Lesson 3: Ethics and Artificial Intelligence¶

Lesson Overview

Duration: 50 minutes

Structure:

Introduction (5 min)
Core Concepts (25 min)
Discussion Activity (15 min)
Wrap-up (5 min)

Learning Objectives¶

After completing this lesson, you will be able to:

Explain the difference between "Ethics of AI" and "Ethical AI"
Identify sources of bias in AI systems
Apply bias detection and mitigation strategies
Use AI tools responsibly and ethically in research
Recognize transparency and accountability issues
Evaluate AI systems using ethical frameworks
Make informed decisions about AI integration in your work

Introduction (5 minutes)¶

The AI Revolution¶

Dartmouth AI Workshop 1956 — Dartmouth Summer Research Project on Artificial Intelligence, 1956. A new field of science had begun. (Credit: IEEE Spectrum, The Minsky Family)

In 1956, a small group of scientists gathered at Dartmouth for a Summer Research Project on Artificial Intelligence. They proposed:

"Every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it."

Nearly 70 years later, we are witnessing the practical realization of their vision. AI systems now:

Assist in scientific research and data analysis
Support medical diagnosis and drug discovery
Power search engines and recommendation systems
Generate text, images, code, and music
Make decisions affecting human lives

With this power comes profound responsibility.

AI in Environmental Health Research

DUST researchers are already using AI for:

Image Analysis - Automated quantification of lung tissue damage, fibrosis, and cellular changes in histopathology slides
Exposure Modeling - Machine learning to predict arsenic dispersal patterns from mine tailings across landscapes
Risk Assessment - AI models predicting health risks for communities near contaminated sites
Spectroscopy Analysis - Neural networks identifying arsenic species from X-ray absorption spectra
Plant Stress Detection - Computer vision analyzing hyperspectral imagery to identify metal-stressed vegetation
Literature Review - LLMs assisting with systematic reviews of toxicology research

Each application raises ethical questions about accuracy, bias, transparency, and appropriate use.

Two Critical Questions¶

Framing Our Discussion

Ethics of AI - What principles and regulations should govern AI development and deployment?
Ethical AI - How should AI systems behave to align with human values?

As researchers, we must consider both: the ethical frameworks surrounding AI and the ethical use of AI in our work.

Core Concepts (25 minutes)¶

Understanding AI Bias¶

Key Definitions

AI Bias occurs when an AI system produces systematically prejudiced or unfair results, typically due to biased training data or flawed assumptions during development.

Algorithmic Discrimination occurs when AI use results in unfair or illegal treatment of individuals or groups based on protected characteristics (age, disability, race, religion, sex, socioeconomic status).

Fairness includes equalized error rates across groups and parity of outcomes, though defining fairness remains contextual and contested.

Sources of AI Bias¶

AI systems inherit and often amplify human biases through multiple pathways:

1. Data Bias¶

The most common source of AI bias stems from flawed, unrepresentative training data:

Selection Bias

Training data not representative of entire population
Example: AI trained only on data from one demographic group

Measurement Bias

Data systematically differs from true values
Example: Using zip code as proxy for income, introducing geographic bias

Exclusion Bias

Certain groups omitted from data collection
Example: Medical AI trained predominantly on male patients

Labeling Bias

Subjective judgments during data annotation
Example: Different labelers having different standards for "appropriate" content

Historical Bias

Data reflects past discrimination
Example: Hiring AI trained on historical data that reflects gender imbalances

2. Algorithmic Bias¶

The algorithm itself may prioritize certain features inappropriately:

Optimization techniques favoring majority over minority groups
Feature selection that encodes protected characteristics
Loss functions that do not account for fairness constraints
Models optimizing for average performance, ignoring subgroup disparities

3. Human Decision Bias¶

Human biases influence every stage of AI development:

Confirmation Bias

Over-reliance on pre-existing beliefs when designing systems
Seeking data that confirms hypotheses

Stereotyping Bias

Perpetuating harmful generalizations
Example: Associating certain occupations with specific genders

Out-Group Homogeneity Bias

Treating underrepresented groups as more similar than they are
Failing to capture within-group diversity

Empathy Gap

Inability to incorporate nuanced human experiences
Reducing complex realities to quantifiable metrics

4. Synthetic Bias¶

When models trained on biased data generate synthetic datasets, they perpetuate and potentially amplify those biases into new systems.

