8 Technology & Society IIb

Technology and Society - Intermediate

Part 2 - Artificial Intelligence and Large Language Models

No technology in human history has been deployed at scale as quickly as AI and large language models (LLMs). Within roughly two years of ChatGPT’s public release in late 2022, AI tools became embedded in:

Search engines and web browsers
Email and productivity software
Educational platforms and homework help
Customer service and business operations
Healthcare diagnostics and treatment planning
Hiring and employee evaluation systems
Creative tools for writing, art, music, and video
Legal research and contract analysis
Financial advising and fraud detection
Government services and decision-making

This creates an unprecedented problem: Billions of people are using, being evaluated by, or having decisions made about them by AI systems they don’t understand. The gap between deployment speed and public understanding is wider than for any previous major technology.

This matters because AI systems are making consequential decisions about who gets hired, who receives medical treatment, who gets approved for loans, who is flagged as suspicious by law enforcement, and what information people see. These decisions affect individual lives and shape society—yet most people using or affected by these systems don’t know how they work, what their limitations are, or what biases they might contain.

The goal of this section isn’t to make you an AI expert—it’s to give you enough understanding to:

Recognize what AI can and cannot do reliably
Use AI tools effectively when you choose to use them
Identify situations where AI use is inappropriate or risky
Evaluate claims made by companies selling AI products
Make informed decisions about AI in your personal, professional, and civic life
Understand the ethical implications of AI development and deployment
Participate meaningfully in conversations about AI policy and regulation

What you’ll learn:

First, how AI and LLMs actually work—not the deep mathematics, but enough to understand what these systems are doing and why they behave the way they do.

Second, what they can and cannot do—their genuine capabilities and fundamental limitations, cutting through both hype and dismissiveness.

Third, the ethical considerations—the environmental, social, and economic costs of AI that aren’t visible in the interface, so you can make informed decisions about whether and how to use these tools.

Fourth, how to use AI effectively when you decide it’s appropriate—practical guidance based on understanding how these systems work.

Fifth, common pitfalls—the mistakes people are making right now that damage their education, careers, and relationships, and how to avoid them.

Finally, how to evaluate AI claims critically—recognizing hype, understanding what questions to ask, and making informed judgments about AI products and policies.

A note on terminology: “AI” is a broad category including many different technologies—machine learning, neural networks, expert systems, robotics, and more. This section focuses primarily on large language models (LLMs)—the technology behind ChatGPT, Claude, Gemini, and similar tools—because these are what most people encounter directly. Many of the principles apply to other AI systems as well, but LLMs are the most immediately relevant for most readers.

How AI/LLMs Actually Work

Large language models don’t think, understand, or know anything. They are extremely sophisticated pattern-matching and prediction systems. Understanding this fundamental fact explains both their impressive capabilities and their significant limitations.

The basic mechanism: predicting the next token

At their core, LLMs do one thing: predict what comes next in a sequence of text.

A “token” is roughly a word or part of a word. When you type a prompt, the model analyzes the patterns in your text and generates a response by repeatedly predicting “given everything so far, what token is most likely to come next?”

It’s similar to your phone’s autocomplete feature, but vastly more sophisticated. Your phone might predict the next word based on the previous few words. An LLM predicts based on thousands of tokens of context and patterns learned from billions of examples.

How they learn: training on massive text datasets

Before an LLM can predict anything, it must be trained on enormous amounts of text—books, articles, websites, code, conversations, and more. During training, the model:

Reads billions of examples of human-written text
Learns statistical patterns: which words tend to follow which other words, what sentence structures are common, how different topics are typically discussed
Adjusts billions of internal parameters (called “weights”) to better predict the next token in the training examples
Gets better at prediction through repeated exposure and adjustment

Think of it like this: Imagine reading millions of mystery novels. You’d start recognizing patterns—how clues are planted, how red herrings work, when the detective typically solves the case. If someone asked you to write a mystery novel, you could produce something that looks like a mystery novel by following these patterns—even without understanding what makes a good mystery or caring whether your clues actually make logical sense.

That’s essentially what LLMs do with language. They’ve “read” so much text that they can produce text that follows the patterns of human writing remarkably well. But following patterns isn’t the same as understanding.

Why they seem intelligent but aren’t

LLMs can produce text that sounds knowledgeable, thoughtful, and coherent. This creates a powerful illusion of intelligence and understanding. But:

They have no concept of truth or accuracy. They generate what’s statistically likely based on their training data, not what’s actually true. If their training data contains misinformation (and all large datasets do), they’ll confidently reproduce it. They can’t fact-check themselves or verify information.

They have no understanding of meaning. When an LLM writes “the sky is blue,” it’s not accessing any concept of sky or blue or color. It’s predicting that these words frequently appear together in this pattern. It would generate “the sky is blue” and “the sky is green” with equal confidence if the patterns in its training data supported both.

They have no goals, intentions, or beliefs. They’re not trying to help you, deceive you, or accomplish anything. They’re predicting tokens based on patterns. Any appearance of personality, helpfulness, or intention is a reflection of patterns in their training data (including human feedback during training), not actual mental states.

They have no memory between conversations. Unless specifically designed with external memory systems, LLMs don’t remember previous conversations. Each interaction starts fresh. (Some implementations save conversation history and feed it back as context, but the model itself has no persistent memory.)

Key limitations that follow from how they work

Understanding the mechanism explains why LLMs have specific, predictable limitations:

Hallucinations: LLMs sometimes generate information that sounds plausible but is completely fabricated. This happens because they’re predicting likely text patterns, not retrieving facts. If the pattern suggests a citation should appear, it will generate one—even if that citation doesn’t exist. If asked about a person’s biography, it will generate plausible-sounding career details based on patterns from similar biographies, regardless of whether those details are true.

This isn’t a bug to be fixed—it’s fundamental to how these systems work. They can’t distinguish between “patterns that happen to be true” and “patterns that happen to be false.” Both are just patterns.

No real-time information: LLMs are trained on data up to a specific cutoff date. They know nothing about events after that date unless specifically provided with that information. They can’t browse the internet (unless given that capability as an external tool), access databases, or know what’s happening right now.

Inconsistency: Ask the same question multiple times and you’ll often get different answers—sometimes contradictory. This is because prediction involves some randomness (by design, to make output less repetitive). The model isn’t consulting a consistent internal knowledge base; it’s generating likely continuations, which can vary.

Inability to reason reliably: While LLMs can sometimes produce text that looks like logical reasoning, they’re not actually reasoning—they’re pattern-matching on what reasoning looks like in their training data. Give them novel problems that don’t match common patterns, and they often fail. They’re particularly bad at:

Multi-step mathematical reasoning (though they can handle simple math that appears frequently in training data)
Novel logical puzzles
Anything requiring genuine understanding rather than pattern recognition
Recognizing when they’re being asked something impossible or nonsensical

Bias reproduction: LLMs learn patterns from their training data, including biases. If the training data contains gender stereotypes, racial biases, or cultural assumptions (and all large datasets do), the model will reproduce these patterns. The model can’t evaluate whether a pattern is fair or accurate—only whether it’s common.

Lack of common sense: Humans have embodied experience of the physical world—we know that objects fall, water is wet, and you can’t be in two places at once. LLMs have only seen text about these things. They can generate text about common sense without having it, leading to occasional absurd errors that no human would make.

No self-awareness or uncertainty calibration: LLMs can’t actually evaluate how confident they should be. They can generate text that sounds confident or uncertain based on patterns, but this doesn’t correlate with actual accuracy. They’ll state completely fabricated information with the same confident tone as well-established facts.

What they’re actually good at

Despite these limitations, LLMs are genuinely useful for tasks that align with their actual capabilities:

Pattern-based text generation: Writing that follows common patterns (emails, summaries, first drafts, formatting)
Text transformation: Rewriting text in different styles, translating between languages (with caveats), summarizing documents
Brainstorming and exploration: Generating ideas, alternative perspectives, or questions to consider (with human evaluation)
Code assistance: Suggesting code based on common programming patterns (again, requires human review)
Explanation of common concepts: Explaining well-documented topics in accessible language (though verification is essential)
Structured information extraction: Pulling specific information from text and organizing it

The key is matching the task to what the system actually does—pattern recognition and generation—rather than treating it as a knowledge base, reasoning engine, or intelligent assistant.

Why this matters for users

Understanding how LLMs work helps you:

Use them effectively by asking for tasks they can actually do well
Verify outputs appropriately by knowing where errors are likely
Avoid inappropriate uses where accuracy and reasoning matter
Evaluate vendor claims about what AI can do
Understand privacy implications (your prompts become training data in many systems)
Recognize AI-generated content by spotting characteristic patterns

You don’t need to understand the mathematics or computer science details. But understanding “sophisticated pattern matching, not thinking” is essential for using these tools wisely and evaluating their role in society.

Capabilities and Limitations

Understanding the mechanism behind LLMs helps predict what they’ll be good or bad at. This section gives you practical guidance for evaluating whether an LLM is appropriate for a specific task.

Tasks LLMs handle well

These are tasks where pattern recognition and generation align with the actual need:

Writing assistance and editing:

First drafts and outlines: LLMs excel at generating starting points based on common writing patterns. They can produce a basic structure you then refine with your actual knowledge and judgment.
Rewriting for clarity or tone: Given existing text, they can suggest alternative phrasings, simplify complex sentences, or adjust formality level.
Grammar and style checking: They recognize common grammatical patterns and can catch errors (though not perfectly).
Format conversion: Turning bullet points into paragraphs, reorganizing information, creating tables from text.

