Summary of "Algorithmic Decision Making - 2"

AI assistant vs AI agent

AI assistant
- Intelligent support tools such as Alexa and chatbots.
- Provide help but do not autonomously decide or act on the user’s behalf.
- Typically have session-limited memory (no persistent user memory across sessions unless explicitly designed).
- Prompt-driven and often rule-based or algorithmic; prompt engineering is important (covered later).
AI agent
- Autonomous systems that design and execute workflows on behalf of users (examples: automated trading systems, autonomous vehicles).
- Characterized by multicomponent autonomy, persistent memory (tracks historical interactions), and decision-making capability.
- Act as substitutes for human agents/decision-makers.

The speaker frames these five axes as the main dimensions scholars use to compare machine and human decision-making:

Specificity of decision/search space
- Algorithms generally require well-specified inputs and goals. Humans can often act on vague or underspecified instructions (until AGI advances).
Interpretability / explainability
- Many machine-learning models (especially deep neural nets) are black boxes; human decisions are often more explainable. Explainability (XAI) is a critical managerial issue.
Size of alternative set and bounded rationality
- Algorithms can evaluate far larger solution sets than humans, who are boundedly rational.
Decision speed and replicability
- Machines are faster and more reproducible (same inputs → same outputs); humans vary.
Overall comparison framing
- These axes together are used to weigh when machines should make decisions versus humans.

Important warning: models may learn correlations that are not causal; using such predictors for decisions can be misleading or harmful.

Examples discussed:

Facebook / Starbucks (HBR example)
- Users who “liked” Starbucks spent 8% more — this could reflect pre-existing preferences (confounding), not a causal effect of the like.
Running speed vs sneaker color
- A learned correlation between sneaker color and performance was actually an artifact of supplier color codes tied to shoe size; shoe size correlates with speed. This is a local spurious correlation that harms generalization.

Practical implications:

Require careful causal thinking, experimental controls, and validation before using correlations for managerial decisions.

Overfitting
- Flexible models can learn spurious, local patterns and fail to generalize.
Bias–variance trade-off
- Less flexible models can be biased but may generalize better; highly flexible models reduce bias but risk overfitting.
Data and model quality issues
- Algorithms produce biased or wrong outputs if predictors, data, or model choices are poor. Developers and managers must attend to data quality, feature selection, and model validation.

Primary speaker: course lecturer (unnamed) presenting literature and examples
Cited organizations and works mentioned in the talk:
- IBM (practical insights on AI assistants vs agents)
- California Management Review paper (referred to as “Shaa and others”)
- Harvard Business Review example/study (referred to as “John and others” about Facebook/Starbucks)
- Barukas and others — chapter from Fairness and Machine Learning (“When is automated decision-making legitimate?”) discussing correlational pitfalls

No further action requested.