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

The Soap Dispenser That Couldn't See

Why training data is the original sin β€” and how a machine raised on skewed examples becomes permanently tilted.
If an AI only ever learned from one kind of world, what does it do when it meets a different one?

In August 2017, a video went viral without any AI company's permission. A Black man named Chukwuemeka Afọlabi stood at a motion-activated soap dispenser in a hotel bathroom. He waved his hand. Nothing. He waved again. Nothing. Then a white colleague walked up, held their hand in the same spot β€” and the dispenser immediately released soap.

The video spread because the dispenser was not broken. It worked exactly as designed. The problem was what it had been trained to detect: reflected infrared light, calibrated on hands with lighter skin tones. Darker skin absorbs more infrared. The sensor never "learned" that darker hands existed. So to the machine, they were invisible.

The company had not programmed racism into its sensor. But racism β€” or at least, a world that had mostly tested this technology on one group of people β€” had found its way in anyway. Through the data.

What Is Training Data, Really?

Every AI system learns from examples. You show it thousands β€” sometimes billions β€” of pictures, texts, measurements, or decisions, and it finds patterns. The collection of examples you use is called training data. It is the entire world the AI has ever seen.

Here is the problem: the real world is not a neutral, perfectly balanced place. Some groups have been photographed more. Some voices have been recorded more. Some languages appear far more often on the internet. When you scoop up a giant pile of "the world" and hand it to an AI, you're actually handing it a very specific slice of the world β€” one shaped by who had power, who had cameras, who had access to technology first.

The AI doesn't know this. It just sees what it sees. And it assumes that what it sees is everything.

Training dataThe collection of examples an AI system learns from. Whatever is missing, skewed, or wrong in this data becomes missing, skewed, or wrong in the AI's behavior.
Bias (in AI)When an AI system systematically produces worse results for some groups than others, usually because those groups were underrepresented or misrepresented in the training data.
The Face Recognition Problem That Went National

In January 2020, a man named Robert Williams was arrested in Detroit, Michigan. Police officers came to his home, handcuffed him in front of his wife and daughters, and drove him to a detention center. He spent 30 hours in jail before investigators realized their lead had come entirely from a face recognition algorithm β€” and that the algorithm was wrong.

The system had matched Williams's driver's license photo to a blurry surveillance image from a watch store robbery. Williams had never been to the store. The AI had made a false identification. But investigators had trusted it enough to make an arrest.

Later research revealed why this kind of error was predictable. A 2018 study by researcher Joy Buolamwini at MIT β€” published under the name Gender Shades β€” tested face recognition systems from Microsoft, IBM, and Face++ on thousands of faces. The results were stark: the systems were up to 34% less accurate on dark-skinned women compared to lighter-skinned men. The training datasets had simply contained far more lighter-skinned faces. The machines had learned a world that didn't reflect everyone equally, and they performed exactly as well as that skewed world had equipped them to.

Robert Williams was eventually cleared, but the experience β€” the arrest, the night in jail, the terrifying scene in front of his daughters β€” was not something he got back.

Why This Isn't Just A Tech Problem

Police departments in at least 24 U.S. cities were using face recognition AI as of 2020. Most of those departments had no formal policy about how to verify an AI match before making an arrest. The technology moved faster than the rules designed to govern it.

Garbage In, Garbage Out β€” At Scale

Programmers have a saying that predates AI by decades: garbage in, garbage out. It means a system can only be as good as the information you feed it. But AI makes this problem worse in a specific way: AI systems take small patterns in data and amplify them into confident predictions applied to millions of people.

Imagine a hiring AI trained on 10 years of rΓ©sumΓ©s from a tech company where 85% of people hired were men. The AI doesn't know that this happened because of historical bias in who was recruited or who felt welcome applying. It just sees the pattern: men got hired. So it learns to rank male applicants higher. Amazon actually built and then scrapped exactly this kind of system in 2018 after discovering it was systematically downgrading rΓ©sumΓ©s that included the word "women's" β€” as in "women's chess club" β€” because the training data had taught it that women rarely got hired.

The bias wasn't in the algorithm itself. It was in the reflection of a biased world that the data captured. But the algorithm took that reflection and turned it into policy, at scale, applied to thousands of applicants.

You Can Now See What Most People Miss

When you read a headline like "AI hiring tool gives unfair results," most readers assume someone programmed discrimination in on purpose. You now know the more common truth: no one had to. The data carried the discrimination in silently, and the AI learned it without anyone noticing β€” until the harm was already done.

The Question No One Has Answered

Here is the ethical question sitting at the center of all of this, and it doesn't have a clean answer:

If an AI reflects the world accurately β€” including the world's real inequalities β€” is that fair or unfair?

Think about it. A crime prediction AI might correctly learn that certain zip codes have historically had more arrests. But those arrests were partly the result of police being deployed to those neighborhoods more. So the AI is learning "where police found crime" not "where crime actually happens." The data is accurate. But accurate data from a biased system produces biased conclusions.

Some researchers argue you should "debias" training data β€” deliberately add more examples of underrepresented groups to balance things out. Others argue that artificially adjusting data introduces a different kind of distortion. Still others say the real problem isn't the data at all β€” it's using AI in high-stakes decisions like hiring and policing when we know these errors occur.