Limitation: LLMs don’t know if the content is accurate, appropriate, or achieves your actual goals. They generate plausible-sounding text, not necessarily good text. You must provide the judgment, expertise, and verification.

Information synthesis and summarization:

Summarizing documents: Condensing long text into key points, though they may miss nuance or include hallucinated details not in the original.
Explaining common concepts: For well-documented topics with lots of training data (basic science, mainstream history, popular technology), they can generate accessible explanations.
Comparing options: They can list common differences between similar things (programming languages, historical events, etc.) based on how these comparisons appear in training data.

Limitation: Summaries may omit critical details or introduce errors. Explanations are only as good as the patterns in training data—if the training data contains misconceptions, the output will too. Always verify with authoritative sources for important information.

Brainstorming and exploration:

Generating ideas: Lists of possibilities, alternative approaches, questions to consider, creative prompts.
Exploring perspectives: “What might someone argue from X perspective?” or “What are counterarguments to Y?”
Question formulation: Helping articulate what you’re trying to figure out.

Limitation: Ideas are based on common patterns, not necessarily good or original ideas. The “perspectives” are stereotypical versions based on training data, not genuine diverse viewpoints. Treat output as starting points for your own thinking, not answers.

Code assistance:

Boilerplate and common patterns: Generating standard code structures, common functions, repetitive tasks.
Syntax suggestions: Helping remember syntax for languages you know but don’t use daily.
Explaining code: Describing what a code snippet does (with verification).
Debugging help: Suggesting common causes of errors based on error messages.

Limitation: Generated code often contains subtle bugs, security vulnerabilities, or inefficient approaches. Code may work for common cases but fail on edge cases. Never use AI-generated code without understanding it and testing it thoroughly. This is especially critical for security-sensitive or safety-critical applications.

Language translation:

Common languages with lots of training data: Translations between major languages (English, Spanish, French, Chinese, etc.) often capture basic meaning.
Informal communication: Translating casual messages where perfect accuracy isn’t critical.

Limitation: Translations can miss nuance, cultural context, idioms, and tone. Less common languages or specialized vocabulary have worse performance. Use professional human translators for important documents, legal texts, medical information, or creative works.

Text-based data processing:

Extracting structured information: Pulling names, dates, or specific facts from unstructured text and organizing into tables or lists.
Categorizing content: Sorting text into predefined categories based on patterns.
Sentiment analysis: Identifying whether text is generally positive, negative, or neutral (with caveats about accuracy).

Limitation: Extraction accuracy varies based on text clarity and how well the task matches training patterns. May miss edge cases or misclassify ambiguous content. Verify accuracy on sample data before trusting at scale.

Tasks LLMs handle poorly or dangerously

These are tasks where pattern matching fails to meet actual requirements:

Factual research and information retrieval:

LLMs are not search engines or encyclopedias. They don’t retrieve information—they generate plausible text based on patterns. For any factual question:

They may generate completely fabricated “facts” that sound authoritative
They may confidently state outdated information (from before their training cutoff)
They may confuse similar topics or combine details from different sources
They can’t tell you which facts are well-established vs. controversial vs. false

If you need accurate information, use actual research tools: Search engines, academic databases, authoritative sources, expert consultation. You can use an LLM to help phrase your research questions or organize information you’ve already verified, but never trust an LLM as a primary source.

Medical, legal, or financial advice:

LLMs can generate text that sounds like professional advice. This is extremely dangerous:

Medical advice requires understanding individual circumstances, current research, and potential interactions—pattern matching on general medical discussions can kill people
Legal advice requires knowing current law in specific jurisdictions and how it applies to specific situations—getting this wrong can result in lost rights, financial harm, or criminal penalties
Financial advice requires understanding your specific situation, current markets, and regulatory requirements—bad advice can destroy your financial security

Always consult qualified professionals for medical, legal, or financial decisions. LLMs can help you formulate questions to ask your doctor or lawyer, but they cannot replace professional judgment.

Mathematical reasoning and calculation:

While LLMs can handle arithmetic that appears frequently in training data (2+2=4), they struggle with:

Multi-step calculations
Problems requiring actual mathematical reasoning
Novel mathematical puzzles
Anything beyond pattern recognition of common problems

They may generate confident-looking mathematical work that contains fundamental errors. Use calculators, spreadsheets, or mathematical software for actual calculations. Use human mathematicians for genuine mathematical reasoning.

Logical reasoning and novel problem-solving:

LLMs can reproduce reasoning patterns they’ve seen in training data, but they can’t actually reason. Give them:

Novel logical puzzles
Multi-step reasoning where each step depends on the previous one
Problems requiring genuine understanding rather than pattern matching

They often fail or produce plausible-sounding but logically invalid results. Don’t trust LLMs for decisions requiring genuine logical reasoning.

Real-time or recent information:

LLMs don’t know:

Current events (after their training cutoff date)
Current prices, availability, or business hours
What’s trending or happening now
Recent research or discoveries

Some implementations add web search capability, but the LLM itself has no access to current information. Check publication dates and verify time-sensitive information.

Detecting AI-generated content:

Paradoxically, LLMs are poor at detecting whether text was AI-generated. They can recognize some patterns of AI-generated text, but so can humans—and both produce many false positives and false negatives. Claims about “AI detection software” are largely unreliable. Don’t trust AI detectors for high-stakes decisions.

Evaluating truth, ethics, or quality:

LLMs can generate text about ethics or quality, but they can’t actually evaluate these things:

They can’t tell you if something is ethical—only what ethical arguments commonly appear in their training data
They can’t evaluate whether creative work is good—only whether it matches common patterns
They can’t judge source credibility—they’ll treat conspiracy theories and peer-reviewed research the same if both appear in training data

These judgments require human values, expertise, and critical thinking.

Tasks requiring accountability:

LLMs can’t be held responsible for errors or harm. If you use AI output for important decisions and it goes wrong:

You can’t sue the AI
The AI developers typically disclaim liability
You bear the responsibility

Never use AI for decisions where you’re not prepared to take full responsibility for errors.

Common misconceptions to avoid

Misconception: “AI is always improving, so limitations are temporary”

Reality: Some limitations come from the fundamental approach (pattern matching vs. reasoning) and won’t be solved by more data or bigger models. Other limitations may improve, but new ones often emerge. Evaluate tools based on current capabilities, not promises about the future.

Misconception: “If it sounds confident, it’s probably right”

Reality: Confidence in AI output has no correlation with accuracy. LLMs generate confident-sounding text about fabrications and facts with equal certainty.

Misconception: “Expensive/premium AI is more accurate”

Reality: Paid versions often have better performance, longer context windows, or additional features, but they still hallucinate, still have the same fundamental limitations, and still require verification. Price doesn’t eliminate the need for human judgment.

Misconception: “AI can learn and improve from our conversation”

Reality: In most implementations, the model doesn’t learn from individual conversations. It generates responses based on its fixed training. Some systems save conversation history to provide context, but this isn’t learning—the underlying model remains unchanged.

Misconception: “AI understands what I mean, even if I’m unclear”

Reality: LLMs respond to patterns in your text. Unclear prompts lead to responses based on what similar unclear prompts looked like in training data, not what you actually meant. Clarity in, clarity out.

How to evaluate appropriateness for a task

When considering whether to use an LLM for a specific task, ask:

Does this require accuracy that I can verify? If you can check the output against reliable sources, an LLM might help. If you can’t verify, don’t use it.
Is pattern-based generation acceptable for this purpose? For creative brainstorming or first drafts, patterns work. For factual accuracy or original reasoning, they don’t.
What’s the cost of error? For casual emails, errors are minor annoyances. For medical decisions, legal documents, academic work, or professional communications, errors can cause serious harm.
Am I treating AI as a tool or a substitute for expertise? Tools extend capability you already have. Substitutes replace capability you lack—and that’s where danger lies.
Would I be comfortable taking full responsibility for this output? If not, don’t use it or don’t publish it.

Understanding these capabilities and limitations helps you use AI tools effectively when appropriate—and avoid them when they’re not.

Ethics of AI Use

Before learning how to use AI effectively, you should understand the costs that aren’t visible in the interface. Using an AI tool feels free and effortless—you type a question, get an answer, and move on. But behind that simple interaction are significant environmental, social, and economic costs that users rarely see or consider.

This section isn’t telling you not to use AI. It’s giving you information to make informed decisions about whether and how to use it, just as you might consider the environmental impact of driving vs. public transit, or the labor practices of companies you buy from. Understanding the full cost of a technology is essential for informed choice.

Environmental costs: Energy, water, and carbon

AI systems consume extraordinary amounts of resources:

Training costs: Creating a large language model requires running thousands of processors for weeks or months. Training GPT-3 (an earlier, smaller model than current ones) was estimated to consume 1,287 megawatt-hours of electricity and produce 552 tons of CO2 emissions—equivalent to driving an average car for over 1.2 million miles. Larger, more recent models require significantly more.

