L1
·
Quiz
·
Lab
L2
·
Quiz
·
Lab
L3
·
Quiz
·
Lab
L4
·
Quiz
·
Lab
Module Test
Module 5 · Lesson 1

The Score That Follows You

Credit algorithms, FICO, and what your financial data actually decides
When an algorithm approves or denies your loan, what is it really measuring — and who built the rules?

In November 2019, Apple launched its Goldman Sachs–backed credit card. Within weeks, software engineer David Heinemeier Hansson posted on Twitter that his Apple Card credit limit was 20× higher than his wife's — despite her having a better credit score and sharing assets. Steve Wozniak, co-founder of Apple, reported the same pattern with his wife. The New York Department of Financial Services opened a formal investigation. Goldman Sachs stated the algorithm did not use gender — but regulators found it used proxies that correlated with gender, producing discriminatory outcomes anyway.

How Credit Scoring Became Algorithmic

The FICO score — introduced by Fair Isaac Corporation in 1989 — turned creditworthiness into a three-digit number computed entirely by an algorithm. Today approximately 90% of top U.S. lenders use FICO scores. The score weighs five categories: payment history (35%), amounts owed (30%), length of credit history (15%), new credit inquiries (10%), and credit mix (10%). These weights were set by analyzing historical lending data — which itself encoded decades of discriminatory lending practices.

In 2007, researchers at the National Consumer Law Center documented that African-American and Latino borrowers were systematically steered into subprime mortgages even when they qualified for prime rates. The algorithm was trained on patterns produced by human discrimination, so it reproduced those patterns at scale and with the veneer of mathematical objectivity.

Modern lenders go far beyond FICO. Upstart, a lending platform, incorporates over 1,600 variables including your education, your job title, and how long you spent filling out the application. ZestFinance has used social-media behavior as a proxy for creditworthiness. These systems are largely unregulated and their exact inputs are trade secrets.

Documented Impact

In 2021, the Consumer Financial Protection Bureau (CFPB) warned that algorithmic credit models using "alternative data" — shopping patterns, phone usage, location — may systematically disadvantage protected classes. A 2022 University of California Berkeley study found that fintech mortgage algorithms still charged Black and Latino borrowers 11 basis points more than white borrowers with identical risk profiles.

The Opacity Problem

The Equal Credit Opportunity Act (ECOA) requires lenders to give "specific reasons" for adverse actions — but regulators have struggled to apply this to black-box models. When Upstart's algorithm denies your loan, it may cite a factor like "insufficient credit experience" — a description that tells you nothing about which of its 1,600+ variables actually drove the decision.

In 2023, the CFPB issued guidance stating that "I relied on a complex algorithm" is not an acceptable reason to deny a loan applicant specific, actionable explanations. The guidance put lenders on notice that opacity is not a legal defense. But enforcement remains uneven, and the underlying models are still proprietary.

The European Union's AI Act, which began phasing in during 2024, classifies credit scoring as "high-risk AI" and requires human oversight, transparency, and the right to explanation. The United States has no equivalent federal law as of this writing.

Proxy discriminationWhen an algorithm does not directly use a protected characteristic (race, gender) but uses correlated variables that produce the same discriminatory outcome.
Adverse action noticeA legally required explanation given to consumers when credit is denied or terms are less favorable than requested.
Alternative dataNon-traditional information — rental payments, utility bills, phone usage, social media — used by newer credit models beyond FICO inputs.
Key Takeaway

Credit algorithms do not eliminate bias — they automate it, scale it, and make it harder to challenge. When the training data reflects historical discrimination, the model learns to discriminate. The mathematical form of the output doesn't make it neutral.

