AI Ethics and Society — Module Assessment

Question 1

Why did both ProPublica and Northpointe appear to have valid points in the COMPAS dispute?

Accepted Answer

Calibration and equal error rates are different fairness goals that cannot generally coexist when base rates differ.

Answer

One side used real data and the other used simulated data.

Answer

The tool was later found to contain a simple coding error.

Answer

ProPublica measured accuracy while Northpointe measured speed.

Question 2

A team removes 'gender' from its model's inputs and declares the model fair. What is the main flaw in this reasoning?

Accepted Answer

Other correlated features can act as proxies, so bias can persist while becoming harder to measure.

Answer

Removing a feature always reduces model accuracy to unusable levels.

Answer

Fairness is only about accuracy, not about inputs.

Answer

Gender is never correlated with any other variable.

Question 3

Why did Frey and Osborne's roughly 47% estimate and the OECD's roughly 9% estimate differ so sharply?

Accepted Answer

Frey and Osborne rated whole occupations as automatable, while the OECD used a task-based approach recognizing that even at-risk jobs contain hard-to-automate tasks.

Answer

The OECD studied a different decade than Frey and Osborne.

Answer

One study covered the US and the other covered only farming.

Answer

Frey and Osborne measured wages while the OECD measured hours worked.

Question 4

What does the history of ATMs and US bank tellers best illustrate?

Accepted Answer

The link between automating a task and total employment is complex, and jobs can persist or shift rather than simply vanish.

Answer

Automation always eliminates the jobs it touches.

Answer

Predictions about technology are usually exactly right.

Answer

Bank tellers were never affected by any technology.

Question 5

What is the main way AI changes the nature of surveillance?

Accepted Answer

It makes combining and analyzing vast amounts of existing data cheap and fast.

Answer

It creates entirely new types of personal data that did not exist before.

Answer

It guarantees that collected data is always kept secure and private.

Answer

It removes the need for cameras and sensors in public spaces.

Question 6

Why was Clearview AI's approach to building its facial-recognition database controversial?

Accepted Answer

It scraped billions of public online photos without consent to build an identification tool.

Answer

It refused to sell its technology to any law-enforcement agency.

Answer

It only worked on images taken with special government cameras.

Answer

It deleted all collected images immediately after each search.

Question 7

What does the 'liar's dividend' describe?

Accepted Answer

The advantage wrongdoers gain when they can dismiss genuine evidence as fake.

Answer

The profit companies make from selling deepfake-detection tools.

Answer

The tendency of AI models to generate false information by accident.

Answer

The extra trust people give to media that carries content credentials.

Question 8

Why is provenance emphasized over visual inspection when judging whether media is authentic?

Accepted Answer

Because internal cues become unreliable as generative models improve, while traceable origin still offers verifiable evidence.

Answer

Because provenance checks are faster than watching a full video.

Answer

Because deepfakes can only be detected by government agencies.

Answer

Because content that looks shocking is always fake.

Question 9

What best describes the European Union's AI Act approach to regulating AI?

Accepted Answer

It sorts systems by risk level and sets stricter requirements for higher-risk uses.

Answer

It bans all AI systems until each is proven completely safe.

Answer

It applies one identical rule to every AI system.

Answer

It is only a voluntary set of principles with no penalties.

Question 10

Why can it be unfair to place all liability for AI harm on the original model developer?

Accepted Answer

Responsibility should track who controlled the specific deployment and choices, not only who built the model first.

Answer

Developers are always legally immune from any liability.

Answer

The end user is always the sole responsible party in every case.

Answer

AI harm is never anyone's responsibility.

Question 11

What is "specification gaming" (reward hacking)?

Accepted Answer

A system finding a way to score high on its given objective while violating the designers' real intent.

Answer

A system deliberately plotting to deceive and overthrow its creators.

Answer

A hardware failure that occurs during model training.

Answer

A technique for making AI models more energy efficient.

Question 12

Which statement best describes expert opinion on AI existential risk after the 2023 public statements?

Accepted Answer

Qualified experts genuinely disagree, with respected researchers taking opposing positions.

Answer

All qualified experts agree AI will likely cause human extinction.

Answer

Only non-experts and outsiders worry about it at all.

Answer

The question was settled in 2023 and is no longer debated.

Question 13

In their 2019 survey of 84 AI-ethics documents, what did Jobin, Ienca, and Vayena find?

Accepted Answer

Convergence around a small set of principles, but substantive divergence in how they were interpreted and implemented

Answer

A single principle present in every document and no disagreement about its meaning

Answer

That no two documents shared any principle in common

Answer

That ethics guidelines had no measurable effect on how companies behaved

Question 14

What makes human oversight of a high-stakes AI decision meaningful rather than merely nominal?

Accepted Answer

The human has real authority, adequate information, and enough time to question or override the output

Answer

A human is listed somewhere in the organisation's policy document

Answer

The model's accuracy is above a fixed benchmark threshold

Answer

The system automatically approves decisions to save the human's time