Who decides? How do you weigh accuracy against fairness? These are questions that companies, governments, courts, and researchers are actively arguing about right now β€” without agreement.

Module 5 Β· Lesson 1

Quiz: The Soap Dispenser That Couldn't See

5 questions β€” test your reasoning, not your memory.
1. The hotel soap dispenser failed to detect darker skin tones primarily because:
Correct. The sensor worked as trained β€” the problem was what it had been trained on. It learned from a narrow slice of the world and generalized that slice to everyone.
Not quite. The sensor wasn't broken and no one intentionally excluded anyone. The failure came from skewed training data β€” the sensor simply hadn't learned to recognize what it hadn't seen enough of.
2. Joy Buolamwini's 2018 Gender Shades study found that commercial face recognition systems were least accurate on which group?
Correct. The systems performed up to 34% worse on darker-skinned women β€” the group least represented in the training datasets.
Actually, the study found accuracy was worst for darker-skinned women. Training datasets had far more lighter-skinned male faces, so those were the faces the systems learned best.
3. A school uses an AI to predict which students are at risk of failing. The AI was trained mostly on data from wealthy school districts. A teacher uses it in a lower-income school and notices it keeps flagging students who are actually doing fine. What is the most likely cause?
Correct. This is a training distribution mismatch β€” the AI learned what "at risk" looks like in one type of school, and those patterns don't map cleanly onto a different context.
This is a training data mismatch problem. The AI learned what "at risk" looks like in wealthy schools β€” maybe different attendance patterns, different grade distributions β€” and those patterns don't transfer directly to a different environment.
4. Amazon scrapped its AI hiring tool in 2018 because it penalized rΓ©sumΓ©s that included the word "women's." What does this reveal about how bias enters AI systems?
Correct. No one wrote a rule saying "penalize women." The AI inferred that rule from the data β€” because the data reflected a world where women had historically been hired less. The bias was implicit in the training signal.
The bias wasn't programmed β€” it was learned. The AI saw 10 years of hiring data where men were selected more often and learned to replicate that pattern. It turned historical inequality into a prediction rule.
5. Which of the following best describes why "garbage in, garbage out" is a bigger problem for AI than for a simple calculator?
Correct. A calculator applies a fixed rule to your specific input. An AI generalizes from patterns in massive datasets and then applies those generalizations to new situations β€” amplifying any errors or biases in the original data across every decision it makes.
The key difference is scale and generalization. An AI doesn't just process your input β€” it applies patterns it learned from millions of examples to make predictions. Any tilt in those patterns gets multiplied across every future decision.
Module 5 Β· Lab 1

Bias Auditor

You're reviewing an AI system before it gets deployed. Your job is to catch what the engineers missed.

Your Scenario

A city government wants to deploy an AI system to help decide which neighborhoods get priority for road repairs. The system was trained on 20 years of repair request data and resident complaint logs. The city says it's "objective" because it uses data instead of human judgment. You've been asked to audit it before launch.

Your lab partner has reviewed the technical specs. They're pushing back on your concerns. Make your case β€” and defend it against their challenges.

Start by telling your partner: what's the first question you'd ask about the training data before trusting this system?
Lab Partner β€” Bias Audit
AI Peer
I've looked at the technical documentation. The model accuracy is 91% on the test set. The engineers are saying that's good enough to deploy. What's your first concern β€” and why does it matter more than the accuracy number?
Module 5 Β· Lesson 2

When Confident Is Wrong

AI systems don't just make mistakes β€” they make mistakes with certainty. Understanding why confidence and accuracy are different things changes how you read every AI output.
Why does an AI that's wrong 1% of the time still cause serious harm β€” and why does it keep insisting it's right?

On May 14, 2013, a judge in Barron County, Wisconsin sentenced a man named Eric Loomis to six years in prison. The judge cited, in part, a score generated by a computer program called COMPAS β€” Correctional Offender Management Profiling for Alternative Sanctions. COMPAS had rated Loomis as high risk for reoffending.

Loomis appealed. He argued that he had a right to understand and challenge the score. The company that made COMPAS, Northpointe, refused to reveal how the algorithm worked, calling it a proprietary trade secret. The Wisconsin Supreme Court upheld the sentence in 2016, ruling that the score was "one factor among many." His case eventually reached the U.S. Supreme Court, which declined to hear it.

Meanwhile, journalists at ProPublica had been running their own investigation. In May 2016, they published a landmark analysis of COMPAS scores for over 7,000 people in Broward County, Florida. Their finding: Black defendants were nearly twice as likely as white defendants to be falsely flagged as high risk β€” meaning they were rated dangerous and then did not reoffend. White defendants were more likely to be falsely rated low risk β€” and then did go on to reoffend. The algorithm's errors were not random. They were systematically distributed along racial lines.

Why Confidence Is Not Accuracy

COMPAS didn't say "this person might reoffend." It produced a numerical score β€” typically 1 to 10 β€” that felt precise and authoritative. Judges, who are busy humans under enormous caseloads, received a number that looked like certainty.