Module 5 · Lesson 1

Quiz: Credit Scoring Algorithms

3 questions · select the best answer for each
1. The 2019 Apple Card controversy revealed that Goldman Sachs's algorithm produced different credit limits for men and women even when women had better credit scores. What was the most significant finding from the regulatory investigation?
Correct. The investigation found proxy discrimination — variables that didn't explicitly encode gender but correlated with it. This is the core challenge of algorithmic fairness: a model can be "gender-blind" in its inputs and still produce gender-biased outputs.
Not quite. The investigation found no evidence of intentional programming. The problem was subtler: proxy variables that correlated with gender produced disparate outcomes without any explicit intent to discriminate.
2. A 2022 UC Berkeley study found that fintech mortgage algorithms charged Black and Latino borrowers approximately how much more than white borrowers with identical risk profiles?
Correct. The study found an 11 basis point premium charged to Black and Latino borrowers even after controlling for risk factors. On a 30-year mortgage, that difference compounds into thousands of dollars.
The study found an 11 basis point difference — not zero, and not as extreme as 100 points. The finding showed that algorithmic lending reduced but did not eliminate racial disparities compared to traditional lending.
3. The CFPB's 2023 guidance on algorithmic credit models stated that lenders:
Correct. The CFPB made clear that opacity is not a legal defense. Lenders must provide specific reasons, not hide behind the complexity of their models. Full source code disclosure is not required, and no algorithmic ban was proposed.
The CFPB guidance went the other direction — it explicitly stated that algorithmic complexity does not excuse lenders from providing specific, actionable explanations for credit denials.
Module 5 · Lab 1

Credit Algorithm Interrogation

Explore how lending algorithms encode bias — at least 3 exchanges to complete

What you're doing

You're going to interrogate how credit scoring algorithms work, where their biases come from, and what "proxy discrimination" looks like in practice. Ask the AI about the Apple Card case, FICO's history, alternative data risks, or anything else from Lesson 1.

Start with: "Explain how a credit algorithm can discriminate without ever looking at race or gender directly."
Credit Algorithm Lab
AI TUTOR
Ready to dig into credit algorithms. What would you like to explore — proxy discrimination, the history of FICO, alternative data, or something else from Lesson 1?
Module 5 · Lesson 2

Sentenced by Software

COMPAS, predictive policing, and the use of AI in criminal justice decisions
When an algorithm predicts you will commit a future crime, what evidence does it use — and can you challenge it?

In 2013, Eric Loomis was arrested in La Crosse, Wisconsin, after being found driving a car used in a shooting. He pleaded guilty to charges and was sentenced to six years in prison. At sentencing, the judge cited his COMPAS risk score — a proprietary algorithm that had classified Loomis as "high risk." Loomis argued he had a constitutional right to know how the algorithm worked so he could challenge it. In 2016, the Wisconsin Supreme Court ruled against him, holding that COMPAS could be used at sentencing even though its formula was a trade secret and he could not inspect it.

What COMPAS Actually Measures

COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) was developed by Equivant (formerly Northpointe). It generates risk scores based on a 137-question survey covering criminal history, age, education, employment, and — critically — neighborhood characteristics and family criminal history. The algorithm does not ask about race directly, but a 2016 investigation by ProPublica found that COMPAS was twice as likely to falsely flag Black defendants as high-risk compared to white defendants, and twice as likely to falsely label white defendants as low-risk.

ProPublica analyzed 7,000 criminal defendants in Broward County, Florida. Of defendants who were labeled high-risk but did not reoffend, 44.9% were Black and only 23.5% were white. The company disputed the methodology, and academics debated the statistical approach — but the underlying disparate outcomes were not refuted.

COMPAS is used in at least 17 U.S. states to inform decisions about bail, sentencing, and parole. In most jurisdictions, defendants are not told their score and have no formal mechanism to challenge the algorithm's output.

Predictive Policing

PredPol (now Geolitica) sold "predictive policing" software to over 150 U.S. police departments, claiming its algorithm could predict where crimes would occur. A 2021 investigation by The Markup found that the algorithm disproportionately directed police to low-income Black and Latino neighborhoods — not because those areas had more crime, but because they had historically received more police surveillance, creating a feedback loop. More policing produced more arrests, which produced more data pointing to more policing. Santa Cruz, California banned predictive policing in 2020; Los Angeles stopped using PredPol in 2020 after community protests.

The Feedback Loop Problem

Criminal justice algorithms share a structural flaw with predictive policing: they are trained on arrest and conviction records, not on actual crime rates. Because policing has historically been concentrated in certain communities, those communities have higher arrest records — not necessarily higher crime rates. An algorithm trained on arrests will predict more arrests in those communities, directing more police there, creating more arrests, and reinforcing its own predictions in an endless loop.

