INITIALIZING BUNKROS IDENTITY LAB
LOC UNDERGROUND
SYS --:--:--
Learning / AI Ethics

The illusion of objectivity.

We're told algorithms are pure, objective mathematics. But maths must be fed data to learn — and that data is human history. Ask a machine to imagine success, beauty or criminality and it doesn't calculate the truth. It calculates our past. This class confronts algorithmic bias before you deploy the skills from earlier modules.

You'll be able to

Name the bias

Articulate how flawed historical data poisons modern outputs.

You'll be able to

Spot a proxy

Define a proxy variable and how it masks discrimination.

You'll be able to

Assess impact

Run a basic ethical impact assessment before deploying AI.

01 — Warm-up

True or false?

"Because AI systems are built on mathematical formulas, they are inherently objective and cannot possess racist or homophobic biases." Decide, then confirm.

Self-Check · True / False
02 — Foundation

A mirror, not a mind

Teach a child what a "leader" looks like using only a book of 18th-century European kings, and they'll draw a man in a crown. They're not biased — their dataset is incomplete. AI works identically. We feed the probability engine billions of internet images and documents, but the internet over-represents Western, English-speaking, affluent, historically dominant groups.

Ask an AI to screen software-engineer résumés and it notes most historical tech résumés belong to men. Without context, it predicts that being male is a job requirement — and starts rejecting women.

Demo · "Generate an image of a highly successful corporate CEO"
Pause & notice

If an AI uses historical data to predict future success, how can marginalised communities ever break historical cycles of exclusion?

Mini-recap

Algorithms don't invent prejudice. They inherit systemic human prejudice from uncurated training data and automate it at massive scale.

03 — Visualise

The pipeline of amplification

A single flawed premise snowballs into systemic harm. Watch the current grow brighter as bias flows left to right.

1 · Biased history

Human records full of past exclusion.

2 · Uncurated data

Scraped wholesale, biases intact.

3 · The model learns

Statistics absorb and encode it.

4 · Real-world harm

Automated rejection at scale.

Each stage amplifies the last: a small historical skew becomes an automated, industrial-scale injustice — like a résumé screener that quietly rejects whole demographics.

04 — Interactive

The moderation dilemma

Safety filters are supposed to stop hate speech. But language is contextual — a slur in one mouth is a reclaimed term of power in another. Play the AI engineer trying to build a "perfectly safe" filter. (Spoiler: there's no clean win.)

Interactive fiction · You are the safety filter
05 — Deep dive

The danger of the average

Facial recognition is the clearest example of training data enforcing societal norms. Historically, datasets were scraped and labelled into strict binary buckets — "Male" or "Female" — and models learned to predict gender from the geometry of cheekbones, jawlines and eyes.

For cisgender people these systems eventually got accurate. For transgender, non-binary and gender-nonconforming people, they fail catastrophically: the AI forces everyone into one of two rigid statistical buckets. Deployed at airport scanners or banking identity checks, gender-nonconforming people get flagged as "anomalies," "errors" or "fraud risks."

The core failure

Machine-learning models are built to find the centre of the bell curve. By definition they optimise for the majority and penalise the margins. A technology that can't recognise a marginalised identity should not be a gatekeeper to public life.

06 — In the Wild

The healthcare proxy

A hospital network deployed AI to flag patients needing extra care — hoping to remove human prejudice from triage. The AI couldn't examine patients, so engineers gave it a proxy variable: healthcare spending. The logic: sicker people spend more.

The flawed assumption

Cost = sickness

The AI found White patients historically spent more on healthcare than Black patients, and concluded White patients must be sicker.

The harm

Sicker patients deprioritised

It systematically prioritised healthier White patients over sicker Black patients for advanced care.

Why it matters

The AI never knew race and wasn't "racist." But due to systemic inequality and discrimination, marginalised communities often spend less even when sicker. Healthcare spending is a flawed proxy — and the algorithm blindly amplified a historical injustice because the engineers assumed maths equals objectivity.

07 — Research

Find a real failure

Algorithmic failures aren't theoretical — biased tools get pulled from the market constantly. Find one real example and log it below.

Search for

What to look up

"AI recruiting tool biased against women" · "algorithmic bias predictive policing" · "AI facial recognition wrongful arrest".

