W03Full seminar3 hours

Algorithmic Fairness

Mathematical definitions of justice

Can justice be a constraint you optimize under—and who chooses the constraint?

The class computes competing fairness metrics on the same decisions, sees why several cannot hold together, and treats metric selection as a public-policy choice rather than a purely technical fix.

Join the live room

A pipeline archaeology diagram tracing evidence back through an AI system — Visual field note · Week 3 · use this diagram to keep the model inside its surrounding system.

Before class

Complete the short confusion-matrix worksheet and bring a 400-word memo answering: which error should a hiring screen treat as most serious, and why?

After class

Revise your memo after the impossibility result. Name the metric you would choose, the duty it serves, and the recourse required for people it disadvantages.

The promise

By the end of this room…

01Compute and explain demographic parity, equal opportunity, equalized odds, calibration, and predictive parity.
02State the conditions that create fairness impossibility results.
03Connect each metric to the stakeholder and harm it prioritizes.
04Separate measurement uncertainty from normative disagreement.
05Design a decision and appeal process around a selected metric.

Why this week now

Signals, not scene-setting.

Fairness disputes increasingly reach courts and regulators through workflows, vendor roles, and adverse effects—not through a single accepted metric.

Generative and ranking systems make the outcome space less tidy, but the normative choice of who bears error remains.

A metric without an appeal route can make disparity legible while leaving affected people powerless.

Run of show

Provoke → frame → work → argue → synthesize.

Open student room ↗

0:00
provocation
Choose your error
Students choose which error a high-stakes screening system should equalize before seeing the group's base rates or stakeholder roles.
Live activity · choose error
0:15
frame
The metric is a moral commitment
Confusion matrices, base rates, thresholds, calibration, demographic parity, error parity, and the different social worlds each metric assumes.
0:45
discussion
Metric stations
Groups calculate a different fairness condition on one shared decision table, then translate the number into who gains, who waits, and who bears error.
Live activity · metric choice
1:10
frame
Why fairness goals collide
Chouldechova and Kleinberg impossibility results, followed by a practical move: from 'make it fair' to an accountable choice with evidence and recourse.
1:35
break
Break
Ten minutes. Leave the room's unresolved question visible.
1:45
forensics
COMPAS evidence room
Groups reconstruct the ProPublica/Northpointe dispute from the underlying cohort, outcomes, and metrics before writing any moral conclusion.
Deliverable · Claim → metric → evidence → affected party → limitation.Live activity · compas finding
2:25
controversy
ProPublica v. Northpointe
Structured academic controversy: defend one metric, switch sides, then recommend a decision rule and an appeal route for an institution—not a benchmark.
Live activity · fairness recommendation
2:50
synthesis
The sentence behind the number
Complete: 'We chose this metric because the institution has a stronger duty to prevent…'
Live activity · week 3 exit