This was documented in a 2019 paper by researchers at Rutgers University, who found that predictive policing models amplify historical racial biases by approximately 2× to 3× in communities that were historically over-policed. The models are not predicting where crime will occur — they are predicting where police will find it if they look.

The Human Rights Watch, in a 2021 report, documented that COMPAS and similar tools are used in juvenile justice in states including Pennsylvania and Ohio, meaning teenagers as young as 12 may have their futures shaped by algorithmic risk assessments they cannot see or challenge.

RecidivismThe tendency of a convicted criminal to reoffend. COMPAS and similar tools attempt to predict recidivism risk — but accuracy, fairness, and what "risk" means are deeply contested.
Feedback loopWhen an algorithm's outputs become inputs for future decisions, amplifying initial biases over time. In policing, more surveillance → more arrests → more training data → more surveillance.
Trade secret defenseThe legal argument that proprietary algorithm details need not be disclosed, even when used in high-stakes government decisions affecting liberty. Courts have largely upheld this argument.
Key Takeaway

Criminal justice algorithms make decisions about freedom and incarceration based on data that encodes historical policing patterns. They are shielded from challenge by trade secret protections. The people most affected — often the poorest and least powerful — have the fewest tools to contest them.

Module 5 · Lesson 2

Quiz: Criminal Justice Algorithms

3 questions · select the best answer for each
1. ProPublica's 2016 analysis of COMPAS scores in Broward County, Florida found that the algorithm was twice as likely to do what to Black defendants compared to white defendants?
Correct. The key finding was about false positives: Black defendants who did not reoffend were labeled high-risk at nearly twice the rate of white defendants who also did not reoffend. This is a serious fairness problem — people losing liberty based on incorrect algorithmic predictions.
The ProPublica finding was specifically about false positives for Black defendants — being incorrectly labeled high-risk when they did not subsequently reoffend. This happened at roughly twice the rate compared to white defendants in the same situation.
2. In State v. Loomis (Wisconsin Supreme Court, 2016), the court ruled that COMPAS could be used at sentencing even though:
Correct. This is the core constitutional problem: a proprietary algorithm influenced a sentencing decision, but the defendant had no right to examine how it worked. The Wisconsin Supreme Court held this did not violate due process, a ruling that remains controversial among legal scholars.
The Loomis case centered on the trade secret problem: COMPAS was proprietary, Loomis couldn't inspect its formula, yet it was used to help sentence him to six years. The Wisconsin Supreme Court upheld this practice.
3. What is the core "feedback loop" problem identified in research on predictive policing algorithms like PredPol?
Correct. This self-reinforcing cycle is the key structural problem. The algorithm isn't predicting where crime occurs — it's predicting where police will find crime if they look. Because it sends police to historically over-policed areas, arrests there rise, "confirming" the predictions and perpetuating the cycle.
The feedback loop is the opposite of beneficial self-improvement. It's a self-reinforcing bias: police go where the algorithm says, make arrests there, which the algorithm counts as evidence it was right, sending police back again. Crime in un-policed areas goes undetected.
Module 5 · Lab 2

Criminal Justice AI Debate

Interrogate the ethics of algorithmic sentencing — at least 3 exchanges to complete

What you're doing

You're going to work through the ethical and legal problems with using algorithms in criminal justice decisions. Consider the COMPAS case, predictive policing feedback loops, or the trade secret problem. Push the AI on the hardest questions.

Start with: "Should a defendant have the right to examine the algorithm that influenced their sentence? What would that even require in practice?"
Criminal Justice AI Lab
AI TUTOR
Let's dig into algorithmic criminal justice. Ask me about COMPAS, predictive policing, due process, or the feedback loop problem — what's your first question?
Module 5 · Lesson 3

The Resume You Never Know Was Rejected

How AI screening, automated interviews, and algorithmic management decide your career
When an algorithm decides your resume isn't worth a human's time, what does it actually know about you?

Between 2014 and 2017, Amazon built and then quietly abandoned an AI recruiting tool it had hoped would automate resume screening. The system was trained on 10 years of Amazon's hiring decisions. Researchers discovered it had learned to penalize resumes that contained the word "women's" — as in "women's chess club" or "women's college" — and downgraded graduates of all-women's colleges. Because Amazon's historical hiring had been male-dominated, the algorithm concluded that male was the correct profile for a successful candidate. Amazon disbanded the project in 2017 without ever deploying it in final hiring decisions, but the episode became one of the most cited examples of algorithmic bias in hiring.

