COMP90087 · The Ethics of Artificial Intelligence
Algorithmic Bias, Accessibility, and Equity
Week 9 defines training-data bias and why it is ethically problematic, using the canonical facial-recognition and AI-recruitment examples the exam rewards, and separates equality from equity and accessibility. The signature short-answer here is the 3-mark “define training-data bias with an example” item, marked one point per component. It also covers the impossibility of a single fair model (Narayanan’s many fairness definitions) and the PNG digital-divide case, giving you both recall and applied material.
What this chapter covers
- 01Training-data bias — data over- or under-representing groups skews model outputs; why it is ethically problematic (unjust discrimination)
- 02Accessibility — usable as effectively by people with disabilities as by those without; universal usability; situational and temporary impairments
- 03Equality vs equity — same resources for all vs resources adjusted so all can actually succeed
- 04The impossibility of a single fair model: Narayanan’s “21 definitions of fairness”; provably incompatible metrics, so some bias is unavoidable
- 05Bias-after-deployment gallery: facial recognition on darker skin, gender bias in AI recruitment (Njoto et al.), word-embedding bias, Amazon’s résumé tool
- 06Three forms of bias in hiring algorithms: bias in datasets, bias in the system, bias in human decisions (Cheong et al.)
- 07The digital divide (access; skills/usage; disparities in returns) and the PNG case (Cheong et al., BRIDGES) — open-source software as a leveller
- 08Mitigations: debias/diversify data, mask sensitive attributes, diverse teams, bias audits, humans in the loop
Short answer: define training-data bias, say why it is wrong, give an example
- +1Mark 1 — definition. Training-data bias occurs when the data used to train a model over- or under-represents certain groups (or encodes past social prejudice), so the model’s outputs are systematically skewed for those groups.
- +1Mark 2 — why it is ethically problematic. It produces unjust discrimination: people are treated systematically worse on the basis of group membership, often invisibly and at scale, and the opacity of the model can hide it.
- +1Mark 3 — concrete example. A facial-recognition system trained mostly on lighter-skinned faces performs worse on darker-skinned faces; or an AI recruitment tool trained on past hires (mostly one gender) down-ranks candidates from an under-represented group. Give one, specifically.
Key terms
- Training-data bias
- When training data over- or under-represents groups, or encodes past social prejudice, so the model’s outputs are systematically skewed — producing unjust discrimination, often invisibly and at scale.
- Equality vs equity
- Equality gives everyone the same resources; equity adjusts resources so that everyone can actually succeed, taking account of and addressing existing inequalities. The exam expects the distinction, not the conflation.
- Accessibility
- Technology is accessible if it can be used as effectively by people with disabilities as by those without. It extends to situational impairments (a busy parent) and temporary ones (a broken arm) — the curb-cut effect.
- The impossibility of a single fair model
- There are many competing mathematical definitions of fairness (Narayanan’s “21 definitions”), some provably incompatible — so no single model can satisfy every reasonable fairness definition at once, and some bias is unavoidable.
- Digital divide
- Van Deursen & Helsper’s three levels: access, skills and usage patterns, and disparities in returns (who benefits most from being online). The PNG case shows Big-Tech pricing excluding the Global South.
- Three forms of hiring-algorithm bias
- Cheong et al.’s split: bias in the datasets, bias in the system (design assumptions/errors), and bias in human decisions — useful for locating exactly where a recruitment tool goes wrong.
Algorithmic Bias, Accessibility, and Equity FAQ
How should I structure the training-data-bias short answer?
As three tight sentences worth one mark each: (1) definition — data over/under-represents groups, skewing outputs; (2) why it’s wrong — unjust, often invisible discrimination at scale; (3) a concrete example — facial recognition performing worse on darker skin, or an AI recruitment tool down-ranking an under-represented group. Keep it to about 40 words and do not pad.
What does ‘the impossibility of a single fair model’ mean?
Researchers have defined many mathematical notions of fairness (Narayanan lists 21), and some are provably incompatible — satisfying one can force you to violate another. So no single model can meet every reasonable fairness definition simultaneously, which means some bias is unavoidable and “fairness” requires a value choice about which definition to prioritise. This is a strong essay point.
What’s the difference between equality and equity here?
Equality distributes the same resources to everyone; equity adjusts resources so everyone can actually succeed, explicitly taking account of existing inequalities. The EqualShareAlgorithm thought experiment makes the point: dividing a resource equally looks fair but becomes inequitable once real needs differ. Accessibility and the digital divide are applied cases of the same equality-vs-equity gap.
Can AI help me revise algorithmic bias for COMP90087?
Yes. Sia can drill the training-data-bias short answer, quiz you on equality vs equity and the fairness-impossibility point, and set fresh bias-after-deployment cases to analyse — explaining each step and checking your reasoning. It mirrors how the University of Melbourne assesses this and does not complete graded work; the subject’s integrity and GenAI rules apply.
Exam move
Over-rehearse the training-data-bias short answer as three one-mark sentences (definition, why-wrong, example) with a precise example you can state fast — it is the signature Week-9 item. Learn the equality-vs-equity distinction and the fairness-impossibility point (Narayanan, incompatible metrics) as clean recall and as an essay argument that fairness is a value choice. Keep a small gallery of real cases (facial recognition, Njoto et al. recruitment bias, the PNG digital divide) so you can supply concrete examples on demand. Because the exam is a closed-book hurdle, make the definitions and one example per concept automatic rather than leaving them to SWOTVAC.
Working through Algorithmic Bias, Accessibility, and Equity in COMP90087? Sia is AskSia’s AI AI Ethics tutor — ask any COMP90087 Algorithmic Bias, Accessibility, and Equity question and get a clear, step-by-step explanation grounded in how COMP90087 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
