AI Ethics

225 flashcards · Other · by Repo-SK

What is a primary benefit of using AI scribes in clinical settings, even if they don't save time due to the need for corrections?
They can reduce the cognitive load and administrative burden on clinicians
What is a key ethical dilemma when considering whether to use AI to enable less-skilled clinicians to perform tasks in developed countries, a practice already being used to increase capacity in developing countries?
The dilemma of resource allocation: Should we prioritize reducing costs in developed nations or providing essential care in developing ones? It also raises questions about whether this lowers the standard of care.
The observation that most FDA-approved AI tools in healthcare are in fields like radiology and cardiology is largely attributed to what factor?
The relative ease of obtaining large
Which of the following are proposed initiatives to improve and regulate transparency in AI systems? (Select one or more answers)Open Sourcing AI models Conducting AI Impact Assessments Establishing certification and labeling standards Keeping all algorithms proprietary to encourage competition
Open Sourcing AI models Conducting AI Impact Assessments Establishing certification and labeling standards
What is 'contractual trust' as it applies to AI systems?
It is the trust a human places in a machine to fulfill a specific
An XAI method that can be applied to any type of machine learning model
regardless of its internal structure
In the famous example of an AI classifier mistaking a husky for a wolf, the post-hoc explanation revealed the model was focusing on the snow in the background, not the animal's features. This illustrates a failure to learn the correct causal relationship.
causal
A saliency map that highlights important pixels in an image after a deep neural network has classified it is an example of a post-hoc explanation.
post-hoc explanation
What is the 'inherent trade-off' in algorithmic fairness?
It refers to the trade-off that often exists between the fairness of an algorithm (e.g.
Amazon's historical attempt to build an AI recruitment tool resulted in a system that penalized female candidates. The tool was trained on 10 years of company data, which was dominated by male resumes. This is a well-known case study of data-driven bias.
data-driven bias
What is 'procedural fairness' in the context of AI systems?
Ensuring that the process of making and implementing algorithmic decisions is morally and politically acceptable to those affected
A simple decision tree with clear if-then rules is an example of an interpretable model.
interpretable
What is the key difference between interpretability and post-hoc explainability in XAI?
Interpretability refers to models that are inherently understandable by design (e.g., a decision tree), whereas post-hoc explainability refers to creating an explanation for an opaque, 'black box' model after it has made a decision.
According to the lecture
AI transparency is crucial for promoting trust
A person follows the advice of a new, untested AI financial advisor and loses money. Their trust in the AI could be described as unwarranted trust.
unwarranted trust
The concept that "clinical data are formed, not found" implies that medical data is not a pure reflection of a patient's state, but is shaped by what?
It is shaped by the context of care
What is the core aim of AI or algorithmic transparency?
To make the functionalities
An explanation for why a specific individual's loan application was denied by an AI model is an example of a local explanation.
local
Which of the following is NOT one of the four ethical 'red flags' for AI use discussed in the lectures? Options:The decision-maker has power over the subject. The subject distrusts the AI. The AI system is not 100% accurate. The decision-maker would not accept the AI being used on themselves.
The AI system is not 100% accurate.
In the context of AI accountability, what does transparency primarily refer to?
Being open
A significant challenge in AI is that models are adept at finding statistical associations, but not necessarily the underlying reasons for those associations. This is the challenge of distinguishing correlation from causation.
correlation; causation
What ethical concern was raised by the use of an AI voice emulator in the documentary about Anthony Bourdain?
The AI generated his voice to speak words he had written but never spoken aloud, raising issues of authenticity, consent, and a lack of clear disclosure to the audience.
According to Paakk Wong's Accountability for Reasonableness framework, the condition that decisions about algorithmic fairness and their rationales must be publicly transparent and accessible is known as the publicity condition.
publicity condition
An operator at a nuclear power plant begins to ignore an alarm system because it has a history of frequent false alarms. This behavior is a prime example of what concept? Options: * Misuse * Disuse * Abuse * Automation Bias
Disuse
Which ethical theory
focusing on the greatest good for the greatest number
What is the key idea behind the "impossibility theorem" regarding algorithmic fairness?
It is mathematically impossible for a single algorithm to simultaneously satisfy several different technical definitions of fairness (like statistical parity and equal opportunity).
