Example 1 — Applying the concept of power

Explain how the concept of power applies to the use of ranking algorithms by social media platforms. [4]

Ranking algorithms determine which content each user sees, so the platform decides what information reaches whom — a form of power over public attention exercised without any formal political authority. ✓ This power is asymmetric: the platform can observe and adjust the system continuously, whereas users cannot see the ranking criteria, cannot verify why they were shown something, and have no route to contest it. ✓ It also extends beyond users. Publishers and small businesses that depend on the platform for reach are effectively governed by a change they do not control — a ranking adjustment can remove most of an organisation's audience overnight, as many news publishers experienced when Facebook deprioritised news content in its feed. ✓ Accountability is therefore weak: the party with the power to change the system is not the party that bears the consequences of it, and no external body reviews the decision. ✓

4/4.

The structure is the lesson. Identify the power, show it is asymmetric, show who else is affected beyond the obvious group, and end on accountability. Note that "who can change it" and "who is affected by it" being different parties is the whole argument.

Example 2 — Two perspectives that both hold

Discuss the use of facial recognition technology in public spaces. [6]

Facial recognition in public space is a question of power, space and the conflict between privacy and security.

The case for. Police forces argue that identifying suspects in crowds prevents serious harm and does so faster than any human process. Deployments at large transport hubs and stadiums are defended on the grounds that public space carries no expectation of anonymity, and that the technology only flags people already on a watchlist. Operators point to consented uses too — border control gates that travellers choose to use.

The case against. Public space is exactly where anonymity has historically been possible, and continuous identification changes what it is to be in public: people behave differently when identifiable, which has a measurable chilling effect on protest and assembly. Accuracy is not evenly distributed — several evaluations, including the US National Institute of Standards and Technology's 2019 study, found substantially higher false-match rates for women and for people with darker skin, so the burden of error falls on groups already over-policed. Consent is effectively impossible: you cannot opt out of a public square. And function creep is well documented, with systems deployed for one purpose extended to others without fresh authorisation.

Judgement. The technology's acceptability depends far less on the technology than on the governance around it — whether watchlists are independently audited, whether error rates are published and disaggregated, whether there is a route to contest a false match, and whether use is limited by law to defined purposes. Where those conditions hold, targeted deployment is defensible. Where they do not, the asymmetry of power is too great, and the harms fall on people with the least capacity to challenge them.

Why this reaches the top band. Both sides are argued at their strongest, not caricatured. There is a named, dated source for the accuracy claim. Concepts are named and used. And the judgement is genuinely conditional — it says what would have to be true for the technology to be acceptable, which is a far stronger conclusion than approval or condemnation.

Example 3 — Explaining algorithmic bias precisely

Explain how bias can arise in an automated recruitment system. [4]

The system is trained on historical hiring data — records of which past applicants were selected. ✓ If past hiring favoured a particular group, that pattern is present in the data, and the model learns it as though it were a legitimate predictor of success rather than a record of past preference. ✓ Bias can also enter through proxy variables: even with protected characteristics removed, a feature such as postcode, university attended or a gap in employment can correlate strongly with them, so the model reproduces the same effect indirectly. ✓ The result is applied at scale and with an appearance of objectivity — decisions are attributed to a neutral system rather than to a person, which makes the bias harder to notice and much harder to contest. ✓

4/4.

Note what is absent: any suggestion that someone did this deliberately. Attributing algorithmic bias to malice is the standard weak answer. The marks are for the mechanism — historical data, proxy variables, scale, and the false appearance of neutrality. The last point is the one strong candidates include and others miss.

Example 4 — Evaluating an intervention

Evaluate the effectiveness of requiring users to give consent before their data is collected. [5]

Strengths. Consent requirements establish that data about a person belongs, in some sense, to that person, and they create a legal basis on which collection can be challenged. They have forced disclosure: organisations must now state what they collect and why, which makes practices visible that were previously invisible and gives regulators something to enforce against.

Limitations. In practice consent is rarely informed. Agreements are long, written in legal language, and presented at the moment the user wants to do something else, so they are not read. Consent is usually bundled — accepting the service means accepting all processing, including purposes unrelated to the service — and there is frequently no meaningful alternative, since declining means losing access to a platform that peers, employers or public services rely on. Consent fatigue from repeated prompts leads people to accept reflexively, so the mechanism trains the very behaviour it was designed to prevent. It also places the burden of protection on the individual, who is least equipped to assess the risk.

Judgement. Consent is necessary but not sufficient. It works as a transparency and accountability mechanism — it gives regulators a standard to enforce — but it fails as a means of individual protection, because the conditions for genuine choice are largely absent. Effective regulation therefore has to constrain what may be collected at all, regardless of consent, rather than relying on individuals to negotiate terms they cannot change.

Why this scores. The limitations are specific and mechanical rather than vague complaints. The judgement distinguishes what the intervention is good for from what it fails at, which is what evaluate asks. "Necessary but not sufficient" is a genuinely useful formula in this subject — provided you then say sufficient for what.

Educerie · original worked examples written against the published IB syllabus structure for Digital Society Topic 1, first assessment 2024. Original text; the NIST 2019 facial recognition evaluation is cited as a real published study. Last reviewed 5 September 2026.