cs1302 midterm

Which of the following best defines Human-Computer Interaction?

An interdisciplinary field focused on designing systems that are usable and effective, understanding how people interact with technology, and studying how technology shapes people and society.

Which of the following is not typically part of HCI research?
A) UX design
B) Usability testing
C) Quantum computation modeling
D) Participatory design

Quantum computation modeling

What distinguishes HCI from computer science in general?

HCI centers on the human in the computing loop — it combines technical design with psychology, sociology, and design principles to ensure systems meet real-world human needs, not just technical efficiency.

Give an example of an HCI research method and its goal.

Usability testing—researchers observe users performing tasks to identify barriers and improve interface design for clarity, efficiency, and satisfaction.

Why is “considering the human” key to creating impactful technology?

Even powerful technologies fail without trust, accessibility, or alignment with human contexts. For example, an accurate cancer-detection model achieves little if clinicians don’t trust it or patients can’t understand results. HCI bridges technical capability and social adoption by ensuring systems fit into users’ workflows and values.

What differentiates sociotechnical systems from purely technical systems?

They include not only technical components but also human, cultural, institutional, and regulatory contexts that shape and are shaped by the technology.

Which is not an element of a sociotechnical system?

A) Law & policy B) Social norms C) CPU clock speed D) Culture

C – CPU clock speed

Provide an example of a sociotechnical analysis of a refrigerator.

Beyond mechanical cooling, consider privacy (smart-fridge data), environmental impact (energy use), and social factors like affordability or food-sharing programs — all demonstrate how design interacts with social systems.

Why might studying sociotechnical systems be essential for ethical technology design?

It reveals how technologies can reproduce or mitigate inequality, enabling designers to anticipate unintended consequences and align systems with societal good.

Explain how sociotechnical analysis complements HCI.

HCI focuses on usability and individual interaction, while STS widens the lens to institutions, laws, and cultural norms. Together they show that a design’s success depends on both user experience and broader systems—e.g., privacy law compliance, accessibility policy, or algorithmic bias governance.

What design issue occurs when users cannot tell two identical knobs apart on a microwave?

Poor signifiers. Labels or indicators are missing, so users can’t infer the correct action.

Which term refers to the perceived possibilities for action provided by an object?

Affordances

Which concept ensures users know their action had an effect?

Feedback

Differentiate between affordance and signifier.

Affordance = what actions are possible (a handle affords pulling). Signifier = the visual cue indicating how to act (a label or arrow showing “pull”).

Give one reason feedback is crucial for usability.

It closes the “gulf of evaluation” by confirming whether the system interpreted an action correctly—preventing confusion or repeated inputs.

How do the “gulfs of execution and evaluation” explain user frustration?

The gulf of execution arises when users can’t figure out what actions are possible; the gulf of evaluation arises when they can’t tell what happened after acting. Good design narrows both via clear signifiers, mapping, and feedback.

Which method is best suited to answering “how” and “why” questions about user behavior?

Qualitative methods (e.g., interviews, ethnography).

Which statement best describes participatory design?

A democratic process that redistributes design power by including stakeholders directly in decision-making and co-creation.

Contrast qualitative and quantitative HCI methods.

Qualitative = deep understanding of experience (context, meaning); Quantitative = patterns and statistical relationships across larger populations.

Give one reason surveys are valuable and one limitation.

They efficiently collect large-scale trends, but lack depth and context for interpreting why users behave a certain way.

How do participatory and speculative design expand the goals of HCI?

They move beyond usability toward empowerment and reflection—participatory design gives marginalized groups agency in shaping technologies, while speculative design imagines alternative futures to critique existing systems.

What question does discoverability address in design?
A) Can I tell what actions are possible?
B) Do I know how the system stores my data?
C) Does the object have good affordances?
D) Can I identify who designed the object?

A — It concerns whether possible actions are visible to the user.