"Philosopher Andy Clark suggests that the human brain is not well-designed for complex planning and logic, but is instead good at pattern recognition and physical actions. He summarized this by saying the brain is ""bad at logic and good at frisbee""."
frisbee
An AI sepsis detector trained on US hospital data performed poorly in Australia. What was a key reason for this failure?
The underlying data was different; a test ordered frequently and early in the US was ordered later and less commonly in Australia
The principle of "databases don't travel" refers to what phenomenon in medical AI?
An AI model trained on data from one specific clinical population or setting may perform poorly when used in a different one due to differences in medical practices
What is the 'moral crumple zone'?
A concept describing the tendency to blame the human operator closest to an AI system's failure
In XAI, what is the difference between a local and a global explanation? Options: * Local explains one model; Global explains multiple models. * Local explains a specific decision for one instance; Global explains the model's overall behavior. * Local is for non-experts; Global is for data scientists. * Local is simple; Global is complex.
Local explains a specific decision for one instance; Global explains the model's overall behavior.
What is a potential negative consequence of an AI tool that provides an "explanation" for its decision, as seen in the mammography example where the AI's outlines of masses were very different from a human's?
The explanation can be so counter-intuitive or ""crazy"" to the human expert that it causes them to lose trust and discount the AI's (potentially correct) recommendation.
According to philosopher Andy Clark, the distinctive feature of human intelligence is our ability to form "information processing mergers" with non-biological tools like pen and paper, or computers.
information processing mergers
A study on AI-assisted mammogram reading in Sweden replaced one of the two human radiologists with an AI. What is the primary ethical and economic justification for this approach?
It maintains the principle of a 'second opinion' to improve accuracy while potentially reducing the cost and workload associated with requiring two highly-paid specialists.
In the context of diagnosing melanomas, the "E" in the ABCDE rule stands for Evolution (how a mole changes over time). Why is this aspect particularly difficult for many AI models to incorporate?
Because historical data (e.g.
Bias that arises from assumptions and simplifications made within the AI model's code itself, such as problematic feature selection, is known as algorithmic bias.
algorithmic bias
Which historical phase of AI in medicine was based on Bayes' theorem to make rational decisions under uncertainty, but was hampered by the difficulty of estimating probabilities?
Probabilistic AI (prominent in the 1960s).
A face detection camera that fails to correctly identify individuals of a certain race because the training data was not diverse enough is a primary example of data-driven bias.
data-driven bias
Which ethical theory would argue against a non-transparent AI grading system on the grounds that it treats students as a 'means to an end' and violates their dignity and autonomy?
Deontological ethics
A situation where an autonomous system causes significant harm, but it is difficult or impossible to assign blame to any single human agent (e.g., the programmer, user, or commander), is known as the responsibility gap.
responsibility gap
What is the 'Uncanny Valley' in the context of robotics and AI avatars?
It is a phenomenon where a humanoid object that is imperfectly human-like provokes feelings of unease or revulsion in observers.
An AI tool that can guide a nurse in a developing country to perform an ultrasound is an example of AI being used for what purpose? Options: * Replacing all doctors * Increasing healthcare accessibility and capacity in resource-limited settings * Enforcing data privacy * Discovering new drugs
Increasing healthcare accessibility and capacity in resource-limited settings
The idea from J.C.R. Licklider in 1960 was not to replace humans, but to create a tight coupling of human brains and computing machines. The goal was that the resulting partnership would think as no human brain has ever thought.
think as no human brain has ever thought
Which type of AI bias occurs when the training data used is not representative of the broader population, leading to skewed or unfair outcomes?
Data-driven bias
Which historical phase of AI in medicine focused on using rules and logic elicited from human experts to build systems that replicate their decision-making?
Logic/Expert Systems (prominent in the 1970s-1980s).
What is the definition of algorithmic bias?
Systematic and repeatable patterns or errors in an AI system that create unfair outcomes, such as privileging or discriminating against a particular group of users.
The epistemological sense of explicability answers the question 'How does it work?', while the ethical sense of explicability answers the question 'Who is responsible for the way it works?'.
Who is responsible for the way it works?
The concept that it is mathematically impossible for an algorithm to simultaneously satisfy all popular definitions of fairness (e.g., accuracy, statistical parity) is known as the impossibility theorem.
impossibility theorem
What is 'automation bias'?