ATS: The Invisible Gatekeeper

Applicant Tracking Systems (ATS) are used by an estimated 98% of Fortune 500 companies and up to 75% of all employers. These systems parse and rank resumes before any human sees them. Studies by Harvard Business School and Accenture (2021) found that ATS systems were filtering out millions of qualified candidates — people with non-linear career paths, those who had taken time off, or those with skills described in non-standard terminology. The report called this the "hidden workers" problem: an estimated 27 million Americans were systematically screened out by algorithms before any human could evaluate them.

A 2019 study from the National Bureau of Economic Research found that when companies adopted ATS software, the probability that a job posting received an application from an older worker dropped significantly. The study found that ATS keyword optimization created what amounted to indirect age discrimination: job descriptions written for ATS parsing used tech-industry vocabulary that was systematically less familiar to workers over 50.

HireVue and Automated Video Interviews

HireVue serves over 700 companies including Unilever, Delta, and Goldman Sachs. Its platform records one-way video interviews and, until 2021, used AI to analyze facial expressions, word choice, and tone of voice to generate a hirability score. In 2019, the Electronic Privacy Information Center (EPIC) filed a complaint with the FTC arguing the practice was unfair and deceptive. HireVue announced in 2021 that it would stop using facial analysis — but continued analyzing voice patterns and linguistic content. Critics note that "voice analysis" for personality inference has no validated scientific basis.

Algorithmic Management: After You're Hired

Amazon's warehouse workers are managed largely by algorithm. The system tracks "time off task" — any period when a worker is not scanning items — and can automatically generate termination notices without human review. A 2019 investigation by The Verge obtained internal Amazon documents showing that hundreds of workers had been fired by algorithm for productivity failures, often without any human ever reviewing their cases. Workers described being fired for taking bathroom breaks or for slowing down when they were injured.

Gig economy platforms operate similarly. Uber drivers' accounts can be deactivated — effectively terminating their income — by algorithm, based on low passenger ratings. A 2021 UK Supreme Court ruling found that Uber drivers were legally "workers" and entitled to labor protections. The ruling noted that the algorithm effectively controlled the terms of employment, substituting for the management decisions a traditional employer would make.

The Equal Employment Opportunity Commission (EEOC) issued guidance in 2023 warning employers that algorithmic employment tools can violate the Americans with Disabilities Act if they screen out qualified candidates with disabilities or penalize reasonable accommodations like non-standard work patterns.

Applicant Tracking System (ATS)Software that parses and ranks job applications automatically, often filtering candidates before any human review. Most large-company hiring flows through ATS.
Algorithmic managementUsing software to monitor, direct, and discipline workers in real time — replacing human supervisors with automated systems that track productivity metrics.
Hidden workersTerm from the 2021 Harvard/Accenture report for qualified candidates systematically excluded by ATS algorithms due to non-standard career paths, employment gaps, or unconventional skill descriptions.
Key Takeaway

Algorithms now control both entry to employment and conditions within it. They filter who gets interviews, score how you perform in them, monitor you on the job, and can terminate you — all with limited human oversight and limited recourse for affected workers.

Module 5 · Lesson 3

Quiz: Hiring & Employment Algorithms

3 questions · select the best answer for each
1. Amazon's AI recruiting tool (2014–2017) learned to penalize resumes mentioning "women's" because:
Correct. This is a textbook case of training data encoding historical discrimination. The algorithm was doing exactly what it was designed to do — learn patterns from past successful hires — but the past was biased, so the model encoded that bias.
The bias was unintentional — it emerged from training on historically biased data. The algorithm learned that successful Amazon employees in the past had been predominantly male, so it penalized signals of femaleness in applications. No deliberate engineering of bias was involved.
2. The 2021 Harvard Business School/Accenture report on Applicant Tracking Systems identified approximately how many Americans as "hidden workers" — qualified candidates systematically screened out by ATS algorithms?
Correct — 27 million. The scale is striking. This includes people with gaps in employment, non-linear careers, those re-entering the workforce, and those with skills described in non-standard terms. Algorithmic screening is creating a massive mismatch between labor supply and demand.
The report found approximately 27 million Americans — a staggering number. These are people qualified for jobs who never get seen by a human recruiter because ATS algorithms filter them out for reasons unrelated to actual job performance.
3. What did HireVue announce in 2021 in response to criticism of its AI interview platform?
Correct. HireVue dropped facial analysis — the most obviously pseudoscientific component — but continued analyzing voice and language. Critics point out that voice-based personality inference also lacks validated scientific backing, meaning the reform was partial at best.
HireVue made a partial concession: dropping facial expression analysis while retaining voice and linguistic analysis. It did not shut down, open its algorithms, or mandate human review. Many critics considered the change insufficient.
Module 5 · Lab 3

