W02Full seminar3 hours

Data Ethics & Bias

Curating datasets and hidden proxies

If the model is a mirror of the data, who curated the reflection?

Students trace harm before modelling begins: problem formulation, consent, categories, labels, proxy targets, collection, documentation, downstream copies, and feedback loops.

Join the live room

A data-to-decision pipeline showing where choices and harms enter — Visual field note · Week 2 · use this diagram to keep the model inside its surrounding system.

Before class

Bring a dataset or data source you have used. Sketch its origin, labels, missing populations, intended use, and one downstream use its creators may not have anticipated.

After class

Draft a one-page datasheet for the dataset you brought. Mark one field as a hard governance gate and explain who has authority to enforce it.

The promise

By the end of this room…

01Locate distinct sources of harm across the machine-learning lifecycle.
02Distinguish a construct from its operational proxy and identify where the relationship breaks.
03Interrogate dataset categories, annotation labour, consent, provenance, and downstream reuse.
04Build a datasheet that functions as a governance boundary rather than a documentation ritual.
05Propose a remedy proportionate to both the original data practice and its model descendants.

Why this week now

Signals, not scene-setting.

Training data is now a governance surface: provenance, consent, licensing, deletion, and model remediation travel together.

Synthetic and generated data do not remove upstream judgments; they can harden them while making lineage harder to see.

The operational question has shifted from whether a dataset is biased to who can inspect, contest, repair, and retire it.

Run of show

Provoke → frame → work → argue → synthesize.

Open student room ↗

0:00
provocation
Delete the data—or delete the model?
Students vote on whether a model trained on data collected without valid consent may remain in service after the source data is deleted.
Live activity · data deletion vote
0:15
frame
The bias cascade
Problem formulation → collection → measurement → representation → aggregation → evaluation → deployment → feedback. Each stage creates a different repair obligation.
0:45
discussion
Construct, proxy, missing world
Pairs unpack a familiar target—engagement, risk, employability, health need—and name what the proxy rewards, excludes, and teaches downstream systems.
Live activity · proxy audit
1:10
frame
Datasets are institutions
Categories have histories; labels contain labour; consent is contextual; documentation and versioning decide whether repair can propagate.
1:35
break
Break
Ten minutes. Leave the room's unresolved question visible.
1:45
forensics
Dataset archaeology
Groups examine Tay and ImageNet through four lenses: provenance, taxonomy, labour, and deployment. Each group builds a causal chain rather than a list of bad outcomes.
Deliverable · A pipeline map with one high-leverage intervention and one unresolved trade-off.Live activity · dataset archaeology
2:25
controversy
Remediation tribunal
The room decides what repair requires when data was collected illegitimately but a socially useful model already exists.
Live activity · remedy ranking
2:50
synthesis
A datasheet that can stop deployment
Each group writes the one datasheet field that should trigger review, refusal, or retirement rather than merely record a caveat.
Live activity · week 2 exit