A cognitive bias where humans tend to over-rely on or put too much faith in automated systems
What is prejudice bias in the context of AI?
A form of bias where human or societal prejudices are embedded into an AI system, for instance, through the prejudiced labeling of otherwise representative data.
The famous case of the Therac-25 radiation therapy machine, which delivered fatal overdoses due to a combination of software errors, poor user interface design, and operator over-trust, is a classic example of what?
The failure of multiple aspects of the human-machine system
The ProPublica investigation into the COMPAS algorithm, used for predicting recidivism, found that it had a disparate impact on African American defendants.
disparate impact
What are the four main principles of the Web Content Accessibility Guidelines (WCAG) from the World Wide Web Consortium (W3C)?
Perceivability Operability Understandability Robustness
In the context of social media
what is a primary transparency concern regarding content moderation?
Which XAI method provides explanations in the form of 'if-then' statements? Options: * Attribution-based explanations * Rule-based explanations * Example-based explanations * Counterfactual explanations
Rule-based explanations
The principle that individuals should have a way to challenge decisions made by an AI system is known as contestability.
contestability
What was the primary ethical issue with the 2014 Facebook "emotional contagion" study?
It manipulated the emotional content of users' news feeds to study their reactions, all without their informed consent.
In the context of criminal justice
what is the potential negative outcome of using 'predictive policing' AI systems?
An AI tool that classifies a skin lesion as 'benign' might be overridden by a clinician for what reason that the AI cannot perceive?
37012
The fact that human working memory is limited to about 5 items, while modern medical decisions can involve hundreds of relevant facts, provides a strong rationale for using AI as a(n) information processing merger.
information processing merger
Which of the following stakeholders is primarily concerned with debugging a model and improving its performance using XAI? Options: * Regulator * Consumer * Business Owner * Data Scientist
Data Scientist
What major ethical and security risk was demonstrated by researchers who were able to add or remove signs of cancer from CT scan images without being detected by radiologists?
The vulnerability of medical data to manipulation and the critical need for data provenance and security.
The case of the UCSF pharmacy robot that dispensed a 39-times overdose illustrates a key limitation of AI systems. What is this limitation?
The lack of common sense to recognize when an instruction is highly unusual or absurd.
The fairness criterion that focuses on ensuring that the true positive rate is the same across different groups (i.e., focusing on fairness among those who are genuinely qualified) is known as Equal Opportunity.
Equal Opportunity
A tech company knowingly deploys a flawed AI system for public use with an interface that is confusing for operators, leading to harmful outcomes. This is best described as an abuse of automation.
abuse
The fundamental theorem of biomedical informatics states that a human and a computer, working together, should make better decisions than either would make alone.
either would make alone
The idea that a human coupled with a computer can make better decisions than either could alone is known as:
The fundamental theorem of biomedical informatics.
What is the key difference between the concepts of equality and equity?
Equality means giving everyone the same resources (e.g., an equal share), while equity means giving individuals the specific resources they need to achieve a fair and equal outcome.
The lack of clarity about the origin and collection methods of data used to train a generative AI model is a problem related to data provenance.
data provenance
Alfred North Whitehead's quote, "Civilization advances by extending the number of important operations we can perform without thinking about them," can be used to justify the goal of AI in what way?
It supports the idea of using AI to automate routine cognitive tasks
According to the lecture, what are two major overarching challenges in healthcare that motivate the use of AI?
Healthcare is both very expensive (often the largest single expense for a developed country) and dangerous (comparable to activities like mountain climbing).
What is the general ethical consensus regarding AI systems
such as customer service chatbots
Which of the following is a key challenge for Explainable AI? Options: * Most AI models are too simple to require explanation. * AI models excel at finding causation, but not correlation. * The internal mechanisms of complex models like deep neural networks are often opaque. * There are too many stakeholders interested in explanations.
The internal mechanisms of complex models like deep neural networks are often opaque.
What is the key distinction between 'trust' and 'trustworthiness' in AI?
Trust is a subjective belief a person has in a system. Trustworthiness is an objective property of the system itself
The principle of explicability is described as having two key components: intelligibility (how it works) and accountability (who is responsible).
explicability
Which of the following describes the fairness criterion of "Statistical Parity" (or Demographic Parity)? Options: * Similar individuals should receive similar outcomes. * The probability of a positive outcome should be the same across all groups. * The true positive rate should be equal across all groups. * The outcome for the worst-off group should be maximized.
