AI Ethics, Data Foundations, and Visualization in Engineering

Ethics in Engineering & AI

Engineers must consider the ethical implications of AI systems, focusing on people, safety, fairness, and accountability.

Predictive maintenance

Using AI to predict when machines will fail, with ethical concerns about overreliance on imperfect predictions.

Cybersecurity

AI's role in defending or attacking systems, with potential misuse or bias endangering user data.

Water quality / environmental monitoring

Concerns about sensor failures or algorithm misreporting that could harm people.

Traffic optimization

AI must balance efficiency and fairness in determining priority.

Job displacement

Automation may replace workers, necessitating consideration of social impacts.

Stakeholders

All parties affected by AI systems, including workers, the public, and regulators.

Duration of concern

The balance between short-term benefits and long-term consequences, such as data privacy erosion.

Predictability

The challenge of unforeseen outcomes requiring caution and transparency in AI systems.

Prioritization

The process of ranking what matters most in tradeoffs between efficiency and fairness.

Knowledge vs. Data

Data are facts or measurements, while knowledge is meaning extracted from data through inference and interpretation.

Quantitative data

Data derived from sensors, surveys, budgets, and experiments.

Qualitative data

Data obtained from interviews, reports, comments, and policy documents.

Data trustworthiness

Questions to assess data validity, honesty, completeness, noise, and bias.

Data Cleaning & Repair

Methods to handle unreliable data, including managing outliers, interpolating missing values, and smoothing noisy data.

Data representation formats

Different ways to represent data, including matrices, graphs, and tables (DataFrames).

Example of knowledge extraction

Data: sensor readings from a bridge; Knowledge: 'Bridge stress is increasing; failure is likely in 3 years.'

Self-driving car data bias

A self-driving car trained mostly on data from sunny cities will perform poorly in snow, highlighting that honest data can still be biased.

Engineers' tradeoffs

Engineers must weigh the tradeoffs between availability, quality, and cost of data.

Nominal

Labels, identifiers

Ordinal

Ordered categories

Interval

Ordered with consistent spacing, no true zero

Ratio

Interval + meaningful zero

Data Science

About understanding data — extracting patterns, meaning, and knowledge.

Artificial Intelligence

About acting intelligently — making decisions or predictions from knowledge.

Data Preparation

Essential for ethical and high-performing AI.

Cleaning

Remove errors, duplicates, outliers

Labeling

Assign correct outcomes or categories

Augmentation

Generate synthetic data to reduce bias or balance classes

Feature Engineering

Combine or transform variables to uncover relationships AI wouldn't see on its own.

Data Science EDiCtS Cycle

Adapted from the AI E.D.I.T. mindset (Explore, Direct, Imagine, Test).

Exploratory Data Analysis (EDA)

Goal: understand the shape and structure of your data.

Univariate Analysis

Look at one variable at a time.

Multivariate Analysis

Examine relationships between multiple variables.

Unsupervised Methods

Used when there are no labels.

Clustering

Groups similar data points (e.g., K-Means)

Dimensionality reduction

Simplifies datasets (e.g., PCA)

Association rules

Find 'if-then' relationships (e.g., market basket analysis)

Data Augmentation

Artificially expand data using transformations (rotation, scaling, synthesis).

Lists

Ordered data

Matrices

2D arrays (NumPy)

DataFrames

Labeled tables (pandas)

File I/O

Reading/writing data files

AI

Acts on a model of the world made up of states and actions.

Perception

senses what's happening (camera, sensor, input)

Knowledge

stores what's known about the world

Reasoning

computes what to do next

Action

performs behavior to change the world

State

external description of the world at a moment

Example of State

A robot's state includes position and orientation; its knowledge includes a map of the environment.

Abstraction

means throwing out information — but doing it thoughtfully.

Example of Abstraction

A self-driving car ignores tree colors but tracks road lines and nearby objects.

Good State Characteristics

Should completely describe everything necessary for the AI to make a decision.

Unambiguous State

multiple world configurations shouldn't map to the same state (if possible).

Compact State

avoid redundancy.

Example Errors in State

Missing data (sensor error → incomplete state), Redundant attributes (wastes computation), Overly complex representation (slows search)

Chess State Representation

All piece locations + turn info {White_King:E1, Black_Queen:D8, Turn:White}

Robot Pathing State Representation

X, Y, and direction (15, 22, North)

Warehouse Robot State Representation

Position, empty/full, goal location (15,22,Empty=True,Goal=(107,62))

Traffic Management State Representation

Time, event status, vehicle count, avg speed (14:43, Event=False, 578, 27.2)

AI as Search

AI problems are often formulated as search problems: Most of AI is state search or traversal.