Which of the following best defines an affordance?
A) A visual label showing how to use an object
B) The relationship between an object’s properties and a user’s abilities that determines how it can be used
C) The feedback a system gives after an action
D) The mental model a person forms of a system

B — Affordances emerge from object–user relationships

Which scenario demonstrates poor mapping?
A) Stove knobs arranged linearly control burners arranged in a square
B) A door with a “PUSH” label and a flat plate
C) A touchscreen button lights up when pressed
D) A phone flashlight toggles on and off with a tap

A — The spatial mapping between controls and functions is unintuitive

Which principle would best fix a user’s confusion about whether a file was deleted after pressing “Delete”?
A) Better conceptual model
B) Added signifiers
C) Improved feedback
D) Simplified affordance

C — Immediate, informative feedback prevents uncertainty

Which of the following most clearly illustrates an anti-affordance?
A) A disabled “Submit” button until all fields are complete
B) A door with a “PULL” handle that must be pushed
C) A screen reader reading “Error: invalid input” aloud
D) A bright “Power” icon indicating on/off

A — It deliberately restricts action, preventing misuse.

Explain the difference between actual and perceived affordances.

Actual affordances are what an object truly allows; perceived affordances are what users think it allows. Design succeeds when these align—for example, a handle both affords pulling and looks pullable.

Define a signifier and provide an example.

A signifier is a perceivable cue that communicates how to act—like a “PUSH” label or an arrow icon. It signals an affordance and guides correct use.

Why are conceptual models important in interface design?

They let users predict system behavior. A clear conceptual model (e.g., phone flashlight mimicking a real one) reduces cognitive load and errors

Describe mapping in your own words and give one example of a “natural mapping.”

Mapping links controls to outcomes. A car window button mapped vertically (up = window up, down = window down) is natural because it aligns with real-world movement

How does feedback reduce the “gulf of evaluation”?

Feedback shows whether an action worked, bridging the gap between user intention and system response (e.g., a visual flash when taking a photo).

Explain how unintended affordances can influence sociotechnical design outcomes.

Unintended affordances arise when users find novel or unplanned uses for a technology, revealing hidden relationships between design and social context. For example, rural Chinese users of Haier washing machines used them to wash vegetables, prompting Haier to redesign the product to accommodate this new function. These cases demonstrate that design is co-shaped by users, and that successful sociotechnical systems must adapt to emergent behaviors and contexts rather than enforcing designer assumptions

• 2 pts: Defines unintended affordance clearly

• 2 pts: Provides a concrete example and explains why it’s unplanned

• 2 pts: Connects example to broader sociotechnical insight (user agency, adaptability, or design ethics)

Compare the affordances, signifiers, and feedback of a search engine versus a conversational chatbot.
Discuss how these design differences influence user expectations, understanding, and trust.

Search engines afford keyword input and ranked results, supported by visible links and formatting cues that emphasize control and transparency. Chatbots, by contrast, afford conversational dialogue with turn-taking signifiers like typing indicators or text bubbles, evoking a sense of human-like understanding. These differing designs shape expectations: users may trust search engines for factual retrieval but overtrust chatbots due to their anthropomorphic feedback, leading to overreliance on generated content

• 2 pts: Correctly identifies key affordances, signifiers, and feedback mechanisms of each system

• 2 pts: Explains how these features shape user experience or expectations

• 2 pts: Connects to sociotechnical implications (e.g., trust, perception of intelligence, or usability)

Discuss how design shapes users’ experiences with technology beyond basic usability.
Use at least one example (e.g., placebo crosswalk buttons, chatbots, or search systems) to illustrate how design can influence user perception, behavior, or societal norms.

Design doesn’t just determine usability—it communicates power, agency, and values. For instance, placebo crosswalk buttons may not function but provide the illusion of control, reducing anxiety and shaping trust in civic systems. Similarly, chatbot design that uses empathetic language influences how users perceive AI as capable of understanding emotion. These examples reveal that design operates at social and psychological levels, affecting not just what people can do, but how they feel and what they believe about technology’s role in society

• 2 pts: Describes how design extends beyond usability (e.g., social/psychological influence)

• 2 pts: Provides a specific, relevant example

• 2 pts: Connects the example to broader implications (user trust, illusion of control, or ethics)

Imagine you are designing a new information-access system for students that combines chatbot-style Q&A with traditional search.
Discuss two design choices that would help balance efficiency and information literacy. Explain why each matters.