Hiring Algorithm Audit

Explore bias in employment AI — at least 3 exchanges to complete

What you're doing

You're going to investigate the specific ways employment algorithms create bias and limit opportunity. Ask about Amazon's recruiting AI, ATS hidden worker problems, HireVue's video analysis, or algorithmic management of warehouse and gig workers.

Start with: "If I have an employment gap on my resume, how would a typical ATS algorithm handle it, and is there evidence this creates disparate impacts on specific groups?"
Hiring Algorithm Lab
AI TUTOR
Let's examine hiring algorithms. I can help you explore ATS screening, the Amazon recruiting AI case, HireVue's methods, or how algorithmic management works at Amazon and Uber. What do you want to dig into?
Module 5 · Lesson 4

When the Algorithm Decides Your Care

Healthcare rationing AI, benefits algorithms, and the new gatekeepers of basic needs
When an algorithm determines how much medical care you receive, or whether your benefits continue, who is accountable for what it gets wrong?

In November 2023, a class-action lawsuit was filed against UnitedHealth Group alleging that its subsidiary NaviMedix used an AI algorithm called nH Predict to systematically deny post-acute care claims for elderly patients. The lawsuit alleged the algorithm had a documented 90% error rate — meaning 9 out of 10 patients who appealed their AI-generated denials won on appeal. The plaintiffs alleged that UnitedHealth knew of the error rate but continued using the algorithm because most denied patients did not appeal, making algorithmic denial financially profitable. At least two named plaintiffs died shortly after their care was cut off.

The Systematic Pattern of Insurance AI Denial

The UnitedHealth case was not isolated. A 2023 investigation by STAT News and ProPublica found that multiple major insurers — including Cigna, Humana, and Aetna — were using AI systems to automate denial of prior authorization requests. In one documented case, Cigna's PXDX system allowed physicians to deny claims at a rate of 50 claims per day per doctor, spending an average of 1.2 seconds per claim, without opening any patient file. The process was described as "assembly-line medical care" driven by algorithmic recommendations that physicians rubber-stamped.

A U.S. Senate investigation published in October 2023 found that three of the largest Medicare Advantage insurers — UnitedHealthcare, Humana, and CVS/Aetna — had denial rates as high as 18%, with AI playing a significant role. The Centers for Medicare & Medicaid Services (CMS) issued a memo warning insurers that coverage decisions must be made on an individual basis and that algorithmic denial of entire categories of care was potentially illegal under Medicare Advantage rules.

Benefits Algorithms: Arkansas and Idaho

In 2016, Arkansas replaced its manual Medicaid home care assessments with an algorithm. Thousands of patients immediately had their hours of personal care reduced — often by 30–50% — without explanation. When patients and advocates investigated, they found the algorithm used an illogical formula: patients with diabetes or ulcers actually received fewer care hours than before because the formula's inputs were misconfigured. In Ledgerwood v. Jeger (8th Circuit, 2019), courts found patients' due process rights had been violated because they were never told how the algorithm worked or how to challenge it. Idaho faced a nearly identical problem with its Medicaid algorithm and similar legal consequences.

Child Welfare and Algorithmic Interventions

At least 30 U.S. states use algorithmic risk assessment tools in child welfare decisions — to predict which families are at risk of child abuse or neglect, and to guide decisions about intervention, removal of children, and family services allocation. The most studied is the Allegheny Family Screening Tool (AFST) used in Allegheny County, Pennsylvania.