The Amplification Problem

AI systems do not just reflect existing biases - they can amplify them. Small biases in training data can lead to large disparities in outcomes, especially when systems are deployed at scale or used in feedback loops.

Real-World Examples of AI Bias¶

Case Studies

Healthcare: Skin Cancer Detection

AI models trained predominantly on light-skinned patients performed poorly when diagnosing melanoma in darker skin tones, potentially delaying critical treatment.

Criminal Justice: COMPAS

Recidivism prediction software showed racial bias, incorrectly flagging Black defendants as higher risk at nearly twice the rate of white defendants.

Hiring: Amazon Recruiting Tool

AI trained on historical hiring data learned to penalize resumes containing the word "women's" (as in "women's chess club"), reflecting and reinforcing gender imbalances.

Natural Language: Word Embeddings

Language models learned associations like "man is to computer programmer as woman is to homemaker," encoding gender stereotypes into AI systems.

Facial Recognition: Misidentification

Systems showed significantly higher error rates for women and people of color, leading to false identifications and wrongful arrests.

Environmental Health AI Bias Scenarios

Exposure Risk Models Trained on Wealthy Communities

An AI model predicting health risks from mine tailings dust is trained primarily on data from well-resourced communities with extensive environmental monitoring. When applied to underserved communities (often disproportionately Indigenous or Latino populations near Arizona mine sites), the model may underestimate risks due to different housing types, occupational exposures, and baseline health disparities.

Lung Tissue Image Analysis Bias

A deep learning model for detecting fibrosis in lung tissue is trained primarily on tissue from one mouse strain. When applied to genetically diverse animal models (or human tissue samples), it may have reduced accuracy, potentially affecting conclusions about arsenic toxicity across populations.

Environmental Justice Data Gaps

AI models predicting contamination patterns rely on historical monitoring data. If sampling has been sparse in low-income or Indigenous communities (a documented environmental justice issue), AI models will be less accurate precisely where vulnerable populations live, perpetuating inequity.

Plant Species Recognition Bias

Computer vision for identifying phytoremediation-suitable plants is trained on images from temperate regions. When deployed in arid Southwest ecosystems, it may misclassify native species, leading researchers to overlook locally-adapted plants with superior remediation potential.

Bias Prevention and Mitigation Strategies¶

Data-Centric Approaches¶

Diverse and Representative Collection

Curate datasets that accurately represent all relevant groups
Oversample underrepresented groups
Document demographic composition clearly

Quality Assurance

Identify gaps and compatibility problems
Address historical underrepresentation
Validate across different subpopulations

Balancing Techniques

Under-sample majority groups
Over-sample minority groups
Generate synthetic data for underrepresented groups (carefully)

Rigorous Labeling

Consistent annotation guidelines
Multiple labelers to reduce individual bias
Masked labeling (hiding sensitive attributes)
Documentation of labeling decisions

Continuous Updates

Data refreshed throughout AI lifecycle
Monitor for dataset drift
Update as populations and contexts change

Algorithmic Techniques¶

Bias Detection Tools

Modern tools help identify bias:

Fairness Metrics

Multiple mathematical definitions of fairness exist:

Demographic Parity - Outcomes equal across groups
Equalized Odds - Error rates equal across groups
Counterfactual Fairness - Decisions unchanged if sensitive attribute changed
Individual Fairness - Similar individuals treated similarly

No Perfect Fairness

Different fairness metrics can conflict with each other. Optimizing for one may worsen another. Context determines which metric(s) matter most.