First, the system could display both chatbot-generated summaries and linked citations, encouraging users to verify sources rather than accepting answers blindly. Second, it could provide confidence indicators or uncertainty labels, teaching students to interpret AI-generated content critically. These strategies preserve the speed and convenience of conversational design while embedding transparency, promoting deeper learning and media literacy

• 2 pts: Mentions two concrete design strategies

• 2 pts: Connects each to IA goals (efficiency, literacy, or trust)

• 2 pts: Explains why these strategies matter in sociotechnical context (user understanding, ethics, education)

Which of the following correctly lists the main stages of the design cycle?
A) Ideate → Implement → Publish
B) Design → Prototype → Evaluate
C) Observe → Model → Deploy
D) Brainstorm → Research → Code

B — The standard design cycle is design → prototype → evaluate

Why is iteration essential in HCI design?
A) It helps systems learn from big data.
B) It compensates for unpredictable human contexts and diversity.
C) It allows designers to skip user testing.
D) It ensures every prototype is high-fidelity from the start.

B — Iteration helps designers progressively refine understanding of users and contexts

What best describes diverge and converge in design processes?
A) Alternating between code debugging and compiling
B) Exploring many ideas broadly, then narrowing down to refine the best ones
C) Switching between qualitative and quantitative methods
D) Using multiple teams to test the same prototype

B — Diverging generates ideas; converging narrows focus for refinement

Why do designers start with low-fidelity prototypes?

Low-fidelity prototypes (e.g., sketches, paper mock-ups) allow quick experimentation and feedback before committing resources. They emphasize idea testing over aesthetics

Explain one difference between design and evaluation stages in the design cycle.

The design stage defines problems and ideates possible solutions; evaluation tests prototypes with users to identify strengths, weaknesses, and improvements

What is the role of parallel prototyping in participatory or co-design?

It involves creating multiple prototypes simultaneously to compare alternatives. This approach fosters creativity, prevents fixation on one idea, and incorporates diverse stakeholder feedback

Describe how iteration and user feedback improve design outcomes.
In your answer, mention:

• Why iteration is necessary in HCI,

• An example of iteration from the lecture or your experience, and

• How user diversity affects the need for iteration.

Iteration allows continuous improvement through cycles of feedback and testing. Because people vary in needs, motivations, and contexts, no design can be perfect the first time. For instance, the co-design of activist data tools required repeated prototyping and critique to adapt to local needs and unreliable data sources. Iteration helps uncover these subtleties, turning assumptions into user-informed insights and more equitable solutions

• 2 pts: Explains what iteration is and why it’s central to HCI

• 2 pts: Provides a relevant example (e.g., activist data tools, iterative testing)

• 2 pts: Connects to human diversity, context, or sociotechnical complexity

In the activist feminicide data case study, what was the main sociotechnical design question?
A) How to automate violence reporting using AI
B) How to monetize data collection platforms
C) How technology can support and sustain civil-society data collection efforts
D) How to visualize data in 3D models

C — The goal was supporting grassroots counterdata initiatives

Which of the following is not a method used in participatory design or evaluation?
A) Co-design brainstorming
B) Wizard-of-Oz prototyping
C) High-fidelity mockups in Figma
D) A/B testing on anonymous web users

D — A/B testing lacks direct collaboration and shared decision-making

What does participatory evaluation add to the design process?

It empowers stakeholders to co-assess usability and impact, ensuring the system aligns with their lived realities rather than external metrics alone

Why is problem framing important before ideation?

It defines user needs, goals, and constraints—helping ensure design solutions address root causes rather than symptoms

Explain how participatory design addresses limitations of traditional top-down design.
Provide two benefits and one potential challenge.

