Class 6 Notes

Student Presentations: AI Applications Across Domains

10 Student PresentationsDiverse AI ApplicationsReal-World Testing

Class Overview

Student-led exploration of generative AI applications across diverse domains

Format: 8-minute presentations with Q&A sessions

Themes ranged from financial prediction to creative writing, education, and social impact

Key Learning: Practical insights into AI capabilities, limitations, and ethical considerations

Key Insights from Student Presentations

Diverse AI Applications

Students explored AI applications across 10 different domains, from finance and education to creative writing and emergency support

Human-in-the-Loop Essential

Every presentation highlighted the critical need for human oversight, expert validation, and ethical considerations in AI implementations

Limitations and Biases Identified

Students discovered significant limitations including cultural biases, overconfidence, and the challenge of assessing qualitative aspects

Student Presentations Summary

AI for Stock Price Prediction

Michael

NVIDIA Stock Forecasting with GPT-4

Key Findings:

  • AI confidently stated prediction was 'Mission Impossible'
  • Modified prompts yielded price ranges but were inaccurate
  • Real closing price ($160.0) fell outside AI's narrow prediction range
  • Highlighted prompt sensitivity and overconfidence issues

Key Lessons:

  • Pattern matching vs. true understanding
  • Market unpredictability
  • Importance of expert judgment

Methodology: Prompt engineering with different question formulations

Conclusion: AI useful for research and analysis but not reliable for specific predictions

AI in Banking and Financial Services

Nancy (Xiu)

Transforming Banking Operations

Key Findings:

  • AI automates time-consuming tasks like SEC filing analysis
  • 24/7 customer service capabilities with immediate responses
  • Enhanced compliance and risk control through data analysis
  • Three lines of defense ensure AI outputs are validated

Key Lessons:

  • Automation vs. human oversight
  • Risk management
  • Institutional safeguards

Methodology: Real-world banking experience and case studies

Conclusion: AI as powerful tool requiring multi-layer verification systems

AI for Learning Mathematics

Ethan

Taylor Series Learning with ChatGPT vs. DataWise

Key Findings:

  • Both AIs explained concepts well and provided additional insights
  • Failed to recognize student's deliberately incorrect answers
  • Math-specific AI (DataWise) promoted better learning through step-by-step guidance
  • 56.6% accuracy rate for ChatGPT as math tutor in studies

Key Lessons:

  • Specialized vs. general models
  • Verification challenges
  • Learning vs. answer-giving

Methodology: Comparative analysis of general vs. specialized AI tutors

Conclusion: AI effective as study partner but not ready as full tutor

AI for Creative Writing

Jason

Language Enhancement and Idea Generation

Key Findings:

  • Excellent for grammar checking and expression improvement
  • Can generate creative prompts and story ideas
  • Raises authorship and originality concerns
  • Grammarly-edited text often flagged as 'AI-generated'

Key Lessons:

  • Creativity vs. hallucination
  • Authenticity questions
  • Human creative input

Methodology: Personal writing experience with Grammarly and ChatGPT

Conclusion: AI as writing partner, not substitute, maintaining authenticity and authorship

AI Emergency Support for Vulnerable Girls

Vivek

Crisis Support in Rural India

Key Findings:

  • AI provided calm, non-judgmental 24/7 support
  • Cultural blind spots in legal advice and emergency numbers
  • Occasionally delivered confident but incorrect legal information
  • Cannot truly feel or discern genuine emergencies

Key Lessons:

  • Technology vs. human empathy
  • Cultural localization
  • Crisis intervention ethics

Methodology: Scenario testing with real-world crisis prompts

Conclusion: Hybrid model needed: AI as first response, connected to human networks

AI Customer Service and Sentiment Analysis

Jeffrey

Automated Review Processing with Prompt Engineering

Key Findings:

  • Successfully categorized reviews by sentiment and priority
  • Generated appropriate response templates
  • Cost-effective solution for small businesses
  • Required careful prompt engineering for professional tone

Key Lessons:

  • Prompt engineering importance
  • Business automation
  • Quality control needs

Methodology: Custom prompt templates with structured outputs

Conclusion: Effective for business efficiency but needs human oversight for quality

AI for Cooking and Recipe Generation

Colin (Callan)

Recipe Creation and Personalization

Key Findings:

  • Great for brainstorming and ingredient-based suggestions
  • Provides structured, easy-to-follow instructions
  • Inconsistent outputs across sessions
  • Cultural bias toward American/Western recipes

Key Lessons:

  • Cultural representation
  • Consistency issues
  • Art vs. algorithm

Methodology: Personal cooking experiments with AI-generated recipes

Conclusion: Useful cooking partner but requires human judgment and cultural awareness

AI's Impact on Student Cognitive Abilities

Oliver

Cognitive Effects and Educational Solutions

Key Findings:

