Algorithmic dilemmas: bias, judgement and accountability

A ranking can be calculated exactly as designed and still cause problems. A system optimized for viewing time may give a viewer content they keep watching without serving their long-term interests. A hiring system may repeat a consistent rule that disadvantages applicants whose circumstances differ. We will distinguish mechanical correctness from justified decisions. The aim is to make a judgement supported by the mechanism, evidence and a clear criterion. The real TikTok case illustrates the tension, but our examples remain bounded by what the sources actually establish.

Bomfim et al., Digital Society for the IB Diploma, Hodder, 2022. Printed pages cited per slide. Original paraphrase. Textbook connection: pp. 86.

Algorithms can reduce repetitive work and apply the same procedure across many cases. Consistency is valuable, but consistently making an unsuitable decision remains a problem. Human judgement can supply context absent from the input, yet human reviewers can also be biased. A meaningful review process requires access to the relevant information and power to overturn an output. Placing a person beside a system is insufficient if they simply accept every score. Ask who makes the final decision and what evidence they use.

Bomfim et al., Digital Society for the IB Diploma, Hodder, 2022. Printed pages cited per slide. Original paraphrase. Textbook connection: pp. 86–87.

Algorithmic bias concerns systematic skew or unfairness in decisions. The textbook identifies biased or incomplete data and choices made by developers. We can examine those choices more precisely. A location variable may stand in for access to resources rather than the ability we want to measure. A target based on previous approvals can repeat past exclusion. A threshold can create different error rates across groups. More data does not automatically solve these problems if it preserves the same omissions or labels.

Bomfim et al., Digital Society for the IB Diploma, Hodder, 2022. Printed pages cited per slide. Original paraphrase. Textbook connection: pp. 87.

These numbers are invented to explain representation. They are not a finding about any company. Suppose a model is tested mostly on Group A. A high overall accuracy can coexist with poor performance for Group B because B contributes fewer test cases. Adding more examples from A increases the size without correcting B’s representation. A fair evaluation needs relevant data and disaggregated results, while taking care with small samples and privacy. Representation matters, but it is only one source of bias.

Bomfim et al., Digital Society for the IB Diploma, Hodder, 2022. Printed pages cited per slide. Original paraphrase. Textbook connection: pp. 87.

Imagine a school choosing volunteers for a media event. It ranks students using the number of previous events they completed. That count is a proxy for reliability. A student who was never offered an event cannot build the same count. The rule may confuse opportunity with willingness or ability. This is a synthetic scenario. To judge the rule, we would need information about access and performance, not just the final ranking. Explain the link between the proxy and the target rather than assuming a numerical score measures the desired quality directly.

Bomfim et al., Digital Society for the IB Diploma, Hodder, 2022. Printed pages cited per slide. Original paraphrase. Textbook connection: pp. 87.

If the first ranking gives one creator more exposure, that creator has more opportunities to receive views and reactions. If the next ranking uses those totals, the original advantage can grow. This is a possible feedback mechanism, not proof of unfairness in every recommender. A creator might also receive exposure because people find the work useful. We need to examine the rule and compare appropriate measures. Explain each link and identify what could interrupt the loop, such as reserving opportunities for items with little previous exposure.

Bomfim et al., Digital Society for the IB Diploma, Hodder, 2022. Printed pages cited per slide. Original paraphrase. Textbook connection: pp. 87–89.

A teacher, moderator or employer might accept a score without asking how it was produced. Over time, that habit can reduce the use of independent judgement. The issue is how people interact with the system, not simply whether software exists. Useful safeguards include recording a reason for a consequential decision, checking cases near a threshold and allowing a person to challenge incorrect input. These are proposals to evaluate, not guarantees. Each requires staff, time and information, and may create new privacy or workload costs.

Bomfim et al., Digital Society for the IB Diploma, Hodder, 2022. Printed pages cited per slide. Original paraphrase. Textbook connection: pp. 86–88.

A black box algorithm produces an output whose internal reasoning is difficult to understand or explain. The textbook focuses on complex learned systems. A simple rule can also be opaque to the person affected if the operator refuses to disclose it. These are different barriers. Publishing a formula may help with secrecy, but a complex model may need other explanations and testing. A black box can still produce useful predictions. The question is whether the available evidence and explanation are adequate for the consequence of the decision.

Bomfim et al., Digital Society for the IB Diploma, Hodder, 2022. Printed pages cited per slide. Original paraphrase. Textbook connection: pp. 88.

A programmer, regulator and affected student need different levels of explanation. A technical model description may be useless to someone trying to correct their record. Transparency can disclose the objective, data source, decision procedure or reasons for a particular output. It can help people contest an error. It cannot by itself establish that the objective is fair or that the system causes no harm. Ask what has been disclosed, to whom, and what action that person can take with the information.

Bomfim et al., Digital Society for the IB Diploma, Hodder, 2022. Printed pages cited per slide. Original paraphrase. Textbook connection: pp. 88.