Participatory design redistributes decision-making power by directly involving stakeholders—especially marginalized groups—in the design process. Benefits include creating contextually appropriate tools and fostering user ownership. For example, activists designing feminicide data platforms ensured technology supported—not replaced—their social mission. However, participatory processes can be time-intensive and require trust-building to manage differing priorities

• 2 pts: Identifies how participatory design differs from traditional methods

• 2 pts: Explains two benefits (empowerment, contextual relevance, trust)

• 2 pts: Notes one challenge (e.g., time, coordination, conflicting goals)

Which of the following is not a core promise of design thinking?
A) Human-centeredness through empathy
B) Quick iteration and experimentation
C) Universal solutions applicable to all contexts
D) Collaboration across disciplines

C — Assuming universality is a critique, not a promise

What is a major critique of corporate design thinking?
A) It encourages empathy and teamwork
B) It often depoliticizes social problems and ignores structural inequality
C) It rejects technology as a tool for change
D) It relies too heavily on government funding

B — Design thinking can oversimplify complex issues by focusing narrowly on usability or innovation

Which underlying value is associated with “solutionism” in corporate design thinking?
A) Viewing design as a political and contextual process
B) Believing complex social problems can be solved by technological fixes
C) Emphasizing long-term community collaboration
D) Prioritizing reflection over implementation

B — Solutionism assumes technology alone can solve deep societal issues

What does “superficial empathy” mean in critiques of design thinking?

t refers to shallow user research that captures surface-level feelings but misses deeper social or cultural contexts, leading to poorly grounded “empathetic” designs

How does design thinking’s optimism differ from speculative design’s goals?

Design thinking imagines bold, achievable solutions for present-day problems, while speculative design explores what-if futures to critique existing assumptions and spark reflection

Compare the promises and critiques of design thinking.
In your response, explain both why design thinking became influential and why scholars argue it can be limited or harmful when applied superficially.

Design thinking rose to prominence for its accessibility and human-centeredness, encouraging creativity, rapid iteration, and empathy in problem-solving. However, critics argue that corporate design thinking often depoliticizes issues, offering quick fixes to systemic problems without addressing underlying inequality. Its values of individualism and universalism assume neutrality and one-size-fits-all solutions. Meaningful sociotechnical design must pair design thinking’s optimism with structural awareness and long-term community engagement

• 2 pts: Identifies key promises (empathy, iteration, collaboration, optimism)

• 2 pts: Identifies core critiques (solutionism, depoliticization, superficial empathy)

• 2 pts: Connects analysis to sociotechnical or ethical implications

What are two alternative approaches beyond design thinking mentioned in the lecture, and how do they differ?

Alternatives include participatory design and speculative design. Participatory design involves users as co-creators, emphasizing power-sharing and contextual fit. Speculative design, on the other hand, challenges assumptions by creating conceptual artifacts that explore possible or undesirable futures. Both extend design beyond quick usability to reflection and empowerment

• 2 pts: Correctly names at least two alternatives

• 2 pts: Describes how each differs from design thinking

• 2 pts: Connects to broader design goals (e.g., empowerment, critique, reflection)

According to Langdon Winner’s argument in Do Artifacts Have Politics?, why is technology not value-neutral?
A) It always produces economic growth.
B) Technological artifacts embody and alter arrangements of power and authority.
C) Engineers typically work for governments.
D) Because users apply it for political campaigns.

B — Technologies carry and reshape social relations of power and control

What does “politics” refer to in Winner’s framework?
A) Partisan disputes in U.S. government
B) Arrangements of power, authority, and control in society
C) Moral values expressed by individual designers
D) Voter participation and elections

B — Politics refers to structures of power and authority, not party politics

Which statement best contrasts social determinism and technological politics?
A) Social determinism focuses on human context; technological politics emphasizes how artifacts themselves distribute power.
B) Social determinism is older than technological politics.
C) Technological politics ignores infrastructure.
D) They are identical concepts.

A — Social forces shape design choices; technological politics shows design itself can enact power relations

Which of the following best illustrates a “strong compatibility” relationship between technology and power arrangements?
A) A small chatbot designed by volunteers
B) Large language models requiring massive computing and corporate data resources
C) A calculator used in classrooms
D) Open-source community projects

B — The size and data requirements of LLMs align with centralized corporate control

Differentiate between “requirement” and “strong compatibility” in technological politics.