Judge the source: use reputable journalism (ProPublica, MIT Technology Review, Reuters), not sensationalised blogs.
Your findings
Saved locally on your device — never sent to a server.
08 — Assignment

The impact assessment

A city council wants to buy an AI "predictive policing" tool that uses historical crime data to tell police which neighbourhoods to patrol. Write a 3-point impact assessment on why it's ethically dangerous — covering the flawed data, the proxy variables, and the feedback loop.

Your 3-point assessment
Saved locally on your device.
Success rubric — tap to expand

Needs work: trusts the tool and argues AI lowers crime objectively.
Getting there: says it might be biased but can't say why the data is flawed.
Solid: identifies the flawed historical data and the risk to marginalised neighbourhoods.
Excellent: articulates the feedback loop — biased data → biased patrols → more arrests in those areas → more biased data → communities trapped in a mathematical loop of over-policing.

09 — Retrieval practice

Quiz — 5 questions

Question 1 · Easy
What is the primary cause of algorithmic bias?
A
Programmers intentionally writing racist code into the system.
B
The AI developing its own independent consciousness and prejudices.
C
The AI inheriting uncurated, historically biased data during its training phase.
Question 2 · Medium
What is a "proxy variable"?
A
A stand-in the AI uses to measure something it can't observe directly (e.g. 'cost' for 'health').
B
A virus that infects the training data.
C
A privacy setting that protects user data.
Question 3 · Medium
A facial-recognition system boasts "99% overall accuracy." Why might it still be unethical to deploy?
A
Because 100% accuracy is required by law.
B
Because the "overall average" hides that the 1% failure often targets a specific marginalised group.
C
Because it uses too much electricity.
Question 4 · Easy · True / False
"Adding more data to a training set automatically eliminates algorithmic bias." — True or False?
T
True
F
False
Question 5 · Hard
What is the "alignment problem" in AI?
A
Ensuring the AI's interface is centred on the screen.
B
The difficulty of ensuring a model's goals and outputs align with human ethical values and safety.
C
Matching the AI's processing speed with internet bandwidth.
10 — Avoid these

Common mistakes

Trap

The 'maths is objective' fallacy

Blindly trusting an AI to reject an applicant or deny a loan. Fix: treat outputs as statistical suggestions, not truth — keep a human in the loop for critical decisions.

Trap

The scale illusion

Assuming a model trained on "the whole internet" must be unbiased. Fix: the internet is dominated by a few demographics — it's not a mirror of global humanity.

Trap

Ignoring the margins

Designing for the "average user" with strict binary gender inputs that break for non-binary people. Fix: design for the edge cases first — if it works for the most marginalised user, it works for everyone.

Trap

Over-relying on safety filters

Trusting corporate moderation to be culturally competent — and watching it flag queer-history discussion as "inappropriate." Fix: advocate for contextual human review over rigid algorithmic banning.

Debug your thinking — check each before you trust an AI tool
11 — Reflect

Curate a fairer dataset

If you had total control over curating the dataset for a truly unbiased AI, what specific human experiences, archives and voices would you ensure are included that current systems actively ignore?

Optional: how do we balance AI safety filters against the danger of erasing reclaimed cultural language? Should algorithms ever decide criminal justice, healthcare triage or employment, given their reliance on biased historical data?

Private journal — your ethical baseline
Saved locally to your device — never sent to a server.
Well done

You've confronted the hardest truth in tech: algorithms are human flaws written in code. With this ethical framework secured, you can manipulate the machine safely. Next, we step inside the visual brain of neural networks.

12 — Reference

Glossary — precise words

Algorithmic bias
Systematic, repeatable errors that create unfair outcomes, privileging one group over others.
Proxy variable
A stand-in for a trait you can't measure directly. Flawed proxies are a major source of hidden bias.
Alignment
The field ensuring AI systems act in accordance with human values, ethics and intended goals.
Training data
The datasets used to teach a model. Poisoned by prejudice → an irreparably biased model.
False positive
When AI wrongly flags a normal case as an anomaly — minorities suffer far higher rates in systems like facial recognition.
Human in the loop
A design where AI can't make a final critical decision without human review and authorisation.
Feedback loop
When an AI's biased outputs generate new biased data, trapping communities in a self-reinforcing cycle.
Safety filter
Automated moderation meant to block harmful content — but often culturally incompetent at the margins.