Algorithmic Adjustments

Pre-processing - Adjust training data (reweighting, resampling)
In-processing - Modify algorithm (fairness constraints, adversarial debiasing)
Post-processing - Adjust outputs (threshold optimization)

Explainable AI (XAI)

Understanding model decisions helps identify bias:

SHAP (SHapley Additive exPlanations)
LIME (Local Interpretable Model-agnostic Explanations)
Attention visualization for neural networks
Feature importance analysis

Using AI Ethically in Research¶

As researchers employing AI tools, we have ethical obligations:

1. Verification and Validation¶

Critical Rule

Never use AI for tasks where you cannot verify accuracy.

If you lack expertise to judge whether AI output is correct, you cannot use it responsibly.

Best Practices:

Verify AI-generated information against primary sources
Use AI as assistant, not replacement for expertise
Fact-check all claims, especially for high-stakes decisions
Maintain critical thinking - do not accept outputs uncritically

2. Transparency and Attribution¶

Disclose AI Use

In methods sections, describe AI tools used
Specify models, versions, prompts when relevant
Explain what AI did and what humans verified

Cite Appropriately

Follow journal guidelines for AI tool citation
Distinguish between AI-assisted and AI-generated content
Do not claim AI output as your original work

3. Privacy and Confidentiality¶

Data Privacy with AI

Do not input confidential, sensitive, or regulated data into AI systems unless:

You have explicit permission
The system complies with relevant regulations (HIPAA, FERPA, etc.)
Data are properly anonymized
You understand data retention policies

Sensitive Data Includes:

Human subjects data
Proprietary information
Personally identifiable information (PII)
Protected health information (PHI)
Student records
Unpublished research data from collaborators

DUST-Specific Privacy Concerns

Never input into public AI systems:

Precise GPS coordinates of contaminated sites (security and trespassing risk)
Unpublished arsenic concentration data from specific locations
Biomarker data that could identify participants in community health studies
Mine ownership or legal information related to ongoing remediation
Tribal community data without explicit permission
Pre-publication toxicology results that could affect property values
Patient-level lung disease data linked to exposure sources

Safe AI uses:

Asking general questions about phytoremediation approaches (no site specifics)
Getting help with R/Python code for statistical analysis (using synthetic example data)
Drafting methods sections based on published protocols
Summarizing published literature on arsenic toxicity mechanisms

4. Bias Awareness¶

When using AI in research:

Understand potential biases in AI tools you use
Consider whether AI recommendations might disadvantage certain groups
Validate AI performance across relevant subpopulations
Report limitations and potential biases in publications

5. Environmental Considerations¶

Large AI models have significant environmental costs:

Training GPT-3 emitted ~550 tons of CO2
Each ChatGPT query uses substantial energy
Consider environmental impact when choosing tools
Use smaller models when appropriate

Transparency and Accountability¶

Model Documentation¶

Responsible AI development requires thorough documentation:

Model Cards (Mitchell et al., 2019)

Intended use and users
Training data characteristics
Performance metrics across subgroups
Ethical considerations
Example: Google Model Cards

Datasheets for Datasets (Gebru et al., 2018)

Motivation for dataset creation
Composition and collection process
Preprocessing and labeling
Uses, distribution, and maintenance

Regulatory Landscape¶

AI governance is evolving rapidly:

International Frameworks

National Policies

US Executive Orders on AI (evolving with administrations)
UK AI Strategy and Regulatory Approach
China's AI Governance Framework

Institutional Guidelines

Many universities and research institutions now have:

AI use policies for researchers
Guidelines for student use of AI
Review boards for high-risk AI applications
Training requirements for AI users

Ethical Frameworks for AI¶

Several frameworks guide ethical AI development:

Asilomar AI Principles (2017)¶

23 principles covering:

Research issues (research culture, funding, science-policy link)
Ethics and values (safety, failure transparency, value alignment)
Longer-term issues (capability caution, importance of humans, AI arms race)

IEEE Ethically Aligned Design¶

Principles:

Human Rights
Well-being
Data Agency
Effectiveness
Transparency
Accountability
Awareness of Misuse
Competence

ACM Code of Ethics¶

Relevant principles:

Contribute to society and human well-being
Avoid harm
Be honest and trustworthy
Be fair and take action not to discriminate
Respect privacy
Honor confidentiality

Discussion Activity (15 minutes)¶

Scenario-Based Ethical Analysis¶

Work in small groups to analyze these scenarios:

