Assignment 2: Audit a deployed suggestion feature
A comparative audit of a real, currently deployed prediction or suggestion feature, using the fixed-input method practised across weeks 6 through 9
Due: 7 May 2027 · Weight: 35%
The brief#
Pick one deployed suggestion feature. Run the same inputs through it, and at least one comparable alternative, and say honestly what you find.
This is the method from the week 3, 6, and 9 labs, done properly: a fixed list of inputs, recorded systematically, compared across more than one system so a finding is a pattern rather than one screenshot. Pick a feature from the second half of the course so far — a smart-reply system, a conversational model, an image-completion tool, or a forecast interface — and audit it the way those labs audited their smaller targets, at assignment depth rather than lab depth.
“Depth” means: enough inputs that a single odd result doesn’t drive your conclusion, a comparison system so your findings are about the feature and not about interfaces in general, and a clearly argued answer to the question every audit in this course eventually asks — given what the feature actually does, whose interest does it appear to be designed around?
This assignment is also the seed of your final project. Choosing a feature you’re willing to keep working with for another three weeks is worth more than choosing the flashiest one.
What you submit#
- your input list and raw results, as an appendix or linked spreadsheet
- a written audit, 1,200–1,800 words, structured around your findings rather than around a feature-by-feature walkthrough
- one paragraph naming what you’d need to do to turn this into the final project’s fuller audit
The spec
The brief poses the problem and leaves room for a response. The spec is the fixed part: what a marker checks. Some lines can be checked mechanically; the rest call for judgement.
- a fixed set of at least ten inputs, run identically across at least two comparable interfaces or products in the same domain
- a record of exactly what each interface returned, not a paraphrase of what it "generally" does
- a stated judgement of whose interest the feature's design choices appear to serve, argued from the recorded evidence rather than asserted
- the audited feature falls in the territory covered by weeks 6 through 9 (suggested replies, chat continuations, image completion, or forecast interfaces)
How it is marked
| Criterion | Weight |
|---|---|
| Method is systematic and reproducible | 30% |
| Findings are specific and evidenced, not asserted | 40% |
| The audit states who the feature's design appears to serve | 30% |