A 2019 study by Virginia Eubanks, published in her book Automating Inequality, found that AFST and similar tools systematically score poor families and families receiving public assistance as higher-risk — not because poverty causes abuse, but because poor families are more visible to government systems. Families receiving food stamps, Medicaid, or housing assistance generate more data points in government databases, making them algorithmically legible in ways that wealthier families who access private services are not. The algorithm scores what it can measure, not what actually predicts harm.

An independent audit of the AFST commissioned by Allegheny County in 2022 found that Black children were referred to the tool at 2.3× the rate of white children, and that the tool had limited ability to predict actual harm — performing only marginally better than a random baseline in some outcome categories.

Prior authorizationThe process by which insurers must approve certain treatments before they are provided. AI is increasingly used to automate denial of prior authorization requests at scale.
Algorithmic legibilityThe degree to which a person or family is visible to algorithmic systems — typically correlated with poverty, because poor people interact more with government data-collecting systems.
Due processThe constitutional right to notice and an opportunity to be heard before the government deprives you of life, liberty, or property. Courts have found algorithmic benefit cuts can violate this right when no meaningful explanation or appeal is provided.
2016
Arkansas Medicaid algorithm cuts home care for thousands of disabled residents. No explanation provided. Courts later find due process violations.
2019
Ledgerwood v. Jeger (8th Circuit): algorithm-based benefit reductions found to violate due process rights when recipients cannot understand or challenge the formula.
2023
STAT News / ProPublica investigation: Cigna's PXDX system documented spending 1.2 seconds per claim, denying 50 claims per physician per day via AI.
2023
Class action against UnitedHealth: nH Predict algorithm alleged to have 90% error rate on post-acute care denials for elderly Medicare Advantage patients.
2024
EU AI Act phased implementation begins: healthcare AI and social benefit systems classified as high-risk, requiring transparency, human oversight, and right to explanation.
Key Takeaway

Algorithms are now rationing healthcare, cutting benefits, and making child welfare interventions — decisions where errors cost lives. The financial incentive structure rewards algorithmic denial: most people don't appeal, making mistakes profitable. Without accountability, transparency, and meaningful appeal rights, algorithmic decision-making in high-stakes human services is a mechanism for harm at scale.

Module 5 · Lesson 4

Quiz: Healthcare & Benefits Algorithms

3 questions · select the best answer for each
1. The 2023 class-action lawsuit against UnitedHealth Group alleged that its nH Predict AI algorithm had what documented error rate on post-acute care claim denials?
Correct. The lawsuit alleged a 90% error rate on claims that were actually appealed — a staggering figure. The lawsuit also alleged UnitedHealth knew about this rate but continued using the algorithm because most patients didn't appeal, making even wrong denials financially beneficial.
The lawsuit alleged a 90% error rate on appealed claims. This is the critical number: of patients who actually went through the appeal process, 9 out of 10 had their denials reversed — suggesting the algorithm was incorrect the vast majority of the time for these patients.
2. Virginia Eubanks's research on child welfare algorithms like the Allegheny Family Screening Tool found they disproportionately score poor families as high-risk primarily because:
Correct. This is the "algorithmic legibility" problem: the algorithm scores what it can measure. Poor families who use public services generate data. Wealthy families who use private services are invisible to government databases. The algorithm sees more data about poor families, generating higher "risk" scores — not because they're actually higher risk, but because they're more legible.
The key insight from Eubanks's research is about measurement bias, not actual causation. Poor families are more visible to government data systems, so the algorithm scores them higher — not because poverty predicts harm, but because poverty predicts data visibility. Wealthy families using private services are invisible to the algorithm.
3. In Ledgerwood v. Jeger (8th Circuit, 2019), courts found that Arkansas's Medicaid algorithm violated what constitutional principle?
Correct. Due process requires that before the government takes away something you have a legal interest in — like Medicaid home care hours — it must give you meaningful notice and a real opportunity to challenge the decision. When the "reason" is an unexplained algorithm, that notice is meaningless.
The constitutional issue was due process. Before a government benefit can be reduced, recipients are entitled to notice of the reason and a meaningful opportunity to challenge it. An algorithm that neither explains its logic nor can be questioned provides neither — which is why the court found a due process violation.
Module 5 · Lab 4

Healthcare AI Decision Audit

Interrogate algorithmic gatekeeping in healthcare and benefits — at least 3 exchanges to complete

What you're doing

You'll examine the real-world consequences of algorithms making healthcare and benefits decisions. Ask about the UnitedHealth case, insurance prior authorization AI, Arkansas's Medicaid algorithm, or child welfare tools. Consider accountability, incentives, and what reform would actually look like.