This is one of the most dangerous properties of modern AI systems: they output confident-sounding results even when their actual accuracy is limited. A large language model β€” the kind that powers chatbots β€” doesn't say "I'm not sure." It generates the next most probable word, with the same smooth fluency regardless of whether the content is accurate or fabricated. A face recognition system doesn't flag its output as "uncertain." It returns a match with a percentage score that looks like proof.

The term for this specific failure mode is overconfidence β€” or in AI research, it's sometimes called miscalibration. A well-calibrated system says "I'm 80% confident" on things it gets right 80% of the time. A miscalibrated system says "I'm 95% confident" on things it only gets right 60% of the time. The system's stated certainty doesn't track with its actual accuracy.

CalibrationHow well an AI's stated confidence matches its actual accuracy. A well-calibrated AI that says "90% confident" should be right about 90% of the time. Most deployed AI systems are miscalibrated β€” often overconfident.
HallucinationWhen an AI generates something that sounds accurate and confident but is completely made up. The term is used especially for text-generating AI that produces false facts, fake citations, or invented events.
The Hallucination Problem Gets Someone Fired

In June 2023, a New York lawyer named Steven Schwartz submitted legal filings to a federal court that cited six specific court cases as precedents β€” Varghese v. China Southern Airlines, Shaboon v. EgyptAir, and four others. The opposing counsel checked the citations. None of them existed. Schwartz had used ChatGPT to help write the brief, and ChatGPT had fabricated six entirely fictional court decisions, complete with convincing case names, dates, and judicial rulings.

When asked about the cases, ChatGPT confirmed they were real. When asked to double-check, it confirmed again. It produced fake screenshots of fake database entries showing the fake cases. The model had no way to know it was wrong. It was generating plausible-sounding legal language based on patterns in its training data, with complete fluency and zero uncertainty. Schwartz was fined and publicly reprimanded. He later said he had not realized the AI could "just make things up."

What makes hallucination especially treacherous is that the false outputs are indistinguishable in style from accurate ones. A made-up court case looks exactly like a real court case citation. A fabricated scientific study looks exactly like a real one. The machine's confidence is constant; only the accuracy varies.

Scale This Problem Up

Schwartz was one lawyer using one AI system once. As of 2024, AI writing tools are used by millions of students, journalists, doctors, and researchers. The same hallucination problem exists in all of them. The question of who is responsible for verifying AI-generated content β€” the user, the company, or no one β€” has not been answered.

The Accountability Gap

Both the COMPAS story and the hallucinating lawyer share a structural problem: no one was clearly accountable when the AI was wrong.

When COMPAS produced a racially skewed score, Northpointe said the algorithm was proprietary. The judge said it was just one factor. The appeals court said the process was legal. Eric Loomis sat in prison. When ChatGPT fabricated court cases, OpenAI said users are responsible for verifying outputs. Steven Schwartz said he didn't know. The judge fined Schwartz, not OpenAI.

This is what researchers call the accountability gap: a space where AI-caused harm happens but where responsibility is endlessly passed from the AI company to the human user to the institution that deployed it, until no one is left holding it. It is not unique to AI β€” corporations have exploited similar gaps for decades. But AI makes the gap wider because the errors can happen at machine speed, across millions of decisions, before anyone notices.

Here is the institutional reality that matters at your age: governments in the European Union began requiring AI systems used in high-stakes decisions β€” hiring, lending, criminal justice, immigration β€” to meet transparency and accuracy standards starting in 2024, under the EU AI Act. The United States, as of that same year, had no equivalent federal law. Different rules apply in different places, right now, to the same AI systems.

The Uncomfortable Question

If an AI makes a decision that harms someone β€” and that decision was made by a system no one fully understands, trained by a company that keeps its methods secret, deployed by an institution that trusted the score β€” who is responsible? The honest answer right now is: we don't have agreement. That is a political and legal question as much as a technical one. And the answer being worked out today will shape how AI is used for decades.

Module 5 Β· Lesson 2

Quiz: When Confident Is Wrong

5 questions β€” apply what you learned to new scenarios.
1. ProPublica's 2016 investigation of COMPAS found that the algorithm's errors were:
Correct. The errors weren't random noise β€” they were patterned along racial lines, with Black defendants more likely to be wrongly rated dangerous, and white defendants more likely to be wrongly rated safe.
The investigation found the opposite of random errors. The mistakes fell disproportionately on Black defendants, who were nearly twice as likely to be falsely flagged as future offenders.
2. What does it mean for an AI to be "miscalibrated"?
Correct. Miscalibration is the mismatch between confidence and accuracy. A well-calibrated system that says "90% confident" should be right 90% of the time. A miscalibrated one might say "95% confident" when it's only right 60% of the time.
Miscalibration specifically refers to the gap between how confident an AI sounds and how often it's actually correct. It's not about always being wrong β€” it's about being wrong more than the confidence level implies.
3. A medical AI diagnoses a rare disease and outputs "98.7% confidence: negative." The patient is actually positive. What does this scenario best illustrate?
Correct. A 98.7% confidence score signals to the doctor "don't look further." That's what makes overconfidence especially harmful in high-stakes domains β€” the certainty discourages human verification at exactly the moment when it's needed most.
The key danger here is the combination of a wrong answer AND high confidence. A doctor who sees "98.7% negative" is less likely to order additional tests. The false certainty actively prevents the correct diagnosis from being found.
4. Why was attorney Steven Schwartz's situation in 2023 particularly significant as an example of AI hallucination?
Correct. The especially alarming detail was that when asked to verify, ChatGPT doubled down and confirmed the fabrications. The model has no mechanism to distinguish its accurate outputs from its invented ones β€” it generates both with equal fluency.
The most important detail was that when Schwartz asked the AI to double-check, it confirmed the fake cases were real. There's no internal warning signal that fires when the model is hallucinating. It produces false and true content with identical confidence.
5. The "accountability gap" in AI systems refers to which of the following?
Correct. It's a structural problem: the AI company points to the user, the deploying institution points to the AI company, the user says they didn't fully understand. The harm is real but the responsibility diffuses until it disappears.
The accountability gap is about responsibility, not accuracy or access. When AI causes harm, there's often a chain of deflection β€” the maker, the deployer, and the user all point at each other β€” and no one ends up accountable for the damage.
Module 5 Β· Lab 2