Inference costs: Every time you use an AI system (called “inference”), it consumes energy. A single ChatGPT query uses roughly 10 times more electricity than a Google search. Multiply this by billions of queries daily, and the energy consumption becomes massive.

Water consumption: Data centers require enormous amounts of water for cooling. Training GPT-3 consumed an estimated 700,000 liters of clean freshwater. Microsoft reported that AI contributed to a 34% increase in their water consumption in 2022 compared to 2021. In regions facing water scarcity, this creates serious ethical concerns.

Carbon footprint: The electricity for AI systems often comes from fossil fuels. Even when companies purchase renewable energy credits, the actual power consumed at peak times often comes from the grid, which includes coal and natural gas.

Scale matters: As AI becomes embedded in more products and services—search engines, email, productivity tools, social media—the aggregate environmental impact grows exponentially. Some estimates suggest that AI could account for 3.5% of global electricity consumption by 2030.

Why this matters for your choices:

Using AI for trivial tasks (having it write a grocery list, generate jokes, create content you don’t really need) wastes resources at scale
Using AI when simpler tools work (calculator instead of asking AI to do math, search engine instead of asking AI for factual information) reduces unnecessary consumption
Some tasks justify the resource cost (accessibility tools, medical applications, research); others don’t
Asking “is this worth the environmental cost?” is a reasonable ethical consideration

Labor exploitation in training and moderation

The “artificial” in artificial intelligence obscures how much human labor goes into these systems:

Data labeling: Training AI requires humans to label millions of examples—tagging images, rating text quality, identifying toxic content. This work is often outsourced to workers in developing countries who are paid extremely low wages (sometimes less than $2/hour) and given minimal benefits or job security.

Reinforcement Learning from Human Feedback (RLHF): Making AI “helpful” and “harmless” requires humans to rate outputs, write preferred responses, and evaluate model behavior. This improves the product but depends on precarious labor.

Content moderation: To prevent AI from generating extremely harmful content, human moderators must review and label toxic material—violence, sexual abuse, hate speech, and worse. This work causes documented psychological trauma. Workers are exposed to disturbing content for hours daily, often with inadequate mental health support, and may develop PTSD, depression, and anxiety.

Investigative reporting has revealed:

Workers in Kenya paid $2/hour to label toxic content for OpenAI
Moderators developing psychological trauma from exposure to child abuse, violence, and other harmful content
Lack of mental health support or adequate breaks
Workers afraid to speak out due to non-disclosure agreements and job insecurity

Why this matters for your choices:

Using AI systems supports business models that may rely on exploitative labor practices
Companies vary in their treatment of workers—some pay fair wages and provide support, others don’t
This is difficult information to access as a consumer, but worth investigating when choosing which AI services to use
Advocating for better labor practices in AI development is part of ethical technology use

Training data: Consent and copyright issues

LLMs are trained on massive datasets scraped from the internet—books, articles, websites, code repositories, social media, and more. Most of this data was used without explicit permission from creators.

Copyright concerns:

Authors, artists, and journalists find their work included in training data without consent or compensation
AI can reproduce styles, techniques, or even near-copies of copyrighted work
Current copyright law doesn’t clearly address whether training on copyrighted material constitutes infringement
Creators whose work trains AI systems that then compete with them economically receive no benefit

Consent concerns:

Personal information, photos, social media posts, and private communications scraped from the internet may be in training data
People didn’t consent to their words being used to train commercial AI systems
Some training data includes sensitive information—medical records, private emails, financial information—obtained from data breaches or improperly secured databases

Cultural and intellectual appropriation:

Training data includes Indigenous knowledge, traditional practices, and cultural expressions—often used without community permission
Minority and marginalized communities’ contributions are used to train systems that may then encode biases against those same communities

Ongoing legal and ethical debates:

Multiple lawsuits are currently challenging AI companies’ use of copyrighted material. Legal frameworks are evolving. The ethics remain contested:

Some argue that training on publicly available data is transformative use, similar to how humans learn from reading
Others argue that commercial AI systems should require licensing agreements and compensate creators
Some propose that non-commercial, research, or accessibility uses have different ethical status than commercial applications

Why this matters for your choices:

Using AI systems may indirectly benefit from unauthorized use of others’ creative work
Some creators ask people not to use their work with AI; respecting these wishes requires conscious choice
Alternative AI systems trained on openly licensed or consented data exist (though with trade-offs in capability)
How you use AI (reproducing someone’s style vs. general-purpose writing assistance) affects the ethical weight

Bias and discrimination amplification

As discussed in “How AI/LLMs Actually Work,” these systems learn patterns from training data—including biased patterns. This creates serious ethical concerns when AI is used in consequential decisions:

Employment: AI screening tools have been shown to:

Discriminate against women (Amazon scrapped a hiring tool that penalized résumés containing the word “women’s”)
Disadvantage older applicants
Penalize employment gaps that disproportionately affect caregivers (mostly women)
Favor candidates whose résumés match privileged patterns

Criminal justice: AI risk assessment tools used in bail, sentencing, and parole decisions have been shown to:

Assign higher risk scores to Black defendants than white defendants with similar criminal histories
Predict recidivism inaccurately, with racial disparities in error rates
Perpetuate historical bias in policing and prosecution

Healthcare: AI diagnostic and treatment recommendation systems may:

Perform worse for underrepresented groups in training data
Encode existing healthcare disparities
Miss conditions more common in populations with less representation in medical research

Financial services: AI credit scoring and loan approval systems can:

Discriminate based on proxies for protected characteristics (like zip code as a proxy for race)
Deny opportunities to people in marginalized communities
Perpetuate historical redlining and financial discrimination

The compounding problem: When biased AI systems make decisions, those decisions create new data that reflects the bias. If an AI hiring tool discriminates against women, fewer women get hired, creating a “feedback loop” where the lack of women in the company reinforces the AI’s patterns. This can entrench and amplify existing inequalities.

Why this matters for your choices:

Using AI for consequential decisions (hiring, admissions, lending, risk assessment) can perpetuate discrimination—often invisibly
The appearance of objectivity (“the algorithm decided”) can obscure bias and make it harder to challenge
Different applications have different risk levels—using AI to draft an email is different from using it to screen job applicants
Understanding bias helps you evaluate whether AI is appropriate for a specific use case

Creativity and attribution questions

When AI generates creative work, thorny questions arise about authorship, originality, and credit:

Who owns AI-generated content?

If you use AI to write a story, compose music, or create art, who is the author? You, the AI company, the creators whose work trained the model, or no one?
Current law is unclear and varies by jurisdiction
Some jurisdictions require human authorship for copyright protection
Terms of service for AI platforms vary—read them carefully

Is AI-generated work original?

AI generates outputs based on patterns in training data
When AI produces something similar to a specific creator’s work, is that plagiarism, inspiration, or transformation?
If you use AI to generate creative work and publish it as your own, are you misrepresenting the authorship?

Attribution and transparency:

Should AI-generated content be labeled as such?
In academic work, journalism, or professional settings, are you obligated to disclose AI use?
How does using AI to generate content affect your credibility or the value of your work?

Impact on creative professions:

If AI can produce “good enough” writing, art, or music cheaply, what happens to human creators?
Does widespread AI use devalue creative labor and expertise?
Are we building a future where human creativity is economically unviable?

Why this matters for your choices:

How you use and credit AI-generated content has ethical implications
Different contexts have different norms (casual use vs. professional vs. academic)
Being transparent about AI use respects audiences and maintains trust
Your choices about when to use human creators vs. AI affects the viability of creative professions

Making informed ethical choices

These ethical concerns don’t necessarily mean “never use AI.” They mean consider the full cost and context:

Questions to ask yourself:

Is this use necessary? Does it solve a real problem, or is it using resources for convenience or novelty?
Are there less costly alternatives? Could a simpler tool, human labor, or different approach work as well?
What are the stakes? Low-stakes personal use (drafting an email) has different ethical weight than high-stakes decisions (screening job applicants).
Am I being transparent? Are you disclosing AI use where appropriate and expected?
Who bears the costs? Environmental costs are distributed globally. Labor costs fall on vulnerable workers. Bias costs fall on marginalized groups. Creative costs fall on artists and writers. Are you comfortable with this distribution for this use?
What systems am I supporting? Your use (and payment, if applicable) supports specific companies and practices. Are those aligned with your values?

You don’t need perfect ethical purity—that’s impossible in complex systems. But informed choice means understanding these costs and making deliberate decisions rather than using AI thoughtlessly because it’s convenient.

The next sections on using AI effectively and avoiding pitfalls assume you’ve considered these ethical questions and decided AI is appropriate for your specific use case.

Using AI Effectively

If you’ve decided AI is appropriate for your task, understanding how these systems work helps you use them effectively. This section provides practical guidance based on the reality of what LLMs actually do: pattern matching and prediction.

Treat AI like a capable but unproven collaborator

A useful mental model: Treat an LLM like you would a new team member whose capabilities you don’t yet know.

When working with someone new, you don’t immediately trust them with critical tasks. You:

Start with small, low-stakes tasks
Check their work carefully
Notice what they’re good and bad at
Give them bigger responsibilities as they prove themselves
Maintain oversight on important work
Never hand off tasks where you can’t verify the results

Apply the same approach to AI:

Start with simple requests where you can easily verify accuracy
Read every word of output before using it—don’t just accept it because it looks plausible
Pay attention to patterns: where does it do well, where does it struggle?
Scale up to more complex tasks only after building understanding of its capabilities
Always maintain final judgment and responsibility
Never use AI for tasks where you can’t evaluate whether the output is good

This approach helps you discover what the tool can actually do for you rather than relying on marketing claims or assumptions.