A technology requires a certain power arrangement when it cannot function without it; it is strongly compatible when it thrives within that structure but could exist differently (e.g., LLMs work best in corporate cloud environments though open alternatives are possible)

How does Winner’s theory challenge the idea that technology is “objective”?

He argues that technological artifacts embody values and arrangements of authority; design decisions reflect social and political choices rather than neutral engineering

Provide one example of how social determinism and technological politics would differ in explaining social media impacts.

Social determinism emphasizes market incentives and culture shaping platform use; technological politics focuses on how algorithmic feeds and design features actively redistribute attention and influence

What core question did Joy Buolamwini pose about facial analysis systems?
A) “Can we improve illumination to fix accuracy?”
B) “Should people change to fit technology or should technology fit people?”
C) “Can AI replace human photographers?”
D) “Do users understand deep learning?”

B — Her critique highlights how systems privilege some groups and force others to adapt

Why is “fixing bias” with diverse datasets insufficient for addressing FRT’s ethical issues?
A) Because bias is impossible to measure
B) Because accuracy improvements can reinforce surveillance and control
C) Because data diversity reduces model size
D) Because companies don’t care about ethics

B — Improving fairness can legitimize and extend surveillance systems rather than challenge them

The MEDS dataset used in facial recognition research primarily contains:
A) Crowdsourced selfies with consent
B) Mugshots of deceased individuals, disproportionately Black men
C) Passport photos from volunteers
D) AI-generated synthetic faces

B — Its origins link FRT to criminalization and racialized surveillance

Which statement best captures the “risk of fair FRT”?
A) It solves all bias issues.
B) It creates equal surveillance of everyone, which is not justice.
C) It makes face data completely private.
D) It reduces algorithmic transparency.

B — Fairness can mean broader and more efficient control, not equity

What historical roots does Simone Browne identify in modern biometric surveillance?

Browne connects current technologies to racialized surveillance practices from slavery and colonialism, showing continuities in how bodies are monitored and controlled

What does studying datasets as “sites of knowledge/power” reveal about FRT?

It shows that datasets aren’t neutral collections of faces but products of institutional and historical power — they shape who is seen and how they are categorized

Evaluate facial recognition technology through a sociotechnical lens. How does it reflect and reproduce systems of power? Propose one design and one policy intervention that could mitigate its harms.

Facial recognition systems emerge from long histories of racialized surveillance and centralized state control. Design choices like large, opaque datasets and limited user feedback reinforce hierarchies of visibility — who is seen and who is targeted. Improving dataset diversity is insufficient without addressing the social context of policing and data ownership. One design response could embed user audit logs and consent dashboards, while a policy intervention could enforce moratoria on public surveillance deployment. Together, these shifts challenge technological politics that concentrate power in states and corporations

• 2 pts: Identifies how FRT embodies power or control

• 2 pts: Connects design and dataset choices to social hierarchies

• 2 pts: Proposes viable design and policy mitigations and explains impact

Which of the following is a major concern about Amazon Ring doorbells?
A) They record too little data to be useful.
B) They enable participatory mass surveillance and racial gatekeeping.
C) They are illegal to sell in the U.S.
D) They cannot connect to law enforcement.

B — Ring spreads citizen-driven surveillance and reinforces inequality

Who ultimately retains significant control over Ring doorbell data?
A) The homeowner alone
B) Amazon (Ring’s parent company)
C) Local police departments
D) Independent review boards

B — Amazon retains storage and access rights, shaping data use and power distribution

List two stakeholders affected by Ring doorbells and how their interests might conflict.

Homeowners seek safety and property protection, while pedestrians and protesters may face privacy violations and profiling. These conflicting interests reflect unequal power over surveillance

How do Ring’s design and pricing contribute to social inequality?

High device and subscription costs limit access for low-income families, while data-sharing with police enhances surveillance in wealthier neighborhoods — reinforcing class and racial divides

Using the Ring Doorbell as a case study, analyze how design choices can redistribute power among stakeholders. Discuss who benefits and who is harmed, and recommend changes to make the technology more equitable.