The Confidence Meter

You're a policy analyst deciding whether to approve an AI system for hospital use. Your partner thinks the accuracy numbers are good enough.

Your Scenario

A hospital wants to deploy an AI system that predicts which patients admitted to the emergency room are at high risk of sepsis (a dangerous infection that kills fast if untreated). The AI flags patients for priority treatment. Its overall accuracy is 89%. The company says that's better than the average doctor's unaided judgment.

Your lab partner thinks 89% is impressive and recommends approval. You're not so sure. Take a position and defend it.

What questions would you ask before deciding whether 89% accuracy is acceptable for this specific use case?
Policy Partner β€” Hospital AI Review
AI Peer
Look, 89% is genuinely good for medical AI. The vendor showed us comparison studies. But you seem skeptical β€” what exactly is your concern? Give me something specific, not just "it could be wrong."
Module 5 Β· Lesson 3

The Loop That Tightens

Some AI errors don't just happen once β€” they feed back into the data and make the next generation of errors worse. This is how a small mistake becomes a permanent system.
What happens when the AI's wrong predictions become the training data for the next AI?

In 2011, the Santa Cruz Police Department became one of the first police agencies in the United States to use predictive policing software β€” an AI called PredPol (later renamed Geolitica). The system analyzed historical crime data and produced maps of small geographic zones β€” sometimes just 500 feet by 500 feet β€” where it predicted crime was most likely to occur on a given day. Officers were directed to patrol those zones.

The logic seemed simple: if crime happened somewhere before, it might happen there again. Send officers, catch crime early. But researchers who studied PredPol deployments found something troubling happening over time. The zones flagged as high-risk were concentrated in neighborhoods that were predominantly Black and Latino. Officers sent there made more stops and arrests. Those arrests fed back into the historical crime data. The data then showed even more crime in those zones. The algorithm flagged them more aggressively. More officers were sent. More arrests were made.

The system had created a feedback loop. It wasn't measuring where crime was highest. It was measuring where police had gone β€” and then using that to decide where police should go next. The algorithm was chasing its own tail, and the tail was made of someone else's neighborhood.

In 2020, Santa Cruz became the first city in the United States to ban predictive policing, citing these exact concerns. The same year, Santa Cruz banned facial recognition by police as well.

Feedback Loops: How Small Errors Compound

A feedback loop happens when the output of a system becomes input to the same system. In music, it's that screech when a microphone gets too close to a speaker β€” sound goes out, gets picked up again, amplified again, louder and louder. In AI, feedback loops happen when an AI's predictions change real-world behavior, that behavior generates new data, and that data gets used to train or update the AI.

The predictive policing example is a clean illustration. But feedback loops appear everywhere AI is deployed at scale:

Content recommendation: YouTube's recommendation algorithm suggests videos you might like. The videos you click get recorded. The algorithm learns to suggest more like them. Over time, it pushes toward more extreme content because extreme content generates more clicks and watch time. The algorithm optimizes for engagement, not accuracy or wellbeing. Researchers at the Centre for Countering Digital Hate documented in 2020 that YouTube's algorithm served conspiracy videos to new users within 70 minutes of them creating a fresh account.

Credit scoring: An AI decides a neighborhood is high credit risk. People in that neighborhood get worse loan terms or are denied loans. Without credit access, some fall into debt or default. The AI sees more defaults from that neighborhood. It decides the neighborhood is even higher risk. The loop tightens.

In both cases, the AI is technically accurate within its own frame. It really is finding patterns. The problem is that it is finding patterns partly created by its own past predictions.

Feedback loopWhen an AI system's outputs influence the data that comes back into the system, causing small errors or biases to compound over time rather than self-correct.
Model Collapse: When AI Learns From Itself

There is a newer version of this problem that researchers began documenting in 2023. As AI-generated text, images, and code flood the internet, future AI systems will inevitably be trained on data that includes large amounts of AI-generated content. Researchers at the University of Oxford and other institutions published studies showing that when AI systems are trained on their own outputs β€” or on data heavily contaminated with AI outputs β€” their capabilities degrade systematically. The technical term is model collapse.