Effective prompting: Being clear and specific

LLMs respond to patterns in your text. Clarity in, clarity out.

Vague prompt: “Write something about climate change” Why it’s ineffective: The model has no constraints, so it generates generic text based on common climate change discussion patterns. You’ll get something that could be a blog post, an essay, a social media rant, or a children’s explanation—and it probably won’t match what you actually wanted.

Specific prompt: “Write a 200-word explanation of how carbon dioxide traps heat in the atmosphere, appropriate for a high school student who understands basic chemistry” Why it’s better: You’ve specified length, topic focus, audience, and assumed knowledge level. The output will better match your needs.

Elements of effective prompts:

State the task clearly: What do you want the AI to do? Write, summarize, rewrite, explain, list, compare?

Provide context: What’s this for? Who’s the audience? What do they already know?

Specify constraints: Length, format, tone, style, what to include or avoid.

Give examples if helpful: “Like this, but for a different topic” or “In the style of X, but about Y”

Break complex tasks into steps: Instead of “write a complete article,” try “first create an outline, then expand each section”

Example comparison:

Weak: “Help me with my email” Strong: “I need to decline a meeting invitation from a colleague. The tone should be polite and professional. Keep it under 100 words. I have a scheduling conflict, not a lack of interest.”

Weak: “Explain photosynthesis” Strong: “Explain photosynthesis to someone who knows basic biology but has never studied it in detail. Focus on the inputs, outputs, and energy transformation. 3-4 paragraphs, avoiding technical jargon.”

Iterative refinement: Working with the output

Rarely will a first attempt produce exactly what you need. Effective AI use is iterative:

Start with a clear prompt
Evaluate the output: What’s good? What’s missing? What’s wrong?
Refine the prompt based on what you learned: “That’s good, but make it more concise” or “You missed X, include that”
Repeat until you have something useful

This is similar to working with a human collaborator—you give feedback, they adjust, you refine together. The difference is that the AI doesn’t actually learn from your conversation (in most implementations), so each new task starts fresh.

Verification: Always check the output

Never trust AI output without verification. The confidence of the text tells you nothing about its accuracy.

For factual claims:

Check against reliable sources
Look up citations (they may be fabricated)
Verify statistics, dates, names, and technical details
Be especially skeptical of confident-sounding claims about obscure topics

For reasoning and logic:

Work through the steps yourself
Check that conclusions follow from premises
Look for logical fallacies or gaps
Test edge cases or counterexamples

For code:

Understand what the code does before running it
Test thoroughly with various inputs
Check for security vulnerabilities
Review for efficiency and best practices

For creative or professional work:

Ensure it matches your actual intent, not just a plausible interpretation
Check tone, style, and appropriateness for audience
Verify it represents your actual views and knowledge
Ensure it doesn’t plagiarize or too closely mimic specific sources

When you can’t verify output, don’t use the output. If you don’t know enough about a topic to check whether the AI’s explanation is accurate, you’re not in a position to use that explanation.

Understanding context windows and limitations

LLMs have a maximum “context window”—how much text they can consider at once (your prompt plus conversation history plus their response). When you exceed this:

Earlier parts of the conversation get cut off
The model “forgets” things you mentioned earlier
Responses may become inconsistent with earlier statements

Practical implications:

Don’t expect the model to remember details from very long conversations
If working on a long document, you may need to break it into sections
Important information should be restated if it’s far back in the conversation
Starting a new conversation can sometimes give better results than continuing a very long one

Different tasks require different approaches

Match your approach to what the task actually requires:

For brainstorming and exploration:

Ask for quantity over quality: “Give me 20 possible approaches”
Request diverse perspectives: “What would someone argue from X viewpoint?”
Use it to expand your thinking, not replace it
Evaluate all suggestions critically—many will be generic or inappropriate

For writing assistance:

Start with outlines, expand incrementally
Use AI for first drafts, but expect to revise substantially
Ask for alternative phrasings when stuck
Provide your own knowledge and expertise—the AI supplies structure and language, you supply content and judgment

For learning and explanation:

Request explanations at your level: “Explain like I’m familiar with X but not Y”
Ask follow-up questions to probe deeper
Verify understanding against authoritative sources
Use it to formulate questions for further research, not as a final answer

For technical tasks (code, data analysis):

Be extremely specific about requirements
Test all generated code thoroughly
Understand the code before using it
Treat suggestions as starting points, not complete solutions

For summarization:

Provide clear guidelines about what to focus on
Compare summary to original to check for omissions or errors
Recognize that summaries reflect the AI’s pattern recognition, which may miss what you consider important
Write your own summary first, then use AI to refine, rather than accepting AI’s judgment about what matters

What not to do: Common misuses

Don’t outsource judgment: AI can generate options, but you must choose. It can suggest approaches, but you must evaluate appropriateness. It can draft text, but you must ensure accuracy and quality.

Don’t treat it as authoritative: Even when it sounds confident, it’s pattern-matching. Confidence is not evidence.

Don’t use it for high-stakes decisions without verification: Medical, legal, financial, academic, professional—anything with serious consequences requires human expertise and accountability.

Don’t assume it understands your intent: It responds to patterns in your text, not your actual meaning. Miscommunication is common.

Don’t trust it with sensitive information: Your prompts may be stored, used for training, or accessed by company employees. Don’t share:

Passwords or credentials
Personal identifying information about yourself or others
Proprietary business information
Confidential data
Private medical or financial details

Don’t use it to impersonate or deceive: Generating content that falsely represents authorship, expertise, or identity is unethical and often against terms of service.

Choosing appropriate use cases

AI is a tool in your toolbox, not a replacement for skill and knowledge. It’s most effective when:

You have expertise to evaluate output: You know enough about the domain to recognize errors, gaps, and quality issues.

The task is pattern-based: Writing in common formats, organizing information, suggesting standard approaches, generating variations.

You can verify results: You have access to authoritative sources, testing methods, or expert review.

Stakes are low to moderate: Errors are inconvenient rather than catastrophic, and you can catch and correct them.

It saves meaningful time or effort: The efficiency gain justifies the environmental and ethical costs.

AI is least effective (or actively harmful) when:

You lack domain expertise: You can’t tell good output from plausible-sounding nonsense.

The task requires genuine reasoning: Novel problems, complex logic, original thinking.

You can’t verify results: No reliable sources, no testing method, no expert available.

Stakes are high: Medical decisions, legal consequences, academic integrity, professional reputation, safety-critical applications.

It’s unnecessary: Simpler tools work as well with fewer costs.

Practical tips for better results

Be conversational but specific: You don’t need to use formal language, but you do need to be clear about what you want.

Correct misunderstandings explicitly: If the AI misinterprets your request, say so: “No, I meant X, not Y. Try again with that understanding.”

Ask for explanations: “Explain your reasoning” or “Why did you suggest that approach?” can help you evaluate whether the logic makes sense.

Request alternatives: “Give me three different ways to approach this” helps you see options rather than accepting the first suggestion.

Set boundaries: “Don’t include X” or “Avoid Y style” helps constrain output to what you actually want.

Use it to augment, not replace, your work: Let AI handle tedious parts (formatting, restructuring, generating variations) while you provide the expertise, judgment, and creativity.

Learning what works for you

Effective AI use is a skill developed through practice and reflection:

Experiment with different prompting styles
Notice what produces useful vs. useless output
Build a sense of where this tool helps you vs. where it wastes time
Develop your own workflow that leverages AI’s strengths while avoiding its weaknesses

Different people will find different use cases valuable. Someone who writes frequently might use AI differently than someone who codes, teaches, or manages projects. There’s no single “right way”—only approaches that work or don’t work for your specific needs.

The key is maintaining agency and judgment. You’re using a tool to extend your capability, not substituting the tool for capability you lack.

Common Pitfalls

People are making predictable mistakes with AI right now—mistakes that damage their education, careers, relationships, and reputations. These aren’t hypothetical risks; they’re happening daily. Understanding these pitfalls helps you avoid them.

Academic misconduct and learning failure

The pitfall: Students use AI to complete assignments, submit AI-generated work as their own, or rely on AI instead of developing their own skills.

Why it happens: AI makes it easy to generate plausible-sounding essays, problem solutions, and project work. The temptation to save time or avoid difficult learning is strong.

The consequences:

Immediate: Academic penalties ranging from failing grades to expulsion. Many institutions now use AI detection tools (imperfect as they are) and sophisticated plagiarism detection. Professors recognize AI-generated patterns.
Long-term: You don’t learn the material. The skills you’re supposed to develop—critical thinking, writing, problem-solving, research—don’t develop. You advance without competence, and this catches up eventually.
Credential devaluation: If widespread AI use in academics becomes known, the value of degrees from institutions that don’t enforce integrity policies may decrease.