Scenario 1: AI-Assisted Toxicology Literature Review¶

The Situation

You are writing an NIH Superfund proposal to study novel mechanisms of arsenic-induced lung disease. You use ChatGPT to:

Summarize recent literature on arsenic toxicity pathways
Suggest experimental approaches for measuring lung fibrosis
Draft sections on statistical analysis for dose-response studies
Generate preliminary hypotheses about gene-environment interactions

You plan to submit the proposal after lightly editing the AI-generated text, without carefully verifying all citations or confirming that suggested methods are appropriate for arsenic inhalation studies.

Discussion Questions:

What ethical issues does this raise? (Consider scientific accuracy, NIH integrity policies)
What could go wrong? (Wrong citations, inappropriate methods, missed critical safety considerations)
How should AI be used ethically in grant writing for toxicology research?
What disclosure is appropriate to NIH reviewers?
What if the AI suggested methods that work for one metalloid but not arsenic?

Scenario 2: Environmental Health Risk Prediction with Biased Data¶

The Situation

You are developing a machine learning model to predict respiratory disease risk for communities near abandoned mine sites. Your training data comes from:

Well-documented mine sites in affluent areas with extensive air quality monitoring
Predominantly from Western mine sites (few data from Southwest dryland ecosystems)
Communities with good healthcare access and documented health outcomes

When you test the model on data from underserved communities near Arizona mine tailings (often Indigenous or Latino populations with limited monitoring data), the model systematically underestimates risks. These are precisely the environmental justice communities that would most benefit from accurate predictions.

Discussion Questions:

What types of bias are present in this training data? (Selection, exclusion, measurement)
What are the potential harms if this system is deployed to guide remediation priorities?
What data-centric approaches could help? (Active sampling in underrepresented areas?)
What algorithmic interventions might work? (Fairness constraints? Domain adaptation?)
Should the system be deployed? Under what conditions? Who should decide?
How do you balance "perfect data" with "timely action" for environmental justice?

Scenario 3: AI-Assisted Phytoremediation Discovery¶

The Situation

You use an AI system to analyze thousands of papers on plant metal uptake, arsenic biochemistry, and dryland ecology. The AI identifies a novel connection: a drought-stress pathway in native Arizona plants may enhance arsenic sequestration in roots, suggesting specific plant species for phytoremediation that hadn't been considered.

You design field experiments based on this AI-generated hypothesis. The experiments are successful, showing that AI-suggested native species outperform traditional hyperaccumulators in arid conditions. You prepare a high-impact publication on "AI-discovered phytoremediation approach."

Discussion Questions:

How should you attribute the AI's contribution? (Co-author? Tool acknowledgment?)
What should you disclose in the methods section? (AI model, version, prompts, literature corpus?)
If AI identified the hypothesis, who "owns" the intellectual contribution?
How does this differ from using PubMed search or a literature database?
What if the AI missed critical papers on plant toxicity, leading you to suggest a harmful approach?
Do you have ethical obligations to the Indigenous knowledge that may underlie some plant use patterns the AI detected?

Scenario 4: Open-Source Contamination Mapping Model¶

The Situation

You develop a machine learning model that predicts arsenic contamination patterns around mine tailings sites using satellite imagery, geology, and meteorological data. The model is highly accurate and could benefit remediation planning nationwide.

However, the same model could be used by:

Property developers to identify and conceal contaminated land
Bad actors to target vulnerable contaminated sites for illegal activities
Those seeking to avoid regulatory oversight by identifying unmonitored sites
Real estate interests to devalue Indigenous or low-income lands

You plan to publish the model open-source with all training data and code to maximize scientific benefit and environmental justice applications.

Discussion Questions:

What are your ethical obligations to both openness and safety?
Should you restrict access to the model? How? (Gated access? Institutional use only?)
How do you balance open science principles with potential for harm?
What documentation, warnings, or terms of use should accompany release?
Should you consult with affected communities before releasing?
Does publishing contamination predictions violate privacy of residents near sites?