Start with: "If an insurance algorithm has a 90% error rate but most patients don't appeal, is the company acting rationally or unethically — or both? What would fix this?"
Healthcare AI Lab
AI TUTOR
Ready to dig into healthcare and benefits algorithms. Ask me about insurance denial AI, the UnitedHealth lawsuit, Arkansas's Medicaid case, child welfare algorithms, or what accountability would look like. What's on your mind?
Module 5

Module Test: When Algorithms Make Decisions About You

15 questions · 80% required to pass · covers all four lessons
1. In the 2019 Apple Card controversy, the New York Department of Financial Services investigated Goldman Sachs. What was the central legal concern?
Correct. Proxy discrimination — the use of correlated variables that produce discriminatory outcomes — was the core issue. The algorithm didn't need to "see" gender directly to discriminate by gender.
The concern was proxy discrimination: the algorithm used variables correlated with gender that produced different outcomes for men and women, without explicitly incorporating gender.
2. FICO scores weight "payment history" as the largest single factor. What percentage of a FICO score does payment history represent?
Correct. Payment history at 35% is the single largest FICO factor, followed by amounts owed (30%), length of history (15%), new credit (10%), and credit mix (10%).
Payment history accounts for 35% of a FICO score — the largest single component.
3. The 2023 CFPB guidance on algorithmic lending models stated that lenders must:
Correct. The CFPB guidance put lenders on notice that "our algorithm is complex" is not a valid substitute for specific, actionable adverse action notices as required by the Equal Credit Opportunity Act.
The CFPB required specific, actionable adverse action notices — algorithmic complexity is not an acceptable reason to withhold meaningful explanation from denied applicants.
4. COMPAS risk scores are used in U.S. criminal justice to inform decisions about all of the following EXCEPT:
Correct. COMPAS is used post-arrest and post-conviction to inform bail, sentencing, and parole — not to determine guilt. Guilt is determined by evidence at trial, not risk scores. But the algorithm influences how long someone stays incarcerated before and after trial.
COMPAS is not used to determine guilt — that remains a matter for trial. But it influences bail (pre-trial detention), sentencing length, and parole release, meaning it affects liberty at multiple stages of the criminal justice process.
5. ProPublica's 2016 analysis found that COMPAS was approximately twice as likely to falsely label Black defendants as high-risk. Of defendants labeled high-risk who did NOT reoffend, what percentage were Black?
Correct — approximately 44.9% of incorrectly labeled high-risk defendants who did not reoffend were Black, compared to 23.5% who were white. This disparity in false positives is the core of ProPublica's finding.
ProPublica found approximately 44.9% of the false-positive high-risk group (those labeled dangerous who did not reoffend) were Black, compared to 23.5% who were white — nearly a 2:1 disparity.
6. The 2019 Rutgers University research on predictive policing found that algorithms like PredPol amplified historical racial bias by approximately what factor in historically over-policed communities?
Correct. The 2–3× amplification factor reflects the feedback loop: historical over-policing generates more crime data in those areas, which algorithms interpret as higher crime risk, directing more policing there, generating even more data — compounding bias over time.
The Rutgers research found approximately 2× to 3× amplification of historical bias in over-policed communities — a substantial effect driven by the self-reinforcing feedback loop between policing data and algorithmic predictions.
7. Amazon's warehouse management algorithm can generate termination notices without human review based on which metric?
Correct. "Time off task" — the system's measure of any period without active scanning — is the documented trigger for automated termination notices. Workers have been fired for bathroom breaks and injury-related slowdowns that fell outside algorithm-defined acceptable pauses.
The documented metric is "time off task" — any period without active scanning activity. The Verge investigation found workers fired for this metric including during bathroom breaks and when injured, with no human ever reviewing their cases.
8. The 2021 Harvard Business School/Accenture report on "hidden workers" found that ATS systems were filtering out which categories of qualified candidates? (Select the most comprehensive answer.)
Correct. The "hidden workers" problem encompasses a broad population: caregivers who left to raise children, workers who changed industries, people who recovered from illness, and anyone whose skills don't match the exact keyword vocabulary of an ATS-optimized job description.