The probability of a positive outcome should be the same across all groups.
What was a key defense offered by North Point, the creators of the COMPAS algorithm, against accusations of racial bias? Options: * The algorithm was more accurate for white defendants. * The algorithm had the same overall accuracy for both Black and white defendants. * The algorithm's source code was fully transparent. * Race was the primary factor used in its predictions.
The algorithm had the same overall accuracy for both Black and white defendants.
The problem where the vast majority of data used to train AI models comes from a few developed countries (like the US) is sometimes described by the phrase: "The Ghost in the Machine has an American accent."
American
What is "accessibility" in the context of technology?
Technology is considered accessible if it can be used as effectively by people with disabilities as by those without.
The concept of designing a product to be as accessible as possible to the widest possible group of people, including those with and without disabilities, is known as universal usability.
universal usability
A ramp designed for wheelchair users that also benefits people with prams, delivery trolleys, and bicycles is a classic example of what design concept?
Universal Usability
According to the lecture, a proposed framework for AI in society adapts four principles from biomedical ethics and adds a fifth. What is this fifth principle, which is seen as enabling the other four? Options: * Profitability * Efficiency * Explicability * Innovation
Explicability
The approach of using typical, representative examples (prototypes) and unusual examples (criticisms) to explain a model's understanding of a category is a form of example-based explanation.
example-based
An AI tells a user, 'The model predicted 'fly' instead of 'beetle' because the input image had 5 eyes, whereas beetles have 2.' This is an example of a contrastive explanation.
contrastive
What is a counterfactual explanation in the context of XAI?
An explanation that describes the smallest change to the input features that would alter the model's decision to a different, desired outcome (e.g., 'Your loan would have been approved if your annual income had been $5,000 higher').
XAI techniques like LIME and SHAP, which determine the contribution of each input feature to a model's prediction, are examples of what type of explanation method? Options: * Rule-based * Attribution-based * Example-based * Prototype-based
Attribution-based
The CORALS algorithm was presented as an example of an interpretable model that could achieve similar performance to a black-box system (Compass) for predicting criminal recidivism. What does this suggest? Options: * There is always a trade-off between a model's accuracy and its interpretability. * The belief in a necessary trade-off between accuracy and interpretability might be a myth, especially for structured data. * Black-box models are always superior. * Interpretable models are only usef…
The belief in a necessary trade-off between accuracy and interpretability might be a myth, especially for structured data.
What is the core principle of Utilitarianism?
To act in a way that maximizes overall good or 'utility'
Deontology is which of the following? Options: * A consequence-based ethical theory * A duty-based ethical theory * A character-based ethical theory * A relationship-based ethical theory
A duty-based ethical theory
What is the derogatory term for feigning virtue to enhance one's social standing, rather than reflecting genuine character?
Virtue signalling
In Aristotle's doctrine of the golden mean, the virtue of courage is the mean between the vices of cowardice and recklessness.
courage
In Aristotle's doctrine of the golden mean, courage is the virtue between which two vices?
The vice of deficiency (cowardice) and the vice of excess (recklessness).
According to virtue ethics, an action is right if and only if it is what a virtuous agent would characteristically do in the circumstances.
virtuous agent
In contrast to utilitarianism
deontology is a non-consequentialist theory.
Deontology asserts that the ends (consequences) do not necessarily justify the means (the actions taken).
means
According to deontology
some actions are considered inherently right or wrong regardless of whether they maximise utility.
The philosopher Peter Singer used utilitarian principles to argue for the moral consideration of non-human animals
on the basis that they can experience pleasure and pain.
A researcher identifies three categories for AI transparency. Which category covers how a system is embedded and used within an organization? Options: * System transparency * Procedural transparency * Transparency of outcomes * Algorithmic transparency
Procedural transparency
The Australian AI ethics principle concerned with ensuring people can challenge the outcomes of an AI system is called Contestability.
Contestability
What is Aristotle's doctrine of the 'golden mean'?
The idea that a virtue is a balanced 'sweet spot' or mean between two vices: the vice of deficiency and the vice of excess.