Real examples:

Students submitting essays with hallucinated citations that professors immediately recognize as fake
Code assignments that include AI-generated comments saying “I can’t complete this task” or containing errors that reveal lack of understanding
Essays in a writing style inconsistent with the student’s previous work, raising immediate red flags
Group projects where one member’s contribution is obviously AI-generated, damaging the entire team’s credibility

How to avoid it:

Understand your institution’s AI policy: Many schools have specific rules about AI use. Know them.
Use AI for learning, not substitution: AI can help explain concepts you’re struggling with, suggest study approaches, or provide practice problems. It shouldn’t do your assignments.
Treat AI like tutoring or collaboration: Most institutions allow you to get help understanding material but require your own work on assignments. Apply the same standard to AI.
Verify and understand everything: If you use AI to help understand a concept, make sure you genuinely understand it and can explain it yourself.
Cite appropriately: If your institution allows AI use and you use it, cite it according to their guidelines—just as you would any other source.

The actual goal of education is learning, not credentials. Cheating with AI gets you a degree without competence, which is ultimately worthless.

Professional and career disasters

The pitfall: Using AI for professional work without adequate oversight, verification, or disclosure.

Why it happens: AI promises efficiency and polish. In professional contexts with tight deadlines, the temptation to let AI handle work is strong.

The consequences:

Job applications: Applicants using AI to write cover letters and résumés often produce:

Generic, cookie-cutter applications that hiring managers immediately recognize and reject
Statements that don’t match the applicant’s actual experience or voice
Fabricated skills, experiences, or achievements (because the AI filled in patterns)
Answers to application questions that miss specific requirements because AI generated plausible-sounding but irrelevant responses

Hiring managers are learning to spot AI-generated applications. Some now actively filter them out. Using AI poorly can make you look worse than not using it at all.

Legal work: Lawyers have submitted legal briefs containing AI-generated case citations that don’t exist. Courts have sanctioned lawyers for this. Hallucinated legal citations can get you disbarred.

Medical contexts: Healthcare providers using AI-generated clinical notes or diagnostic suggestions without verification risk misdiagnosis, inappropriate treatment, and malpractice liability. AI medical errors can kill people.

Business communications: AI-generated emails, proposals, or reports that contain factual errors, inappropriate tone, or misunderstanding of context can:

Lose clients or contracts
Damage professional relationships
Expose confidential information (if you pasted it into an AI prompt)
Reveal incompetence when the error is discovered

Journalism and content creation: Publishing AI-generated content without verification has led to:

Factually incorrect articles damaging publication credibility
Plagiarism of other sources’ work
Fabricated quotes and events
Byline fraud (claiming human authorship of AI work)

Real examples:

Lawyers fined thousands of dollars and facing disciplinary action for submitting briefs with fake case citations
Job applicants rejected because their cover letters were obviously AI-generated templates
Companies publishing AI-generated articles with embarrassing factual errors that went viral
Consultants delivering AI-generated reports containing hallucinated data that clients recognized as false

How to avoid it:

Never submit AI-generated work without thorough review and verification
Verify every factual claim, citation, statistic, and technical detail
Ensure the tone, content, and style genuinely represent you or your organization
Understand what you’re submitting—if you can’t explain it or defend it, don’t put your name on it
Follow professional standards and disclosure requirements in your field
Maintain accountability—you’re responsible for everything you submit, regardless of how it was created

Your reputation is built over years and can be destroyed in moments. AI-assisted work requires the same standards as any other work bearing your name.

Over-reliance and skill atrophy

The pitfall: Depending on AI so much that your own skills deteriorate or never develop.

Why it happens: AI provides easy answers. Over time, reaching for AI instead of thinking through problems yourself becomes habitual.

The consequences:

Cognitive skills decline: If you always use AI to write, your writing skills don’t improve—they may actually worsen. Same for problem-solving, research, analysis, and other skills.
Learned helplessness: You become unable to function when AI isn’t available or when the task requires genuine expertise.
Reduced creativity and original thinking: Relying on AI’s pattern-based suggestions reduces your capacity for novel ideas and approaches.
Professional vulnerability: If your value depends on skills you’ve outsourced to AI, you’re easily replaceable—by AI or by people who maintain their skills.

This mirrors other technology dependencies: People who rely entirely on GPS never develop navigation skills and are helpless when technology fails. People who use calculators for all math often can’t estimate or check whether answers make sense.

How to avoid it:

Use AI selectively, not automatically: Ask “Do I need AI for this, or can I do it myself?” Choose to do things yourself regularly to maintain skills.
Practice skills outside AI use: Write without AI assistance. Solve problems yourself. Research the old-fashioned way. Keep skills active.
Increase difficulty over time: As you get better at something, reduce AI assistance rather than maintaining or increasing it.
Recognize AI as training wheels, not a permanent crutch: It can help you learn faster, but the goal is independent competence.

Skills are assets. Outsourcing them entirely makes you dependent and vulnerable.

Privacy and security failures

The pitfall: Sharing sensitive, confidential, or personal information in AI prompts without understanding how that data is used and stored.

Why it happens: The conversational interface feels private, like talking to a person. People forget they’re interacting with a company’s servers.

The consequences:

Data leaks: Information you share may be stored in company databases, potentially exposed in data breaches
Training data contamination: Your prompts may be used to train future AI models, meaning your private information could appear in responses to other users
Corporate access: Company employees may have access to your conversations for quality control or other purposes
Legal and compliance violations: Sharing regulated information (medical records, financial data, confidential business information) through AI may violate laws or contracts
Competitive disadvantage: Business strategies, product ideas, or proprietary information shared with AI may become accessible to competitors

Real examples:

Employees sharing proprietary code or business plans in AI prompts, potentially exposing trade secrets
Healthcare workers inputting patient information to get AI help, violating HIPAA regulations
Lawyers sharing confidential client information to draft documents, violating attorney-client privilege
Samsung employees sharing internal code in ChatGPT, leading the company to ban AI use after the data exposure risk became clear

How to avoid it:

Assume everything you share may become public: Don’t share anything you wouldn’t want leaked
Never include: Passwords, API keys, personal identifying information, confidential business data, private medical/financial information, others’ private information
Check terms of service and privacy policies: Understand how your data is used. Some services explicitly say they won’t use conversations for training; others do.
Use appropriate tools for sensitive work: Some AI services offer business/enterprise plans with stronger privacy guarantees. Or don’t use AI at all for sensitive work.
Anonymize when possible: Remove identifying details before sharing information with AI

Once information is shared, you’ve lost control of it. Prevention is the only protection.

Accepting bias and misinformation

The pitfall: Taking AI-generated content at face value without recognizing embedded biases, stereotypes, or false information.

Why it happens: Confident, articulate text creates an illusion of authority and accuracy. People don’t instinctively question well-written content.

The consequences:

Perpetuating stereotypes: Using AI-generated descriptions, stories, or examples that encode gender, racial, or cultural stereotypes
Making biased decisions: Relying on AI for hiring, evaluation, or other decisions that disadvantage certain groups
Spreading misinformation: Sharing AI-generated “facts” that are actually hallucinations or misconceptions from training data
Reinforcing your own biases: AI often reflects and confirms existing viewpoints, creating an echo chamber effect

Real examples:

AI-generated job descriptions using gendered language that discourages qualified applicants
Image generation tools producing racially biased outputs (asking for “a professional” consistently generating images of white men)
AI-generated historical or scientific “facts” that are confidently wrong
Biased sentiment analysis marking minority dialects or cultural expressions as negative

How to avoid it:

Actively look for bias: Ask yourself “Who is represented here? Who is excluded? What assumptions are embedded?”
Seek diverse sources: Don’t rely on AI alone for information about people, cultures, or complex social issues
Question stereotypical outputs: If AI generates content that seems to rely on stereotypes, regenerate or rewrite
Verify factual claims: Especially for historical, scientific, or statistical information
Get human perspectives: For anything involving people’s experiences, values, or identities, seek actual human voices from those communities

AI reflects training data, which reflects society’s biases. You must actively counteract this, not passively accept it.

Misunderstanding AI capabilities (“AI said so”)

The pitfall: Treating AI outputs as authoritative expert opinions rather than pattern-based predictions.

Why it happens: The sophistication of AI responses creates false confidence. “AI said X” feels like “an expert confirmed X.”

The consequences:

Bad decisions based on faulty reasoning: Following AI’s logical-sounding but actually flawed advice
Abandoned critical thinking: Accepting conclusions without evaluation because “AI figured it out”
Misplaced confidence: Believing you understand something because AI explained it, without genuinely learning
Dangerous applications: Using AI for medical self-diagnosis, legal advice, financial decisions, or other high-stakes situations

Real examples:

People following AI-generated medical advice that contradicts their doctor’s recommendations
Investors making financial decisions based on AI-generated market analysis that’s fundamentally flawed
Students submitting AI-generated solutions to technical problems without understanding them, then being unable to explain their own work
People ending relationships or making major life decisions based on AI “advice” that’s actually pattern-matched generic responses

How to avoid it:

Remember what AI actually is: Pattern matching, not expertise
Maintain skepticism: “AI suggested this” is a starting point for evaluation, not a conclusion
Seek actual experts for important decisions: Medical, legal, financial, safety-critical
Test AI reasoning: Work through the logic yourself. Does it actually make sense?
Recognize your own expertise: If something contradicts your knowledge or experience, trust yourself and investigate further

“AI said so” is never sufficient justification for important decisions.