The report identified a broad set of filtered-out candidates: those with employment gaps (caregivers, illness), non-linear career paths, and skills described in different terminology than the ATS was programmed to recognize — totaling approximately 27 million Americans.
9. HireVue's response to criticism in 2021 — stopping facial expression analysis while continuing voice and linguistic analysis — was criticized by experts for what reason?
Correct. Critics pointed out that inferring personality, hirability, or risk from voice patterns has no more scientific validation than facial expression analysis. Removing one pseudoscientific component while retaining another doesn't constitute meaningful reform.
The criticism was that voice-based personality inference is also scientifically unvalidated. Dropping facial analysis while keeping voice analysis was seen as removing the most visibly controversial element without addressing the underlying problem of using unvalidated pseudoscience in hiring.
10. The UK Supreme Court ruling on Uber in 2021 found that the algorithmic platform effectively:
Correct. The ruling established that algorithmic control over working conditions constitutes employment, regardless of how the contractual relationship is labeled. The algorithm's direction of drivers — setting routes, managing ratings, controlling deactivation — was effectively management.
The UK Supreme Court found that Uber's algorithm controlled drivers' working conditions in ways that made them employees ("workers" in UK law), not independent contractors. This was significant because it meant labor protections — minimum wage, holiday pay — applied.
11. A 2023 U.S. Senate investigation found that major Medicare Advantage insurers had AI-influenced denial rates as high as:
Correct. Denial rates as high as 18% were found at major Medicare Advantage insurers, with AI playing a significant role. The Centers for Medicare & Medicaid Services issued warnings that algorithmic denial of entire categories of care may violate Medicare Advantage rules.
The Senate investigation found denial rates as high as 18% at some major Medicare Advantage insurers — rates that prompted CMS to warn that algorithmic denial of entire care categories may violate program rules requiring individual coverage determinations.
12. Cigna's PXDX system documented in the 2023 STAT News/ProPublica investigation allowed physicians to deny how many claims per day, spending how much time per claim?
Correct. 50 claims per day at 1.2 seconds each — no physician can meaningfully review a patient file in 1.2 seconds. This is algorithmic rubber-stamping: a doctor's name on a denial generated entirely by AI, providing legal cover for automated mass denial.
The documented figures were 50 claims per day at an average of 1.2 seconds per claim. At that pace, no actual review of any individual patient file is possible — the physician's involvement is purely nominal, providing legal cover for algorithmic mass denial.
13. The Allegheny Family Screening Tool (AFST) 2022 independent audit found that Black children were referred to the tool at what rate compared to white children?
Correct. Black children were referred to the AFST at 2.3× the rate of white children, and the audit found the tool had limited predictive accuracy — performing only marginally better than a random baseline in some outcome categories. High disparity, low accuracy.
The audit found a 2.3× referral rate for Black children compared to white children — and found that the tool's predictive accuracy was limited, raising serious questions about whether the algorithmic disparity was producing any safety benefit.
14. Virginia Eubanks coined the term "algorithmic legibility" to describe a problem where child welfare algorithms disproportionately score poor families as high-risk. The core mechanism is:
Correct. Algorithmic legibility is about data visibility, not actual risk. Families using public services are in government databases. Families using private services are not. The algorithm scores data presence, not danger — creating systematic bias against the poor.
The issue is "algorithmic legibility" — poor families who use public services (Medicaid, food stamps, public housing) are in government databases. Wealthier families using private services are invisible to those databases. The algorithm scores data visibility, not actual risk.
15. Which jurisdiction's legislation classifies both credit scoring and healthcare AI as "high-risk" and requires transparency, human oversight, and the right to explanation?
Correct. The EU AI Act, phasing in from 2024, is the world's first comprehensive AI regulation classifying high-risk AI applications — including credit scoring, healthcare, and criminal justice — with specific requirements for human oversight, transparency, and the right to explanation. The US has no equivalent federal law.
The EU AI Act (phasing in from 2024) is the relevant legislation. The US Algorithmic Accountability Act has not passed as of this writing. The CCPA covers data privacy but not AI risk classification. The UK has taken a different, more sector-specific approach.