Think of it this way: every time you photocopy a photocopy, you lose a little clarity. Do it enough times and the image becomes unrecognizable. When an AI is trained on text written by an AI that was trained on text written by an AI, each generation amplifies errors and narrows variety. The model begins to lose the richness of original human expression and instead converges on a smaller and smaller set of patterns β€” the most average, most common, most synthetic version of language.

This is not a hypothetical future risk. By 2024, researchers estimated that between 15% and 60% of content on major platforms was AI-generated or AI-assisted, depending on the type of content. The web that future AI systems will train on looks increasingly like a hall of mirrors.

The Practical Consequence For You

When you use an AI to help write something β€” an essay, a social media post, an email β€” and that text goes online, it becomes potential training data for the next model. You are, in a small way, participating in shaping what AI learns next. This is not a reason not to use these tools. It's a reason to understand that the system is not static. It evolves based on what people do with it.

Breaking the Loop β€” And Whether Anyone Will

Feedback loops are not impossible to break. The key is independent measurement β€” checking the AI's outputs against sources of truth that are not themselves generated by the AI. For predictive policing, that would mean auditing arrest data against victimization surveys (which ask people directly about crimes, rather than relying on what police recorded). For credit scoring, it would mean regularly testing whether the model's predictions are self-fulfilling by looking at what happens when restrictions are lifted in flagged areas.

The uncomfortable reality is that breaking feedback loops requires time, money, and willingness to find problems. Companies and institutions that deploy AI systems profit from their outputs and have limited financial incentive to conduct audits that might reveal the system is making things worse. Third-party auditing β€” where independent researchers get access to AI systems and data β€” is widely recommended by AI researchers but resisted by most companies on grounds of privacy and intellectual property.

Here is the tension at the institutional level that matters: in 2024, the EU AI Act required companies deploying high-risk AI systems (including those used in policing, employment, and credit) to conduct regular conformity assessments. It is the strongest AI governance law in effect anywhere in the world. Researchers widely consider it insufficient. Companies lobbied hard to weaken it. The United States, China, and most other countries have no equivalent requirement. The feedback loops continue.

You Can See What Most People Miss

When a technology company says its AI is "continuously learning" and "improving over time," that sounds like a feature. You now know it can also be a bug. Continuous learning means continuous feedback. If the system's predictions are shaping the data it's learning from, "improving" might actually mean "becoming more certain of its own errors." The label means nothing without knowing whether the feedback loop is clean.

Module 5 Β· Lesson 3

Quiz: The Loop That Tightens

5 questions β€” reason through feedback and compounding errors.
1. What specific problem did researchers find with predictive policing software like PredPol?
Correct. The algorithm wasn't tracking crime β€” it was tracking where police went, and then directing police to go there again. Each cycle of prediction-patrol-arrest made the next cycle's predictions even more skewed toward those same areas.
The problem wasn't a coding error. The algorithm worked as designed. The flaw was that it created a feedback loop: police presence led to arrests, arrests led to more "crime data" in those areas, more crime data led to more police, and so on.
2. "Model collapse" refers to which phenomenon?
Correct. Like photocopying a photocopy, each generation of AI trained on AI-generated data loses fidelity. The outputs become more narrow, more average, and more prone to compounding earlier errors.
Model collapse is about degradation through self-referential training β€” when AI is trained on AI outputs, it loses the richness of the original human-generated data and begins converging on increasingly narrow, error-prone patterns.
3. A music streaming app uses AI to recommend songs. It notices a new artist getting lots of plays and recommends them more, which causes even more plays, which triggers even more recommendations. Six months later, the artist is inescapable on the platform. What type of AI problem does this best illustrate?
Correct. This is a textbook feedback loop. The AI's output (recommendations) shapes the data (play counts) that feeds back into the AI. The loop can make a modest signal into a dominant one, regardless of whether the underlying quality justifies it.
This is a feedback loop β€” the AI's predictions create behavior that generates data that reinforces those same predictions. It's not hallucination (nothing is fabricated) or miscalibration (the confidence level isn't the issue). The AI is responding to real plays, but it helped create those plays.
4. What is "independent measurement" and why is it the key tool for breaking AI feedback loops?
Correct. The only way to know if a feedback loop has distorted an AI's "learning" is to compare its predictions against a ground truth that was collected independently β€” not generated by the AI's own influence on the world.
Using a second AI doesn't help if both AIs share the same feedback-contaminated data. Independent measurement means comparing predictions against ground truth data that the AI had no role in creating β€” like crime victimization surveys that don't depend on police records.
5. A company says their AI is "constantly learning and getting smarter." Based on what you know about feedback loops, what important follow-up question should you ask?
Correct. "Constantly learning" is only good if what it's learning from is clean and independent. If the AI's predictions shape the data it's learning from, "constantly learning" can mean "constantly compounding its own mistakes with growing confidence."
The key question is whether the learning is based on independent data or on data shaped by the AI's own outputs. Continuous learning with a dirty feedback loop can mean the AI is getting more confidently wrong over time, not more accurate.
Module 5 Β· Lab 3

Loop Detector

You're a researcher investigating whether a social media platform's AI has created a harmful feedback loop. Your partner works at the company.

Your Scenario

A social media platform uses AI to rank what content users see. You've been hired as an independent researcher to assess whether the system has created a feedback loop around health misinformation β€” false medical claims that keep getting recommended because they generate high engagement (outrage clicks, shares, comments).