Search Examples

Pathfinding: robot moving from start to goal, Planning: sequence of actions (manufacturing schedule, game moves), Optimization: minimizing cost or risk, Machine learning: searching over parameters or functions

Search Trade-offs

More Knowledge vs. Less Knowledge, Faster computation vs. Slower computation, Expensive to build vs. Cheaper, easier to build, More assumptions vs. More flexible

Search as Planning

To formalize a problem as a search: Define what you're optimizing (time, energy, cost, risk), Define your state representation, Define your cost function (how you compare options), Define what to keep in memory (visited states, paths), Define how quickly you must solve it (real-time vs. offline)

Ethics

the study of what is the right thing to do (rules, duties, consequences, fairness).

Moral Epistemology

how we know it's right — what evidence, reasoning, or procedures justify our claim.

Engineering mindset

The engineering process (systematic, iterative problem solving) transfers directly to ethical analysis.

Define problem and stakeholders

Identify who is affected by a problem and how (users, bystanders, regulators).

Model trade-offs

Evaluate competing factors such as safety vs. cost and accuracy vs. privacy.

Iterate & test

Validate assumptions and designs through repeated testing and refinement.

Concrete example of predictive maintenance model

Identify workers (safety), owners (cost), and regulators (compliance) while weighing false negatives vs. false positives.

Training data bias

Occurs when the sample isn't representative, such as a face dataset with few darker-skinned individuals leading to poor performance.

Objective / proxy bias

Happens when the optimization target is a proxy that correlates with protected attributes, like using past spending as a proxy for creditworthiness.

Non-technical root causes of bias

Include design choices, labeling decisions, deployment context, and feedback loops.

EV battery predictor example

A 2% low prediction may seem small, but systematic underestimation can lead to customer distrust or incorrect range estimates affecting travel safety.

Tank identification urban legend

A dataset contained pictures of tanks on different backgrounds; the model learned the background instead of the tanks.

COMPAS / recidivism risk

An example where algorithmic predictions affected sentencing, with proxies correlated with race/socioeconomics leading to unfair outcomes.

Testing for bias

Measure performance across subgroups (precision/recall per group) and use counterfactual tests to change only the protected attribute.

Bias severity

A bias can be harmless in isolation but harmful in context, affecting certain groups disproportionately or triggering cascading decisions.

Privacy risks in Big Data

Seemingly innocuous telemetry can reconstruct location traces, leading to privacy risks.

Digital divide

Data-driven tools can advantage those who can pay for data/compute, disadvantaging smaller entities and creating systemic inequity.

Professional vs. software ethics

Professional engineers are bound by legal codes and duty to public safety, while software engineers often lack universal legally binding licenses.

Legal & contractual limits

Most EULAs disclaim warranties and limit damages, shifting risk to users while not absolving engineers morally.

E.D.I.T. framework

A process for responsible AI development: Explore, Direct, and actionable steps to ensure ethical practices.

Explore in E.D.I.T.

Learn the tool including dataset, architecture, and failure modes; run probes for edge cases.

Actionable in E.D.I.T.

Generate adversarial or out-of-distribution inputs; look for hallucinations, biases, and missing coverage.

Direct in E.D.I.T.

Constrain/model behavior through prompt engineering, model guardrails, post-processing filters, and rule-based checks.

Direct

Constrain/model behavior: prompt engineering, model guardrails, post-processing filters, rule-based checks.

Actionable

Add fairness constraints, thresholding, human-in-the-loop points.

Imagine

Ask: what should ideal output look like? Consider stakeholders and possible misuse.

Test

Rigorous testing: metrics across subgroups, stress tests, monitoring in production. Record procedures and results.

Exam tip

The EDIT acronym is an easy 4-step answer for a short-response question on responsible AI.

Define problem & scope

What system, what decisions.

List stakeholders

Primary, secondary.

Enumerate harms & benefits

Safety, privacy, fairness, economic.

Identify evidence

Data sources, validation steps, metrics to measure harms.

Mitigation plan

Data augmentation, fairness constraints, human oversight, opt-outs.

Testing plan

Per-group metrics, counterfactual tests, simulated deployment.

Ethics vs moral epistemology

Definition & example.

E.D.I.T. steps

Explore, Direct, Imagine, Test.