Ring devices shift power toward corporations and law enforcement by outsourcing neighborhood watch functions to private surveillance infrastructures. Homeowners gain a sense of security, but pedestrians and communities of color experience increased policing and loss of privacy. Equity-oriented changes might include default data encryption, explicit opt-in police sharing, and community oversight boards. Such design interventions recognize technology as political and seek to balance power across stakeholders

• 2 pts: Identifies how Ring redistributes power and benefits/harms different groups

• 2 pts: Provides specific examples of design features affecting inequality

• 2 pts: Proposes actionable changes promoting equity or accountability

What is allocative harm in AI systems?
A) When systems stigmatize or stereotype groups
B) When resources or opportunities are denied to certain people
C) When data is measured incorrectly
D) When algorithms produce misinformation

B — Allocative harms arise when access to opportunities is restricted by algorithmic decisions

Which of the following is an example of representational harm?
A) An AI hiring tool rejecting qualified applicants
B) Search results reinforcing gender stereotypes
C) A facial recognition model misidentifying darker-skinned users
D) A navigation app underperforming in rural areas

B — Representational harm occurs when systems perpetuate stigma or stereotypes about groups

Which category of harm describes an algorithm that spreads misinformation and undermines democracy?
A) Quality-of-service harm
B) Interpersonal harm
C) Social system harm
D) Allocative harm

C — Social system harms affect society on a macro level (e.g., democracy, economy, environment)

A speech recognition system fails to understand certain accents. Which type of harm is this?
A) Allocative
B) Measurement
C) Quality-of-service
D) Representation

C — It performs worse for some identity group

Define interpersonal harm and provide one example.

Interpersonal harms arise when systems negatively affect relationships or well-being. For example, sharing non-consensual intimate images causes emotional and reputational damage

What distinguishes social system harms from other categories?

They operate at the macro level — affecting societal institutions like democracy, culture, and the economy — rather than individual users

Compare and contrast allocative and representational harms. Provide one example of each and discuss how both might co-occur in a single AI system.

Allocative harms occur when resources are unequally distributed — for example, loan approval systems rejecting certain demographics. Representational harms arise when systems reproduce cultural stereotypes, such as gendered language in search results. In practice, these often overlap: biased word embeddings can encode stereotypes (representational harm) that later influence hiring algorithms (allocative harm), reinforcing systemic inequality

• 2 pts: Defines allocative and representational harms accurately

• 2 pts: Provides one clear example of each

• 2 pts: Connects the two or explains how they interact in practice

Which type of bias persists even if data is sampled and measured perfectly?
A) Measurement bias
B) Historical bias
C) Aggregation bias
D) Evaluation bias

B — Historical bias reflects preexisting inequalities embedded in the world

Which form of bias arises when training data fails to represent key subgroups of the real-world population?
A) Representation bias
B) Evaluation bias
C) Measurement bias
D) Aggregation bias

A — Representation bias results from underrepresented or excluded populations in the dataset

When an algorithm’s label (like “creditworthiness”) is approximated using a flawed proxy (like “credit score”), which bias is this?
A) Measurement bias
B) Learning bias
C) Representation bias
D) Evaluation bias

A — The proxy misrepresents the underlying construct

If an AI model performs well on benchmark tests but fails on real-world data, what kind of bias is present?
A) Aggregation bias
B) Evaluation bias
C) Deployment bias
D) Measurement bias

B — Evaluation bias stems from benchmarks that don’t reflect deployment contexts

When an algorithm optimized for “healthcare cost” instead of “healthcare need” disadvantages certain racial groups, this illustrates:
A) Historical bias
B) Measurement bias
C) Deployment bias
D) Representation bias

B — The chosen proxy (cost) poorly captures the true construct (need)

Define representation bias and give one example.

Representation bias happens when development samples fail to represent the real-world population. For instance, datasets of college students may omit older or parenting students, leading to non-generalizable models

What does measurement bias mean in the context of AI data?

It occurs when proxies used for abstract concepts (like “success” or “risk”) are inaccurate or unequally applied across groups, such as using medical costs to measure health needs

Why is “more data” not always the solution to bias?