Your contact at the company defends the system. They say the AI is just responding to what users want. You need to make the case that "what users click on" and "what users want" are not the same thing β€” especially when the AI itself is shaping what they click on.

Make your opening argument. Why is "high engagement" an unreliable signal for content quality β€” especially when the AI itself is influencing which content gets seen?
Platform Contact β€” Feedback Loop Investigation
AI Peer
Here's our position: the AI shows people what they engage with. If misinformation gets high engagement, users are choosing to engage with it. That's a user behavior problem, not an AI problem. Why should we override what users are demonstrably interested in?
Module 5 Β· Lesson 4

The Black Box Problem

The most powerful AI systems in the world cannot fully explain their own decisions. This lesson is about why that matters β€” and what, if anything, we can do about it.
If the machine can't explain itself, can we trust it? Should we?

In 2017, the UK government rolled out an automated system to assess visa applications from foreign nationals. The system was developed by Capita under contract with the Home Office. It used an algorithm to flag applications for closer scrutiny β€” or outright refusal β€” based on factors that immigration officers were never fully told.

One of those factors, journalists later revealed, was nationality. The algorithm sorted applicants by country of origin into three groups: green (lower scrutiny), amber (some scrutiny), and red (high scrutiny). Applicants from countries like Nigeria, Ghana, Pakistan, and Afghanistan were consistently funneled into the red stream. The Home Office called it a risk tool. Critics called it an automated nationality-discrimination machine.

In August 2020, after years of legal pressure from civil society groups including the Joint Council for the Welfare of Immigrants, the Home Office quietly scrapped the algorithm. No full technical audit of its decisions was ever made public. Thousands of people had had their visa applications affected by a system whose inner workings the government itself didn't fully understand β€” and refused to explain.

What Is a Black Box?

A black box is any system where you can observe the inputs and outputs but not the process that connects them. You put something in. Something comes out. What happened in the middle is hidden β€” sometimes because of secrecy, sometimes because the system itself is too complex to interpret.

Early AI systems β€” simple decision trees, rule-based expert systems β€” were what researchers call interpretable. You could open them up and read the rules. If a loan was denied, you could find the specific rule that triggered the denial and challenge it.

Modern deep learning systems β€” the neural networks behind facial recognition, language models, image generators, and most cutting-edge AI β€” are fundamentally different. They consist of millions or billions of numerical weights, adjusted through training across vast datasets. There is no single rule to point at. The decision emerges from the combined behavior of an enormous number of interconnected values that no human designed individually. Not the engineers. Not the researchers. Not the company's CEO. No one can look inside and say "this is why the AI decided this."

Black boxAn AI system whose internal decision process cannot be examined or explained, even by the people who built it. You see inputs and outputs, but not the reasoning between them.
Interpretability / ExplainabilityThe degree to which a human can understand why an AI system made a specific decision. Simple rule-based systems are highly interpretable. Deep learning models generally are not.
The Apple Credit Card Controversy

In November 2019, entrepreneur and programmer David Heinemeier Hansson (creator of the Ruby on Rails programming framework) posted on Twitter that Apple Card's credit limit algorithm had given him 20 times the credit limit it gave his wife β€” despite her having a higher credit score and their finances being completely shared. His co-founder's wife reported the same pattern. The story went viral.

Apple's financial partner, Goldman Sachs, issued a statement saying their algorithm did not use gender as an input. This was probably technically true. But the Apple Card algorithm also used factors like spending history, types of purchases, and other behavioral signals β€” signals that can reflect gender without the word "gender" ever appearing in the code. The AI had potentially learned that female spending patterns were associated with lower credit limits because those limits had historically been lower. Again: no one programmed discrimination. The historical reality of discrimination trained itself into the algorithm.

Here is the part that matters: Goldman Sachs genuinely could not fully explain to regulators exactly which combination of factors had produced any individual credit decision. The New York Department of Financial Services launched an investigation. The investigation could not definitively prove or disprove discrimination because the algorithm itself was opaque. No one could read it clearly enough to know.

Why This Is A Civil Rights Issue

The Equal Credit Opportunity Act has prohibited gender-based credit discrimination in the United States since 1974. But a law written for human decisions faces a hard question when the discriminating entity is an algorithm whose reasoning can't be decoded. If you can't read the algorithm, how do you prove it discriminated? This is one of the central challenges AI poses to existing civil rights law.

Explainability Research β€” and Its Limits

Researchers are working hard on what they call explainable AI (XAI) β€” techniques to produce human-readable justifications for AI decisions even when the underlying model is a black box. One common approach is LIME (Local Interpretable Model-agnostic Explanations), which tests how small changes to the input affect the output and uses that to generate an approximate explanation. Another is SHAP (SHapley Additive exPlanations), which estimates how much each input feature contributed to a given prediction.

These tools are valuable, but they have fundamental limits. They don't tell you the actual reason the model decided something β€” they tell you what factors appear correlated with the decision. That's different. It's like asking someone why they chose a restaurant and getting back "you looked at restaurants with good reviews and outdoor seating" β€” that describes a pattern, not a decision process.

