Lecture 7 + Lecture 7 (IDK)

It refers to the way humans shape machine behaviour and machines shape human behaviour as they interact together.

Humans often assign intelligent intentions or emotions to moving objects or AI systems.

Being polite but firm tends to promote better responses from ChatGPT.

Humans’ expectations about how AI should behave influence how much they trust it.

People trust machines for objective tasks that follow rules and explicit reasoning, such as solving numerical problems.

People distrust machines for subjective tasks that require intuition, instinct, or emotional judgment, like hiring decisions or handling emotions.

Tangibility, transparency, and reliability.

It refers to whether the AI system can be perceived or physically interacted with.

It’s the degree to which the AI’s operating rules and inner logic are apparent to the user.

It’s the consistency of the AI’s behavior over time in performing expected tasks.

Non-verbal or psychological cues such as positive affect, showing interest, or eye contact that affect user trust.

It starts low but increases as humans interact with the robot.

It starts high due to high expectations but often decreases with continued use.

It’s when increasing human-likeness in robots first increases empathy but then suddenly causes discomfort or revulsion.

Because people realize the AI’s representation doesn’t match its actual intelligence, lowering confidence in it.

AI designed to help users understand how its decisions are made.

It supports error analysis and reveals biases in data.

Humans often explain decisions after making them, influenced by biases, making explanations unreliable.

It explains how the whole model works.

It explains why the AI made a specific individual prediction.

They can be misleading if they don’t truly match the complex model’s internal reasoning.

“What is the model actually learning?”

A wolf vs dog classifier that learned to distinguish snow from grass instead of the animals themselves.

Because sensitive variables often correlate with other variables like zip codes or income.

The person delivering the training data, the machine learning researcher, or the policymaker.

AI technology can be used for both beneficial purposes, like airport security, and harmful ones, like surveillance abuse.

Using public social media data for research without explicit consent.

AI models like ChatGPT may reuse existing works and sometimes are even listed as authors.

Dehumanization, such as creating fake friends or partners.

The tram (trolley) situation, where AI must decide whom to save in emergencies.

It proposes a legal framework to regulate AI risks and ensure responsible AI development.

Technical/engineering, human-centered solutions, and policy solutions.

A proposed regulation giving users the right to understand how AI made a specific decision.

Tangibility

Transparency (explainable AI)

Reliability

Task characteristics

Immediacy behaviors

LIME (Local Interpretable Model-Agnostic Explanations)

SHAP (SHapley Additive exPlanations)

Identify the decision point and de-bias the actual decision, regardless of what the system learned from unfair data.

Change the biased data until it's more fair before training the model.

By attributing intentions, beliefs, and emotions (Theory of Mind).

For objective, rule-based tasks (vs. subjective, intuitive ones).

Trust in robotic AI may start low and increase, while trust in virtual AI might start high and decrease upon errors.

When trust in an AI decreases after observing its limitations or errors, often starting from a high initial expectation.

The dip in familiarity or trust when humanoid robots look almost perfectly human but have subtle flaws.

AI decision processes can be complex and unlike human reasoning (and humans aren't always rational explainers either).

Using simpler models (like logistic regression) that are easier to understand by nature.

Methods used to explain existing complex models (like neural networks) after they are trained.

It creates perturbations (small changes) of an instance, gets predictions from the complex model, and fits a simple local model to explain that one prediction.

The fact that AI technologies can be used for both beneficial and harmful purposes (e.g., security vs. surveillance).

Investigating human perspectives on moral dilemmas for AI, such as self-driving car crash scenarios.

Because other data features are often correlated with the sensitive attributes anyway.