Accountability concerns responsibility for a decision and what happens when it goes wrong. A provider might explain a rule yet leave nobody responsible for correcting an unfair outcome. A review process needs a responsible decision-maker, a record of what happened and an effective route to remedy. Different organizations may share responsibility for data collection, model development and deployment. Our question is not simply who to blame. It is who can inspect, change or stop the process, and whether the affected person can obtain a meaningful response.

Bomfim et al., Digital Society for the IB Diploma, Hodder, 2022. Printed pages cited per slide. Original paraphrase. Textbook connection: pp. 88.

Purpose: Technical failure, causal explanation and real-world harm are different claims. Say: This creates a bounded case beyond the feed. Specify the task, population, image conditions and error type. A measured disparity warrants investigation; it does not alone establish the designer’s intent or every downstream outcome. Ask: Explain what extra evidence is needed to move from a measured error difference to a judgement about a specific deployment. Practice feedback: use the worked exhibit to check the distinction; keep the first attempt and append a correction. Secure: apply the distinction to a different venue. Partial: isolate the disputed operation and compare the two cases. Unclear: reconstruct the first worked example before retrying. Later return: next lesson, close notes and explain this distinction in a library or community-centre system. These are formative practice responses, not registered assessment items.

Textbook layer: Hodder 2022, supplied chapter pp. 87–89. Original classroom examples are fictional. Historical primary evidence checked 11 Oct 2026: https://www.nist.gov/news-events/news/2019/12/nist-study-evaluates-effects-race-age-sex-face-recognition-software

This is a historical primary source from the organization operating the service. It is useful for the provider’s explanation of inputs and purpose. Its origin also matters because the provider has an interest in confidence in the service. Use the source for what it supports, not as independent proof that the system is safe or unbiased. The exact private implementation and present weights remain outside this account. Open the case file for the locator, provenance and limitations before making a broader claim.

TikTok, 18 June 2020, provider account. https://newsroom.tiktok.com/en-us/how-tiktok-recommends-videos-for-you Textbook connection: pp. 86–89.

The Commission examines risks from the design of the service, including recommendation, scrolling, autoplay and notifications. Its purpose and evidence differ from the provider’s product explanation. This can broaden an inquiry from how a feed works to whether design choices create risks. Keep the word preliminary attached to the finding. Do not attribute every effect to ranking alone or report the investigation as a final judgement. Compare what each source investigates and the evidence available to it.

European Commission, 6 February 2026, preliminary findings. https://digital-strategy.ec.europa.eu/en/news/commission-preliminarily-finds-tiktoks-addictive-design-breach-digital-services-act Textbook connection: pp. 86–89.

Suppose three items receive more views than the others. That is an observation. A rule favouring previous engagement could explain how the pattern develops, but we need evidence about the rule and exposure. Claiming the operator intended to silence another creator adds a further inference. An uneven result does not establish that intention. A source describing a goal, an internal document or another suitable record may help investigate purpose. Keep these levels separate so an inquiry remains critical without exaggerating its evidence.

Bomfim et al., Digital Society for the IB Diploma, Hodder, 2022. Printed pages cited per slide. Original paraphrase. Textbook connection: pp. 87–89.

A conditional judgement states what you recommend, why and under which conditions. For example, a discovery allocation could be justified if testing shows that it improves access for less-seen creators without exposing users to unacceptable content. That proposal could reduce immediate predicted engagement and requires eligibility checks. Transparency may help evaluate it, but outcome evidence is still needed. We have modelled the structure of reasoning on a different example. Your independent response must choose evidence and weigh its own conditions.

Bomfim et al., Digital Society for the IB Diploma, Hodder, 2022. Printed pages cited per slide. Original paraphrase. Textbook connection: pp. 89.

The local practice page contains original partial questions aligned to the current guide. Paper 1 practice asks you to explain or evaluate a system using the case. Paper 2 practice asks you to use sources and integrate relevant knowledge, including direct comparison where the question requires it. Naming bias or listing two sources is insufficient. Explain how the system works and what each piece of evidence supports. Today’s practice is formative. It is not a full paper or a basis for predicting a final grade.

IB Digital Society guide, local April 2026 revision, sections 3.1H, 3.2A–E and 4.2A–D. Assessment authority. Textbook connection: pp. 89, 91.

Close the deck. A community centre allocates its most visible booking slots using previous attendance. Explain how this rule could help organizers and how it could disadvantage a new group. Identify a source you would need and an intervention you would test. This changed case checks whether the knowledge transfers. After answering, return to the relevant teaching slides and correct one reasoning gap. The October 13 source synthesis and October 14 redesign give further opportunities to revisit the same chapter ideas.

Bomfim et al., Digital Society for the IB Diploma, Hodder, 2022. Printed pages cited per slide. Original paraphrase. Textbook connection: pp. 86–91.