maps onto one of the paper's six threats: bias and value erosion sit under "undesirable outcomes", opacity and accountability under "insufficient checks and balances", privacy under "procedural fairness"):

Threat	How it manifested in toeslagenaffaire
Deskilling	Officials deferred to algorithm outputs rather than exercising judgement on individual circumstances.
Opacity	Affected families could not see why they had been flagged.
Bias	Algorithm used ethnic-proxy variables → systematic discrimination.
Privacy	Massive integration of administrative data without proportionality.
Accountability	Diffuse responsibility — no one official could be pinpointed; appeal processes were broken.
Public-value erosion	"Hard line on fraud" optimisation eroded values of fairness, family integrity, and proportionality.

This case is the paradigm IOS example of why algorithmic legitimacy is not academic — the legitimacy of an entire government depended on it.

Key concept 2 — Liefooghe's psychological side: how do humans learn to trust AI?

(Sketch — paper-only, no slides.)

The cognitive-psych side of the lecture likely covers:

Mental models of AI expertise: people anthropomorphise AI ("the algorithm knows X") and over- or under-attribute competence.
Calibration of trust to capability: trust should track actual model performance; in practice it tracks recent salient outcomes (one error can collapse trust; one success can inflate it).
Identity / source effects: same recommendation labelled "by AI" vs. "by an expert" lands differently — and the difference is domain-dependent (people accept AI for objective domains, resist it for moral/subjective ones).
Transparency vs. understandability: showing the user a SHAP plot is transparent but not understandable; only when explanation matches the user's mental model does trust calibrate properly.
Procedural trust: in public-admin contexts, perceived procedural fairness of the decision pathway matters as much as the decision itself.

Why this matters for an open society

This lecture sits squarely on the Democracy & Good Governance pillar, with strong links to Equity & Diversity:

Algorithmic decisions by state agencies are how the constitutional democracy & rule of law element of the Open Society definition is increasingly mediated. If those decisions are illegitimate, the rule of law itself is degraded.
The contestability of state action — central to an open society — depends on a citizen being able to understand and challenge how a decision was made. Opacity directly threatens this.
Algorithmic discrimination concentrates the cost of public-sector AI failures on already-marginalised groups, undermining the equity pillar.

This is also the lecture where Elliott et al.'s TRUST (L1) re-enters: Transparency, Responsibility, Understanding, Stewardship, Truth. Grimmelikhuijsen & Meijer are essentially asking how a public institution discharges each of those five letters when it deploys an algorithm.

Likely essay-question angles

"List Grimmelikhuijsen & Meijer's six threats to algorithmic legitimacy. For each, name one institutional mechanism that mitigates it."
"Apply the six-threats framework to the Dutch toeslagenaffaire. Which threat materialised most acutely, and what institutional reform would specifically address it?"
"Distinguish algorithm aversion from algorithm appreciation. How does this distinction complicate the design of human-in-the-loop ADM systems?"
"Trust is conventionally decomposed into competence, benevolence and integrity. Discuss how each of these is harder to establish for an algorithmic decision-maker than for a human one — and which one matters most for public (as opposed to corporate) AI."

Quick self-test

Three psychological components of trust?
Define algorithm aversion vs. algorithm appreciation; give one moderator.
Three legitimacy categories (input/throughput/output) — what does each evaluate?
List the six threats of Grimmelikhuijsen & Meijer.
Map at least four of those six threats onto the toeslagenaffaire.
What does "calibrated institutional response" mean — why isn't a single mitigation enough?
Connect this lecture back to the Elliott TRUST framework from L1.