Ethical shortcuts and rationalization

The pitfall: Using AI in ways you know are questionable while rationalizing that “everyone does it” or “it’s not really cheating.”

Why it happens: Competitive pressure, time pressure, and widespread AI use create social permission for ethical compromises.

The consequences:

Personal integrity erosion: Small compromises become habits; standards gradually decline
Collective degradation: If everyone takes shortcuts, standards collapse (credential inflation, professional competence decline)
Getting caught: As AI detection improves and human awareness increases, past shortcuts may be exposed
Internal conflict: Knowing you didn’t really earn achievements undermines self-respect and confidence

Common rationalizations:

“Everyone else is doing it” (Doesn’t make it right or beneficial to you)
“I’m just using it as a tool” (Tools extend capability you have; substitutes replace capability you lack)
“I’ll learn it eventually” (Shortcuts now prevent learning that should happen now)
“It’s faster this way” (Efficiency at the cost of competence is false economy)
“No one will know” (You’ll know, and that matters)

How to avoid it:

Clarify your actual values: What kind of person do you want to be? What standards matter to you?
Distinguish tools from substitutes: Honest question: Is this extending your capability or replacing it?
Consider long-term consequences: Will you be glad you did this in five years?
Seek legitimate help: Tutoring, collaboration, mentorship—ways to learn, not skip learning
Accept appropriate struggle: Difficulty is often where learning happens

Integrity isn’t about never making mistakes—it’s about honest effort and genuine development.

Summary: Avoiding pitfalls

Most AI pitfalls come from:

Treating AI as more capable than it actually is
Outsourcing judgment and responsibility
Ignoring verification requirements
Prioritizing efficiency over competence
Underestimating risks and consequences

Avoiding them requires:

Understanding what AI actually does
Maintaining human judgment and oversight
Verifying everything important
Building and maintaining your own skills
Following ethical and professional standards
Taking full responsibility for all output bearing your name

AI is a powerful tool. Like any powerful tool, it can extend capability or cause harm, depending on how it’s used. These pitfalls are predictable and avoidable.

Critical Evaluation: Recognizing Hype vs. Reality

AI is surrounded by extraordinary hype—companies claiming revolutionary capabilities, media predicting utopia or apocalypse, vendors promising to “transform” every industry. Learning to evaluate AI claims critically is essential for informed decision-making as a consumer, professional, and citizen.

Understanding “AI washing”

AI washing is the practice of labeling products or services as “AI-powered” to capitalize on hype, even when AI plays a minimal role or adds little value. It’s similar to “greenwashing” (claiming environmental benefits that don’t exist).

Common AI washing tactics:

Calling traditional automation “AI”: Rule-based systems, databases, or simple algorithms are rebranded as “AI” to sound modern and sophisticated
Vague claims without specifics: “Our AI analyzes your data” without explaining what it actually does or how
Exaggerated capabilities: Claiming the AI “understands,” “thinks,” or “learns” when it’s doing pattern matching
AI as buzzword: Adding “AI-powered” to existing products without meaningful improvement

Why it matters: AI washing wastes your money, creates false expectations, and makes it harder to identify genuinely useful AI applications. Companies benefit from the hype regardless of actual value delivered.

How to recognize it:

Ask “what specifically does the AI do?” If the answer is vague or could be done with simpler technology, it’s likely washing
Look for concrete capability descriptions rather than buzzwords
Check if the “AI” feature costs extra without clear additional value
Compare to non-AI alternatives: Does the AI version actually work better?

Questions to ask about AI products and services

When evaluating any AI product, service, or feature, ask:

1. What problem does this solve, and is AI necessary to solve it?

Sometimes AI is genuinely the best approach. Often, simpler solutions work as well or better. A product marketed as “AI-powered scheduling” might just be a calendar with basic automation—no AI needed.

2. What specifically does the AI do?

Demand concrete descriptions: “The AI analyzes X data to predict Y” or “The AI generates Z based on patterns in A.” Vague claims like “AI optimizes your workflow” tell you nothing.

3. How accurate is it, and how is accuracy measured?

“High accuracy” is meaningless without specifics. Ask:

What percentage accuracy? On what test dataset?
What types of errors occur, and how often?
Does accuracy vary across different populations or use cases?
Who validated these claims independently?

4. What data does it collect, and how is it used?

AI systems require data—yours and others’. Understand:

What data is collected from you?
Is it stored, shared, or sold?
Is it used to train models that benefit the company?
Can you opt out while still using the service?

5. What are the failure modes?

Every system fails sometimes. Critical questions:

What happens when the AI makes a mistake?
How will you know when it’s wrong?
What backup or override options exist?
What’s the cost of errors in your use case?

6. Who built it, and what are their incentives?

Consider:

Does the company have AI expertise, or are they jumping on a trend?
What’s their business model? (Selling AI products, selling your data, subscription revenue?)
Do they have track record of responsible technology development?
Are they transparent about limitations?

7. What does it cost—financially, in privacy, in time?

Consider total cost:

Direct financial cost (purchase, subscription, per-use fees)
Privacy cost (data collection, surveillance)
Time cost (learning to use it, correcting errors, managing the system)
Switching cost (if you later want to stop using it)

8. What are the alternatives, and how do they compare?

AI is one option among many:

What non-AI approaches exist?
What are their trade-offs?
Is there a “good enough” solution that’s simpler or cheaper?
What are you giving up by choosing the AI version?

Red flags in AI marketing and claims

Certain claims should trigger immediate skepticism:

“Revolutionary” or “unprecedented”: Most AI applications are incremental improvements, not revolutions. Extraordinary claims require extraordinary evidence.

“Human-level intelligence” or “understands like humans”: Current AI doesn’t understand anything. This is either hype or deception.

“Eliminates bias” or “completely objective”: AI systems trained on human data inherit human biases. Claims of perfect objectivity are false.

“Learns and improves on its own”: Most deployed AI systems don’t continuously learn from use. They’re static models. If they do learn, ask what they’re learning from and whether that’s desirable.

“Works for everyone” without qualification: AI performance typically varies across populations. Claims of universal effectiveness without acknowledging limitations are suspect.

“Trust the AI” or “let the AI decide”: Responsible AI use maintains human oversight and accountability. Pressure to blindly trust AI is a red flag.

“Proprietary AI that we can’t explain”: While some technical details are legitimately proprietary, companies should be able to explain generally how their system works and what its limitations are. Opacity is often a cover for poor performance or problematic methods.

No mention of limitations or errors: Every AI system has limitations. Companies that acknowledge them honestly are more trustworthy than those claiming perfection.

Evaluating AI in high-stakes contexts

When AI is used for consequential decisions—hiring, lending, medical diagnosis, criminal justice, educational assessment—heightened scrutiny is essential:

Transparency requirements:

How does the system make decisions? Not every technical detail, but the general approach and factors considered
What data was it trained on? Who’s represented, who’s excluded?
Has it been audited for bias? By whom? What did they find?
What recourse exists if it makes an error? Can decisions be appealed? To humans?

Accountability requirements:

Who is responsible for errors? The AI vendor, the organization deploying it, specific individuals?
What liability do they accept? Most AI vendors disclaim liability in terms of service—who bears the risk?
How are errors detected and corrected? What monitoring exists?

Fairness requirements:

Does it perform equally well across different demographic groups? If not, is this disclosed?
What impact does it have on marginalized communities? Has this been studied?
Are there less discriminatory alternatives? Why was AI chosen over them?

Many jurisdictions are developing AI regulation requiring some of these disclosures for high-stakes applications. As a citizen, you can advocate for strong requirements. As a consumer or person affected by AI decisions, you can demand answers.

Evaluating AI policy and regulation proposals

As AI becomes ubiquitous, governments are proposing various regulatory approaches. Critical evaluation helps you participate in democratic decision-making:

Questions for any AI policy proposal:

1. What problem is it solving?

Is there a demonstrated harm that needs addressing? Or is this regulation responding to hypothetical future risks while ignoring present harms?

2. Does it address root causes or just symptoms?

For example, if AI hiring tools discriminate, is the solution better AI, or is it addressing the underlying discrimination in society that the AI learned from?

3. Who benefits and who’s burdened?

Does regulation protect vulnerable people, or does it primarily benefit companies?
Does it impose requirements that only large companies can meet, creating monopolies?
Does it enable accountability, or does it provide legal cover for harmful practices?

4. What trade-offs does it involve?

Innovation vs. safety
Transparency vs. proprietary interests
Privacy vs. functionality
Individual rights vs. collective benefits

There are no perfect answers, but honest acknowledgment of trade-offs is essential.

5. Is it enforceable?

Does the policy include:

Clear standards that can be objectively assessed?
Consequences for violations?
Resources for enforcement?
Mechanisms for affected people to seek remedy?

6. Does it keep pace with technology?

AI develops rapidly. Does the policy:

Apply to existing harms, not just theoretical future ones?
Have mechanisms to adapt as technology changes?
Avoid being obsolete before implementation?

7. Does it include democratic input?

Were affected communities consulted?
Is there public transparency in decision-making?
Can citizens provide input and feedback?

Common regulatory approaches and their trade-offs:

Industry self-regulation: Companies develop voluntary standards and ethics guidelines

Potential benefit: Flexible, informed by technical expertise
Potential problem: No enforcement, conflicts of interest, standards that serve industry rather than public

Sector-specific regulation: Different rules for AI in healthcare, finance, criminal justice, etc.