Larger datasets (e.g., Common Crawl) often reproduce structural inequalities — skewed toward younger, male, Western internet users — rather than fixing underrepresentation

Describe three types of bias in the machine learning pipeline (historical, representation, measurement). For each, explain its cause and one example.

• Historical bias arises from preexisting inequities reflected in data — e.g., word embeddings associating “woman” with “homemaker.”

• Representation bias occurs when the sampled data excludes key groups — e.g., geodiversity gaps in datasets that miss developing-world regions.

• Measurement bias results when proxies like credit scores or healthcare costs misrepresent the true construct, causing racial disparities in outcomes

• 2 pts: Accurately defines each bias type

• 2 pts: Provides one clear example for each

• 2 pts: Connects how bias emerges throughout the ML pipeline

What do word embeddings represent?
A) Image pixel distributions
B) Vectorized relationships among words based on contextual co-occurrence
C) Manual keyword tagging
D) Audio waveform similarities

B — Word embeddings model semantic similarity between words via co-occurrence statistics

Why do word embeddings encode stereotypes?
A) Because they’re trained on diverse data
B) Because they reflect the cultural and historical biases present in text corpora
C) Because they use random vectors
D) Because they rely only on human annotations

B — They mirror the societal patterns embedded in training data

Which study quantified a century of gender and ethnic stereotypes in word embeddings?
A) Buolamwini & Gebru (2018)
B) Garg et al. (2018)
C) Bender et al. (2021)
D) Gray & Suri (2019)

B — Garg et al. analyzed historical biases across 100 years of text data

Why is “debiasing” word embeddings only a partial solution?

Removing specific bias vectors doesn’t address the underlying cultural associations or hidden biases in large corpora — the system still inherits values from data creators

How do feminist scholars critique “universal” representations of knowledge in AI?

They argue all knowledge is situated — shaped by context, culture, and power — making truly universal or objective models impossible

Explain how word embeddings illustrate both the promise and the peril of “learning from data.” What do they reveal about the relationship between data, culture, and power?

Word embeddings demonstrate how machine learning can capture linguistic relationships, enabling translation and search. However, because they are trained on biased web text, they also encode stereotypes and reinforce systemic discrimination. Their “objectivity” is an illusion — they reproduce historical patterns of inequality. This highlights that data-driven systems reflect not only technical but also cultural and political choices

• 2 pts: Explains what word embeddings are and their utility

• 2 pts: Identifies how they encode bias or stereotypes

• 2 pts: Connects to broader social critique (knowledge, power, culture)

What does the term “ghost work” refer to?
A) Automated machine labeling
B) Invisible human labor behind AI systems such as annotation, moderation, and cleaning
C) Machine hallucination in LLMs
D) Synthetic data generation

B — It describes hidden human work enabling AI systems

What ethical issue is raised by ghost work?
A) Overrepresentation of technical workers
B) Exploitative, underpaid, and traumatic data annotation labor
C) Excessive transparency of workers’ identities
D) Overfitting in ML models

B — Data workers often face poor pay, precarity, and emotional harm

How can AI developers better acknowledge data labor?

By offering fair compensation, mental health support, crediting contributions, and enabling collective organizing among annotators

Discuss the ethical implications of “ghost work” in AI. Suggest two ways to make data labor more visible and equitable.

AI systems depend on vast, hidden human labor — content moderation, labeling, and data cleaning — often outsourced under exploitative conditions. Recognizing this labor reframes AI as a collective, not autonomous, achievement. Ethical reforms include requiring dataset transparency that names contributors, and establishing labor protections such as fair pay and collective bargaining. Such practices restore dignity and accountability to AI’s human infrastructure

• 2 pts: Identifies what ghost work is and why it matters

• 2 pts: Discusses ethical consequences (inequality, invisibility, precarity)

• 2 pts: Proposes two realistic strategies for improvementd

What is evaluation bias in machine learning?
A) Bias introduced when collecting training data
B) Bias caused by unrepresentative benchmark datasets or misleading metrics
C) Bias from user feedback loops after deployment
D) Bias that occurs during model optimization

B — Evaluation bias stems from non-representative benchmarks or metrics that hide subgroup errors