  • 40% decline in self-reported cognitive abilities
  • 20% worse performance when over-relying on AI
  • Dopamine-like reward cycle similar to social media
  • Intelligent Tutoring Systems can mitigate negative effects

Key Lessons:

  • Cognitive dependency
  • Educational design
  • Learning vs. shortcuts

Methodology: Research synthesis and school-based observations

Conclusion: AI can enhance learning if designed to promote thinking rather than provide answers

AI for Resume Screening

Larry

Recruitment Automation and Bias Detection

Key Findings:

  • Improved efficiency in processing large volumes of resumes
  • Exhibited biases based on names and non-relevant factors
  • Difficulty assessing soft skills and leadership qualities
  • Risk of missing unconventional but strong candidates

Key Lessons:

  • Algorithmic bias
  • Soft skills assessment
  • Fairness in hiring

Methodology: Testing ChatGPT and Claude with real and AI-generated resumes

Conclusion: Useful for initial screening but requires human oversight to ensure fairness

AI for Learning Design and Curriculum Development

Esther

Custom GPTs for Educational Framework Creation

Key Findings:

  • Custom GPTs significantly outperformed standard ChatGPT
  • Produced properly scaffolded learning outcomes
  • Maintained required institutional structure and format
  • Still required subject matter expert validation

Key Lessons:

  • Custom vs. general models
  • Pedagogical structure
  • Expert validation

Methodology: Custom GPT development with pedagogical frameworks

Conclusion: Effective tool for learning designers when combined with expert knowledge

Cross-Cutting Themes

Prompt Engineering Criticality

Multiple presentations demonstrated how prompt design significantly affects AI output quality

Examples:

  • Michael's stock prediction iterations
  • Jeffrey's customer service templates
  • Esther's custom GPT instructions

Bias and Fairness Concerns

Students identified various forms of bias across different applications

Examples:

  • Larry's name-based hiring bias
  • Colin's cultural recipe bias
  • Vivek's Western-centric emergency advice

Overconfidence Problem

AI systems consistently showed high confidence in potentially incorrect outputs

Examples:

  • Michael's precise but wrong stock predictions
  • Ethan's math error acceptance
  • Vivek's confident legal misinformation

Human-AI Collaboration Models

Most effective implementations involved humans in expert or oversight roles

Examples:

  • Nancy's banking defense lines
  • Esther's learning design workflow
  • Oliver's educational safeguards
Methodological Insights

Comparative Analysis

Several students compared AI tools (Ethan: ChatGPT vs. DataWise, Larry: ChatGPT vs. Claude)

Value: Revealed performance differences between general and specialized models

Real-World Testing

Students used actual scenarios rather than hypothetical examples

Value: Provided authentic insights into practical AI limitations and capabilities

Multi-Criteria Evaluation

Presentations assessed accuracy, bias, usability, and ethical implications

Value: Demonstrated comprehensive approach to AI system evaluation

Iterative Prompt Design

Multiple students refined their prompts based on initial results

Value: Showed importance of prompt engineering in achieving desired outcomes
Professor's Key Reflections

On AI's Paradoxical Capabilities

"It seems like these chatbots can do pretty sophisticated mathematics, but they fail miserably on very simple things... Why is there some data, is there like a lot more data out there for the sophisticated problems than the simpler ones?"

On Creativity vs. Hallucination

"Where is the fine line between creativity and hallucination? Hallucination is something that does not exist, created by the machine, and creativity is, by definition, you want to create something that others have not thought about."

On the Value of Exploration

"AI technology encourages all of us to think more creatively about how to use it. We may end up saying the human system is doing the best. That actually is not a bad outcome. It's a confirmation."

On Human Uniqueness

"We're always looking for something unique, extraordinary. But I cannot tell you what is unique, because if I can tell you, that's not unique. How do you recognize the pattern, which is you don't have a pattern?"

Future Research Directions

Technical Improvements

Development of culturally-aware AI systems that understand local contexts
Better integration of AI with human expertise in professional workflows
Improved confidence calibration to reduce overconfident incorrect responses

Implementation & Ethics

Specialized AI models for domain-specific applications (education, healthcare, etc.)
Enhanced bias detection and mitigation techniques across applications
Standardized evaluation frameworks for AI system assessment in various domains
Course Transition: From Solutions to Problems

Looking Ahead

This session marked the midpoint of the course. The first half focused on "How can we use GAI to solve problems?" while identifying problems generated by GAI itself.

The second half will focus on "How do we solve problems posed by GAI?" - addressing data ethics, bias, transparency, and the complex challenges revealed through these student investigations.

Next Session Preview

Professor Andrew Lo from MIT will discuss "Can you really use ChatGPT for investment?" - continuing the practical exploration with expert academic perspective.