Potential benefit: Tailored to specific risks and contexts
Potential problem: Fragmented, companies may exploit gaps between sectors

Use-case regulation: Regulate “high-risk” AI applications, allow innovation in low-risk areas

Potential benefit: Proportional response to actual harm
Potential problem: Defining “high-risk” is contentious; harms may be normalized

Comprehensive frameworks: Broad requirements for transparency, accountability, fairness across all AI

Potential benefit: Consistent standards, fewer loopholes
Potential problem: May be too rigid, could stifle beneficial innovation

Algorithmic accountability: Requirements to audit, explain, and document AI systems

Potential benefit: Enables oversight and challenge of harmful systems
Potential problem: May not be technically feasible for all AI types; could create compliance burdens

Most proposed regulations combine elements of these approaches. Your role as a citizen is to evaluate whether they actually protect people while enabling beneficial uses.

Understanding genuine capabilities vs. hype

Hype claims: “AI will solve climate change / cure cancer / end poverty / replace all jobs”

Reality: AI is a tool. Tools don’t solve complex problems by themselves—they help humans solve problems when applied thoughtfully within larger strategies. AI can contribute to climate research, medical diagnosis, poverty analysis, and work automation, but none of these are purely technical problems. They’re social, political, economic, and ethical challenges that require human judgment, values, and collective action.

Hype claims: “AI is an existential risk to humanity”

Reality: Current AI systems are not sentient, don’t have goals, and can’t act independently. Real AI risks are immediate and concrete: discrimination, privacy invasion, misinformation, labor exploitation, environmental damage, power concentration. Focus on actual harms happening now, not speculative science fiction scenarios.

Hype claims: “AI can do X as well as humans” (write, create art, diagnose disease, etc.)

Reality: AI can produce outputs that superficially resemble human work, based on pattern matching. This is sometimes “good enough” for specific purposes. But it lacks genuine understanding, creativity, ethical judgment, and accountability. “As well as” depends entirely on how you define the task and what quality standards you apply.

Hype claims: “You need AI to remain competitive/relevant/employable”

Reality: Sometimes true, often fear-mongering to sell products. Evaluate:

Is this actually required in your field, or is it optional?
What specific capabilities does AI provide that matter for your work?
Can you develop those capabilities with or without AI?
Is the person making this claim trying to sell you something?

Developing ongoing critical literacy

AI technology evolves rapidly. Staying informed requires:

Follow reputable sources:

Academic research (not just press releases about research)
Technology journalism from outlets with editorial standards (not just tech company blogs)
Critical voices examining AI harms (civil liberties organizations, affected communities, ethics researchers)
Regulatory developments in your jurisdiction

Recognize conflicts of interest:

AI companies have incentive to hype capabilities and minimize risks
Media outlets depend on tech company advertising and access
Researchers may be funded by companies whose products they evaluate
Politicians may receive donations from tech industry

This doesn’t mean dismiss these sources—just apply appropriate skepticism and seek diverse perspectives.

Interrogate your own reactions:

Does a claim sound too good to be true? It probably is.
Does a claim make you afraid? Fear is often used to drive sales or policy.
Does a claim align with your existing beliefs? You may be less critical of things you want to believe.
Does a claim surprise you? That’s when to investigate most carefully.

Use the skills from other topics:

Critical Thinking (Level 2): Evaluate evidence, recognize logical fallacies, identify manipulation
Science (Level 2): Understand how AI research works, what makes evidence credible
Psychology (Level 2): Recognize cognitive biases affecting your judgment
Communication Skills (Level 2): Identify persuasion tactics in AI marketing

The goal isn’t to become an AI expert—it’s to develop enough literacy to:

Recognize when claims deserve skepticism
Ask informed questions
Seek appropriate expertise when needed
Make reasonable decisions about AI in your life
Participate meaningfully in democratic decisions about AI policy

AI is being rapidly integrated into society. Those who can evaluate it critically will make better decisions for themselves and their communities than those who accept claims uncritically or reject everything reflexively.

How It Connects — Section 2: Artificial Intelligence and Large Language Models

This is the most broadly connected section in the topic — AI touches nearly every domain the program covers, which is part of what makes it so important to understand clearly.

Level 1: What Are People Capable Of? AI is one of the most powerful examples of technology extending human capability — but this section adds essential nuance. The question isn’t just what AI makes possible, but what the difference is between capability genuinely extended and capability replaced or atrophied. Understanding AI clearly is part of understanding your own potential in a world increasingly shaped by it.

Level 1: Internal Barriers Over-reliance on AI, skill atrophy, and the tendency to accept AI output uncritically are all internal barriers in the making. This section’s warnings about common pitfalls — letting AI do your thinking for you, losing the ability to verify claims independently — connect directly to the habits of mind that Level 1 identifies as obstacles to realizing human potential.

Level 1: External Barriers AI creates new external barriers through bias amplification, unequal access, job displacement, and the concentration of power in a small number of companies and governments. At the same time it can reduce barriers through accessibility tools, translation, and democratized access to expertise. This section explores both sides.

Level 1: Horse, Carriage, and Driver AI tools function as an extension of the driver — the intellectual, navigating part of the self. But an AI that navigates for you without your understanding is more like handing the reins to a stranger than developing your own driving skill. The metaphor helps clarify the right relationship: AI as a tool the driver uses skillfully, not a replacement for the driver.

Level 2: Critical Thinking Perhaps the most direct connection in the program. Evaluating AI claims, recognizing AI-washing in marketing, assessing AI products in high-stakes contexts, and understanding the difference between “AI said so” and actual evidence — all of these are Critical Thinking skills applied to a specific and increasingly unavoidable domain. The two topics are mutually reinforcing: better critical thinking makes you a better AI user; understanding AI deeply gives critical thinking a concrete, high-stakes arena to practice in.

Level 2: Psychology Cognitive biases shape how people interact with AI in predictable ways — automation bias (trusting AI output over human judgment), anthropomorphism (treating AI as if it has intentions or feelings), and the tendency to interpret confident-sounding output as accurate output. Understanding the psychology of human-AI interaction is essential for using these tools well.

Level 2: Emotion Management AI is generating genuine anxiety — about job security, about creative identity, about what it means to be human. It’s also being used to manipulate emotions at scale. This section’s content connects to Emotion Management both in terms of managing personal responses to AI-related uncertainty and in terms of recognizing when AI-driven media or communication is targeting emotional responses rather than informing them.

Level 2: Science (as a process) Understanding how AI research actually works — what peer review looks like in this field, how to evaluate AI capability claims, what “benchmark performance” does and doesn’t mean — requires scientific literacy. The hype cycle around AI is one of the clearest examples of why understanding science as a process (not just a set of facts) matters for navigating the modern world.

Level 2: Education (as a concept) AI is transforming learning environments in ways that cut both ways. Used well, it can be a powerful learning tool. Used poorly, it bypasses the learning process entirely — producing outputs without building understanding. The academic misconduct discussion in this section is essentially an Education topic: what is learning actually for, and what does it mean to shortcut it?

Level 2: Communication Skills AI is changing communication in fundamental ways — generating text, images, audio, and video at scale; enabling new forms of manipulation and misinformation; altering what it means to write or speak authentically. Understanding AI’s role in communication is increasingly part of communication literacy itself.

Level 2: Community & Cooperation Individual adaptation to AI is insufficient — the most important decisions about how AI is deployed are collective ones. Labor organizing in response to automation, democratic governance of AI systems, collective action on regulatory standards, and community-level resilience to AI-driven disruption all require cooperation. This section connects to Community & Cooperation in both the practical and the political sense.

Level 2: Long-term Thinking AI lock-in, path dependency, unintended consequences of rapid deployment, and the difficulty of course-correcting once AI systems are deeply embedded in institutions — these are all long-term thinking problems. The pace of AI development makes long-term thinking especially urgent and especially difficult simultaneously.

Level 3: Systems Thinking AI systems are sociotechnical systems — they don’t exist in isolation but are embedded in institutions, markets, social norms, and physical infrastructure. Understanding why AI produces the outcomes it does requires looking at the whole system, not just the algorithm. Systems Thinking provides the conceptual tools for that analysis.

Level 3: Organizational Intelligence Organizations are making decisions about AI deployment that affect workers, customers, and communities — often without adequate understanding of what they’re deploying or why. Organizational intelligence — how groups make decisions, process information, and respond to feedback — determines whether AI adoption in organizations is thoughtful or reckless.

Level 3: Systemic/Institutional Change AI regulation, intellectual property law reform, labor law responses to automation, and democratic governance of powerful AI systems are all systemic and institutional change challenges. This section’s discussion of regulatory approaches connects directly to the frameworks and strategies explored in Level 3.

Level 3: Part-Whole Symbiosis The concentration of AI capability in a small number of powerful actors — a few corporations and governments — is fundamentally a Part-Whole Symbiosis problem. When the benefits and risks of a powerful technology are distributed this unevenly, the whole suffers even as a small number of parts benefit enormously. Open-source AI development and democratic governance are attempts to rebalance that relationship.

