COMP90087 · The Ethics of Artificial Intelligence
Transparency and Explainability
Week 8 separates transparency (openness of the system/process) from explainability (giving an audience intelligible, actionable reasons) and argues why explainability is ethically required — it lets affected parties contest a decision and lets designers detect discriminatory proxies. The exam’s opaque loan-approval item tests exactly this: high accuracy does not discharge the obligation to explain. Know Burrell’s three sources of opacity, Creel’s three levels of transparency, and Miller’s account of explanation.
What this chapter covers
- 01Transparency vs opacity (“black box”): why transparency is wanted differs by stakeholder (data scientist, regulator, affected consumer)
- 02Burrell’s three sources of opacity (exam-core): corporate/state secrecy, technical illiteracy, intrinsic characteristics of machine learning
- 03Creel’s three levels of transparency (exam-core): functional, structural, implementation
- 04Explainability = giving intelligible, actionable reasons to an audience; it is audience- and purpose-relative
- 05Why explainability matters ethically: enables contestation, accountability and detection of discriminatory proxies — accuracy alone cannot
- 06Miller’s account: an explanation answers a “why?” question; good explanations are selective, salient and simple; contrastive (why P rather than Q)
- 07Trust: reliance plus goodwill (Mayer 1995, Jones 1996); warranted vs unwarranted trust/distrust; calibrated trust
- 08The explainability–accuracy trade-off and Rudin’s counter-position: for high-stakes decisions, use inherently interpretable models
Short answer: why explainability matters for an opaque loan model
- +1Define explainability correctly. Explainability means giving the affected party intelligible, actionable reasons for the decision — reasons a person can understand and act on — not merely publishing the code or the accuracy figure. This is the option examiners reward.
- +1Say why it matters ethically. Without such reasons the applicant cannot contest the decision, the bank cannot be held accountable, and neither party can detect whether the model is relying on a discriminatory proxy (e.g. postcode standing in for race). Contestation, accountability and proxy-detection are the three payoffs.
- +1Reject the distractors. High accuracy does not discharge the obligation — a model can be accurate and still give no actionable reason — and explainability is not the same as being open-source, having a low feature count, or being transparent about the training data alone. Name Burrell (opacity sources) or Creel (transparency levels) to show the deeper framing.
Key terms
- Transparency vs explainability
- Transparency is the openness of the system or process; explainability is the ability to give intelligible reasons to a particular audience. Interpretability is how far a human can understand the model’s mechanism. Explainability is audience- and purpose-relative.
- Burrell’s three sources of opacity
- Corporate or state secrecy (justified by security/IP but weighed against rights to information), technical illiteracy (a layperson may not understand even full disclosure), and intrinsic characteristics of machine learning (deep nets are hard to interpret).
- Creel’s three levels of transparency
- Functional (the high-level algorithm/logic implemented), structural (how that algorithm is realised in code), and implementation (how the code actually runs on hardware and is deployed).
- Actionable, intelligible reasons
- The core of explainability: reasons an affected party can understand and act on, which is what enables contestation, accountability and detection of discriminatory proxies — something raw accuracy cannot provide.
- Contrastive explanation (Miller)
- We explain not events per se but “why event P (the target) rather than Q (the contrast case).” Good explanations are selective, salient and simple, responding to what surprises or matters to the explainee.
- Calibrated trust
- The goal in which trust tracks trustworthiness. Trust is reliance plus goodwill (Mayer 1995, Jones 1996); the warranted/unwarranted trust-and-distrust grid distinguishes justified from misplaced trust.
Transparency and Explainability FAQ
Why isn’t a highly accurate model good enough on its own?
Because accuracy and explainability are different things: an accurate model can still give an affected person no reason they can understand or act on. Without intelligible, actionable reasons, decisions can’t be contested, no one can be held accountable, and a discriminatory proxy can hide behind the accuracy figure. Procedural fairness can outweigh raw accuracy — that is the core of the loan-approval item.
What’s the difference between transparency and explainability?
Transparency is openness — how much of the system or process is visible. Explainability is giving a particular audience intelligible reasons for a decision, so it is audience- and purpose-relative (what a data scientist needs differs from what a denied consumer needs). Interpretability, a third term, is how far a human can grasp the model’s mechanism. The exam expects you to keep these distinct.
How do I keep Burrell’s and Creel’s lists straight?
Burrell is about why systems are opaque — three sources: corporate/state secrecy, technical illiteracy, intrinsic ML characteristics. Creel is about levels of transparency — functional, structural, implementation. A quick mnemonic: Burrell = barriers (why you can’t see in), Creel = layers (how deep the seeing goes). Both are exam-core, so learn them as named triples.
Can AI help me with the transparency and explainability material?
Yes. Sia can drill the Burrell and Creel triples, quiz you on the transparency/explainability/interpretability distinctions, and rehearse the loan-approval short answer so you avoid the accuracy trap — explaining each step and checking your reasoning. It mirrors how the University of Melbourne assesses this and never writes graded work; the subject’s GenAI and integrity rules apply.
Exam move
Nail two named triples — Burrell’s three sources of opacity and Creel’s three levels of transparency — and a one-line separation of transparency, explainability and interpretability, because attribution and distinction MCQs are near-certain. Then over-practise the loan-approval short answer: explainability = intelligible, actionable reasons enabling contestation, and accuracy does not discharge it. Add Miller’s “explanation answers a why-question, selective/salient/simple, contrastive” for depth and Rudin’s interpretable-models counter-position for essays. Keep these in active recall for the closed-book hurdle exam rather than leaving them to SWOTVAC.
Working through Transparency and Explainability in COMP90087? Sia is AskSia’s AI AI Ethics tutor — ask any COMP90087 Transparency and Explainability 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.