Some AI researchers argue that the push for explainability is, in some cases, being used as a delay tactic β€” companies can claim they're "working on explainability" as a reason not to impose stronger regulation now. Others argue that for some applications β€” say, approving a mortgage or flagging someone for extra airport screening β€” you shouldn't deploy AI at all unless it can fully explain itself. The EU AI Act takes a middle position: it requires explanations for high-stakes automated decisions, but doesn't specify how thorough those explanations need to be.

The deepest tension is this: the most capable AI systems are often the least interpretable. The same properties that make modern neural networks so powerful β€” their ability to find complex patterns in vast data β€” are exactly what makes them impossible to fully explain. Capability and transparency may be in fundamental conflict. If that's true, then using the most powerful AI systems requires accepting that you can never fully understand why they do what they do.

That's not a technical problem waiting for a technical solution. That's a choice society is making right now about how much we're willing to trust systems we cannot fully see.

The Question That Doesn't Have an Answer Yet

Should an AI system that cannot explain its decisions ever be allowed to make decisions that significantly affect someone's life? There are people who argue yes β€” the system can be accurate even if it's not interpretable, and accuracy should be what matters. There are others who argue no β€” the right to understand why a decision was made about you is a fundamental right, regardless of whether the decision turned out to be correct. Neither side has won this argument. It is being fought in courts, parliaments, and research labs right now. Knowing this puts you ahead of most adults who are affected by these systems daily without ever asking the question.

Module 5 Β· Lesson 4

Quiz: The Black Box Problem

5 questions β€” think about transparency, accountability, and what it means to trust a system you can't see inside.
1. What made the UK Home Office visa algorithm a "black box" problem, rather than simply a case of illegal discrimination?
Correct. The opacity is what created the legal and accountability dead end. Even if discrimination was occurring, it couldn't be decisively proven because no one β€” including the government β€” could fully read and audit the system that was making the decisions.
The key was opacity, not legality. The algorithm's internal workings weren't fully accessible, so when critics argued it was discriminating by nationality, there was no way to fully verify or disprove the claim β€” the system couldn't explain itself.
2. Goldman Sachs claimed their Apple Card algorithm did not use gender as an input. Why did this defense fail to satisfy critics?
Correct. This is the proxy variable problem. You don't need to input gender to get a gender-discriminatory result. If the data you input correlates with gender β€” types of stores shopped at, spending patterns, historical credit behavior shaped by past discrimination β€” the model can learn to discriminate without the discriminating variable ever being explicitly named.
The problem is proxy discrimination. A model doesn't need the variable "gender" to learn gender-correlated patterns from data. Other variables β€” purchase categories, credit history shapes, behavioral data β€” can act as proxies for gender and produce discriminatory results without the word ever appearing in the code.
3. Explainability tools like LIME and SHAP provide which type of information?
Correct. LIME and SHAP approximate, they don't reveal. They tell you "changing this input seemed to matter most for this output" β€” which is useful but not the same as reading the actual reasoning. It's correlation-based attribution, not causal understanding.
These tools don't open the black box β€” they build a simpler approximation around it. They identify which inputs appear most influential for a given output, but they can't show you the actual internal reasoning of a deep neural network, because that reasoning isn't stored in any human-readable form.
4. A researcher argues: "The most capable AI systems and the most interpretable AI systems are fundamentally different β€” and we may not be able to have both at once." If true, what would be the hardest implication of this for society?
Correct. If the capability-interpretability tradeoff is real, then every decision to use a powerful black-box AI in high-stakes contexts (hiring, sentencing, medical diagnosis) is also a decision to accept unaccountable power. That's not a technical decision β€” it's a political and ethical one.
The deepest implication is a genuine tradeoff between capability and accountability. If you can't have both maximum power and full transparency, then every deployment of a powerful AI in a high-stakes context is also a choice about how much unaccountable decision-making you're willing to allow.
5. You are applying to a university and get rejected. An administrator tells you the decision was partly based on an AI scoring system and that the system cannot explain why it scored you the way it did. Based on this module, what is the most important right that this situation potentially violates?
Correct. The right to explanation is both procedural (you can't challenge what you can't see) and systemic (audits for discrimination require being able to examine decisions). An AI system that can't explain itself removes both protections simultaneously.
The core issue is explainability as a right. If a decision significantly affects your life and you can't understand why it was made, you can't contest errors and no one can audit for discrimination. The opacity removes your ability to get a fair review β€” regardless of whether the underlying decision was actually correct.
Module 5 Β· Lab 4

The Explainability Hearing

You're testifying before a government committee on whether a black-box AI should be allowed in criminal sentencing. Your partner represents the AI company.

Your Scenario

A state government is deciding whether to allow a new AI risk-scoring system in criminal sentencing. The AI company says the system is 92% accurate at predicting reoffending β€” more accurate than any human judge. They say the internal model is proprietary. They are willing to provide summary explanations ("the top 3 factors in this score were...") but not the full model weights or training data.

You are testifying that this is insufficient. Your lab partner represents the AI company. They argue that accuracy should be the standard, not explainability. Make your case β€” and respond to their counterarguments.