Advanced Practice Exercises — Section 2: Artificial Intelligence and Large Language Models

Comprehension Check

In your own words, explain what it means to say that an LLM is doing “pattern matching and prediction” rather than thinking or reasoning. What specific behaviours or failure modes does this explanation predict — and does your experience with AI tools confirm or complicate that explanation?
The section distinguishes between tasks LLMs handle well and tasks they handle poorly. What is the underlying principle that separates the two categories? It’s not arbitrary — there’s a logic to it. Can you articulate what it is?
What is “AI washing,” and why does it work as a marketing strategy? What does it exploit about how people think about technology?
The section describes several regulatory approaches to AI governance, each with trade-offs. Name at least two approaches and explain what each one prioritizes and what it risks sacrificing.
The section argues that using AI for tasks where you’re still learning can undermine your own development. What is the mechanism by which this happens, and under what conditions might AI use support learning rather than replace it?

Reflection Exercises

Think honestly about your current relationship with AI tools — or if you don’t use them, your relationship with other technologies that do tasks on your behalf. Do you notice any signs of over-reliance, skill atrophy, or uncritical acceptance? What habits have you built, and are they the ones you’d choose deliberately?
The section covers the environmental costs of AI — energy consumption, water use, carbon output. Did you already know about these costs before reading this section? If not, does knowing change anything about how you feel about using AI tools? Why or why not — and be honest about any rationalization you notice in your own thinking.
The section discusses AI and creative attribution — questions about what it means to create something, who owns it, and whether using AI as a tool is meaningfully different from using other creative tools. Where do you personally draw that line, and why? What values are shaping where you draw it?
Consider the “capable but unproven collaborator” framework for thinking about AI. Does this framing feel accurate to you? Does it change how you’d approach using AI tools — and if you already use them, does it describe how you actually use them, or how you think you should?

Application Exercises

Capability test: Choose a task you know well — something in your area of expertise or daily life — and give the same task to an AI tool. Evaluate the output not as a newcomer but as someone who actually knows the subject. What did the AI get right, get wrong, or miss entirely? What does this tell you about where AI is and isn’t trustworthy in your specific context?
Hallucination hunt: Ask an AI tool a question where you can verify the answer — a historical fact, a scientific claim, a specific statistic. Check the response against a reliable source. Then ask a follow-up that requires the AI to be specific (names, dates, citations). Document what you find. How confident did the AI sound relative to how accurate it was?
AI washing analysis: Find a real product, service, or news article that makes significant claims about AI capabilities — a company announcement, a tech news piece, or a product advertisement. Apply the critical evaluation framework from this section: What specific claims are being made? What evidence is provided? What questions aren’t being answered? What red flags, if any, do you notice?
Ethical audit: Choose one AI tool you use or are considering using. Research, as best you can, its environmental footprint, the company’s data practices, its known biases or controversies, and its labor practices. You won’t find complete information — notice what’s easy to find and what’s deliberately obscured. What does that asymmetry tell you?
Prompting experiment: Take a task where AI output quality matters to you and deliberately compare two approaches: one minimal-effort prompt and one carefully constructed prompt using the techniques from this section. Document the difference in output quality. What does the gap tell you about what the tool actually needs from you to perform well?

Discussion Exercises

(Partner or group) Share your hallucination hunt results. Did different people find different failure modes, or similar ones? What patterns emerge? How does this experience compare to how AI is typically discussed in media and marketing?
(Partner or group) The section raises the question of academic misconduct and what it means to shortcut the learning process. As a group, try to define where the legitimate uses of AI in learning end and the problematic ones begin. Is the line the same for everyone, or does it depend on context, goals, or stage of learning? Try to reach a shared position — and notice what makes that difficult.
(Solo journaling or group) The labor exploitation discussion covers the people who do data labeling, content moderation, and RLHF work — often in low-income countries, often exposed to disturbing content, often for very low pay. This work is largely invisible in how AI tools are marketed and discussed. How does knowing about it affect your relationship to the tools? Is there a tension between knowing something is ethically problematic and continuing to use it anyway? How do you navigate that kind of tension in your own life?
(Partner or group) If you were designing a policy framework for AI governance — at a national or international level — what would your priorities be? What harms would you most urgently try to prevent, and what values would you be willing to trade off to do it? Present your framework to the group and defend it against challenges.

Research & Evidence — Section 2: Artificial Intelligence and Large Language Models

Foundational Sources

Kate Crawford — Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence (2021, Yale University Press) The most comprehensive critical examination of AI available to general readers. Crawford traces AI systems from the mines where rare earth minerals are extracted to the warehouses and data centers that power them, through to the human labor and surveillance infrastructure they enable. Covers environmental costs, labor exploitation, and political economy in depth — directly relevant to this section’s ethics content. Essential.

Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell — “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” (2021, ACM FAccT Conference) The academic paper that introduced the “stochastic parrot” framing — the idea that LLMs are sophisticated pattern-matching systems that manipulate language without understanding it. One of the most cited and debated papers in AI ethics. Freely available online; moderately technical but readable by motivated non-specialists. Directly supports this section’s explanation of how LLMs actually work.

Cathy O’Neil — Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy (2016, Crown) Examines how algorithmic systems — including but not limited to AI — encode and amplify bias while presenting themselves as objective. Written for general audiences, evidence-based, and full of concrete examples. Predates the current LLM wave but remains one of the most accessible and rigorous treatments of algorithmic harm. Directly relevant to the bias amplification discussion.

Safiya Umoja Noble — Algorithms of Oppression: How Search Engines Reinforce Racism (2018, NYU Press) A detailed examination of how search algorithms encode racial and gender bias, with implications that extend well beyond search. Provides important historical and sociological context for understanding why AI bias isn’t accidental — it reflects the biases present in training data and the values of the people who build systems. Accessible and well-documented.

Key Studies & Reports

Emma Strubell, Ananya Ganesh, and Andrew McCallum — “Energy and Policy Considerations for Deep Learning in NLP” (2019, ACL) The paper that first brought serious attention to the energy costs of training large language models, quantifying the carbon footprint of training a single large NLP model and comparing it to the lifetime emissions of multiple cars. Sparked significant debate in the AI research community. Freely available; the core findings are accessible even if the technical details aren’t.

TIME — “OpenAI Used Kenyan Workers on Less Than $2 Per Hour” (January 2023) Investigative journalism documenting the labor conditions of the workers in Kenya hired to do content moderation and RLHF (reinforcement learning from human feedback) for OpenAI — the work that makes AI tools less likely to produce harmful content. Workers were exposed to disturbing material with inadequate psychological support and paid very low wages. A primary source for the labor exploitation discussion in this section.

AI Now Institute — AI Now Report (annual) The AI Now Institute publishes an annual report on the social implications of artificial intelligence, covering labor, bias, surveillance, and policy. Each report is well-researched, freely available, and current — making it one of the best resources for staying up to date on AI’s real-world impacts without relying on industry sources.

Reputable Organizations & Ongoing Resources

AI Now Institute (ainowinstitute.org) An independent research institute focused on the social implications of AI — labor, bias, power concentration, and accountability. One of the most rigorous and credible non-industry sources on AI policy and ethics. Their publications are freely available and regularly updated.

Algorithmic Justice League (ajl.org) Founded by researcher Joy Buolamwini, the AJL focuses specifically on bias and discrimination in AI systems. Their research on facial recognition bias — showing dramatically higher error rates for darker-skinned and female faces — has directly influenced policy. A good resource for the bias amplification content and for ongoing developments in AI accountability.

Electronic Frontier Foundation (eff.org) The EFF covers AI policy, surveillance, intellectual property, and digital rights from a civil liberties perspective. Their analysis of AI legislation, copyright disputes involving training data, and surveillance technology is rigorous and freely accessible. A reliable ongoing resource for the regulatory and rights dimensions of this section.

Mozilla Foundation — Internet Health Report (annual, foundation.mozilla.org) Mozilla’s annual report on the state of the internet includes substantial coverage of AI, data privacy, and platform power. Freely available, accessible, and written from a pro-open-web perspective that aligns well with the program’s values.

Accessible Entry Points

3Blue1Brown — Neural Networks series (YouTube) Grant Sanderson’s visual explainer series on how neural networks actually work is widely regarded as the clearest introduction available anywhere. Uses animation to build genuine conceptual understanding rather than surface-level analogies. Free; no technical background required. Directly supports this section’s explanation of how LLMs function at a mechanical level.

Coded Bias (documentary, 2020 — directed by Shalini Kantayya) Follows Joy Buolamwini’s research into facial recognition bias and its deployment in policing, hiring, and public surveillance. Accessible, well-made, and grounded in real evidence and real-world consequences. One of the best entry points for understanding AI bias for people who prefer documentary to reading.

Your Undivided Attention (podcast — Center for Humane Technology) Hosted by Tristan Harris and Aza Raskin, this podcast covers the intersection of technology design, human psychology, and societal harm. Covers AI topics regularly with a focus on persuasion, manipulation, and what technology does to human attention and behaviour. Accessible and thought-provoking without being technically demanding.

Gary Marcus — Substack and public writing (garymarcus.substack.com) Marcus is one of the most prominent and readable critics of AI capability claims, consistently pushing back on hype with specific technical arguments. His writing is accessible to non-specialists and provides an important counterweight to AI industry marketing. Useful for developing ongoing critical literacy about AI claims.

Continue to Technology Intermediate Part 3 → Return to the main page.