Group Reporting¶

Each group shares:

Key ethical issues identified
Proposed solutions or guidelines
Remaining uncertainties or disagreements
Broader principles that emerged

Wrap-up (5 minutes)¶

Key Takeaways¶

Remember These Concepts

Bias is pervasive - AI systems inherit human biases from data, algorithms, and decisions
Multiple approaches needed - Address bias at data, algorithm, and deployment stages
Verification is essential - Never use AI where you cannot validate outputs
Transparency matters - Disclose AI use, document systems, understand limitations
Context is everything - Ethical AI use depends on application, stakes, and alternatives
Ongoing learning - AI ethics is an evolving field requiring continuous education

Ethical AI Checklist¶

Use this checklist when integrating AI into research:

Before Using AI

Do I have expertise to verify AI outputs?
Am I using appropriate tools for this task?
Do I understand potential biases in this AI system?
Have I checked institutional policies on AI use?
Is my use consistent with AI's intended purpose?

When Using AI

Am I critically evaluating all outputs?
Have I verified facts and citations?
Am I protecting confidential/sensitive data?
Am I documenting what AI does vs. what I do?
Would I be comfortable explaining this use publicly?

After Using AI

Have I disclosed AI use appropriately?
Have I cited AI tools correctly?
Have I evaluated AI impact on outcomes?
Have I considered potential biases introduced?
Have I documented limitations?

Self-Assessment Quiz¶

What is the difference between 'Ethics of AI' and 'Ethical AI'?

Ethics of AI refers to principles and regulations governing AI development and deployment - the frameworks, laws, and guidelines surrounding AI.

Ethical AI focuses on how AI systems behave and whether that behavior aligns with human values - the conduct and impacts of AI systems themselves.

Both are important: we need good governance (Ethics of AI) and well-designed systems (Ethical AI).

True or False: Selection bias occurs when the algorithm prioritizes certain features over others

False

Selection bias is a type of data bias that occurs when training data are not representative of the whole population. For example, training a medical AI predominantly on data from one demographic group.

Algorithmic bias refers to issues with how the algorithm itself processes information, such as optimization techniques that favor majority groups.

Which is an example of 'Explainable AI (XAI)' in bias mitigation?

Understanding which inputs drive model decisions, revealing hidden biases

XAI techniques like SHAP, LIME, and attention visualization help make AI decision-making transparent. This transparency is crucial for identifying biases - for example, discovering that a hiring AI inappropriately weighs gender or that a medical AI relies on spurious correlations rather than causal factors.

True or False: It is acceptable to use AI to write a research proposal in a field where you have no expertise

False

Using AI where you cannot verify accuracy is unethical. If you lack expertise to judge whether AI-generated information is correct, truthful, or appropriate, you cannot use it responsibly in research contexts.

AI should assist, not replace, domain expertise. Always use AI only in areas where you can critically evaluate its outputs.

Looking Forward¶

Ethics in AI and research is not a one-time lesson but an ongoing practice:

Continue Learning:

Follow AI ethics research and discussions
Engage with institutional review boards
Participate in ethics training offered by your institution
Stay current on regulatory developments

Take Action:

Implement ethical practices in your work
Advocate for responsible AI in your field
Share knowledge with colleagues
Contribute to ethical guidelines in your discipline

Critical Engagement:

Question AI systems and their design
Consider who benefits and who might be harmed
Think about long-term societal implications
Balance innovation with responsibility

Completing the Training¶

Congratulations on completing all three lessons! You now have foundational knowledge in:

Open science principles and practices
Research data management and FAIR principles
Ethical considerations in AI

Next Steps:

Review Additional Resources for deeper learning
Apply concepts to your current research projects
Share knowledge with your research group
Revisit lessons as you encounter new situations
Provide feedback to improve this training

Additional Resources¶

Frameworks and Guidelines:

Tools and Platforms:

Further Reading:

Weapons of Math Destruction by Cathy O'Neil
Artificial Unintelligence by Meredith Broussard
Atlas of AI by Kate Crawford

Courses:

Previous: ← Lesson 2: Data Management | Home: Training Home →

Last updated: 2025-10-14