Open your testimony: why is "92% accurate" an insufficient basis on which to deploy this system, and what standard would you require instead?
AI Company Representative β€” Sentencing Hearing
AI Peer
Our position is simple: 92% accuracy means this system makes fewer errors than human judges, who are subject to their own biases, mood, and fatigue. The research on human sentencing disparities is damning. Our system is more consistent. If accuracy is what protects defendants, why isn't accuracy enough?
Module 5 β€” Final Assessment

Module Test: When the Machine Gets It Wrong

15 questions across all four lessons. Score 80% or higher to pass. Apply concepts β€” don't just recall definitions.
1. Robert Williams was wrongly arrested in Detroit in January 2020. The underlying cause was:
Correct.
The cause was a face recognition AI match that investigators trusted without adequate independent verification.
2. Joy Buolamwini's Gender Shades study demonstrated that face recognition accuracy gaps were primarily caused by:
Correct.
The gaps came from skewed training data β€” not intentional design choices. The systems learned from predominantly lighter-skinned faces and were therefore less accurate on those underrepresented in training.
3. Amazon scrapped its AI hiring tool in 2018. What does this case best illustrate about where bias enters AI systems?
Correct.
The Amazon case showed that bias emerges from historical patterns in training data β€” not explicit code. The AI learned that men were hired more often and turned that historical fact into a prediction rule.
4. A well-calibrated AI system is one that:
Correct.
Calibration is about matching confidence to accuracy. A well-calibrated system's stated certainty tracks with how often it's actually right β€” not always right, but honest about its own uncertainty.
5. When ChatGPT was asked to verify the fake court cases it had generated for lawyer Steven Schwartz, it:
Correct.
ChatGPT confirmed the fabrications were real. This is the core of the hallucination problem β€” the model generates true and false content with identical fluency and has no internal alarm for when it is making things up.
6. The "accountability gap" in AI refers to a situation where:
Correct.
The accountability gap is the structural tendency for responsibility to be passed around β€” between makers, deployers, and users β€” until no one is left holding it when harm occurs.
7. PredPol's predictive policing feedback loop worked like this: police were sent to high-flagged zones β†’ more arrests occurred β†’ arrest data fed back into the system β†’ zones were flagged even higher. What was the fundamental flaw in this cycle?
Correct.
The flaw was that the data tracked enforcement, not actual crime. More police presence created more arrests, which looked like more crime, which justified more police presence β€” a loop that validated itself.
8. "Model collapse" is a risk that arises specifically from:
Correct.
Model collapse is the photocopy-of-a-photocopy problem: each generation of AI trained on AI-generated content loses fidelity, variety, and accuracy compared to the generation before it.
9. A new AI system for approving student loan applications claims to use "hundreds of behavioral and financial variables." A critic argues this might discriminate even without using race as an input. The most accurate explanation of this concern is:
Correct. Proxy discrimination is a documented phenomenon. Variables like zip code, spending categories, credit history, and school attended can all correlate strongly with race due to historical structural inequality β€” allowing a system to discriminate effectively without using a protected category directly.
Proxy discrimination is the key concept here. Variables that aren't race can still be correlated with race because of historical inequality β€” meaning a system can discriminate by race through indirect routes, without the word "race" appearing anywhere in its code.
10. What is the key difference between how a simple rule-based AI and a modern deep learning system handle explainability?
Correct.
Rule-based systems are interpretable β€” you can read the rules. Deep learning systems are not β€” the decision emerges from billions of numerical weights that no one designed individually and no one can fully decode after training.
11. The UK Home Office scrapped its visa algorithm in 2020 after what kind of pressure?
Correct.
Civil society organizations spent years applying legal pressure before the algorithm was quietly dropped in 2020. No full technical audit was ever made public.
12. What does the COMPAS case reveal about how AI scores can be misused in high-stakes decisions?
Correct.
COMPAS combined three problems: racial bias in outcomes, opacity in methodology, and legal weight in sentencing β€” while the company hid behind trade secret protection to prevent any external audit.
13. Why might a company use "we're working on explainability" as a strategy to resist stronger AI regulation?
Correct. Promising future work can function as a delay tactic β€” it signals good intentions without actually delivering accountability, and it can be used to argue against regulation being "needed right now."
The strategic value is delay. Claiming to be working on explainability creates the impression of responsibility while deferring any actual binding requirements. Meanwhile, deployment continues on existing opaque systems.
14. An AI content moderation system flags posts for removal. After six months, researchers notice it removes posts in minority languages at twice the rate of English posts, even for similar content. The most complete explanation is likely:
Correct. This scenario combines two failure modes from this module: training data bias (less data for minority languages means worse performance) and potential feedback loops (early false removals become training signal for future removals).
This combines training data bias (less minority language data means worse performance) with potential feedback loops β€” once the system starts removing minority language content at higher rates, those removals can become training signal reinforcing future removals.
15. Which of the following correctly summarizes the deepest tension at the heart of the black box problem?
Correct. This is the deepest tension in the module: capability and interpretability may be in genuine conflict, not just an engineering problem waiting to be solved. If that's true, deploying the most powerful AI systems requires accepting unaccountable decision-making β€” which is a political and ethical choice, not just a technical one.
The deepest tension is that power and interpretability may conflict fundamentally. The same complexity that gives deep learning systems their capabilities is what makes them opaque. You may not be able to have both maximum performance and full transparency at the same time.