Session 9: The Dark Side — AI Risks, Bias, and Ethics

AI is powerful but flawed — hallucinations, bias, deepfakes, and privacy concerns are real. Here's what you need to know.

We've spent eight sessions singing AI's praises. Time for the other side. AI is powerful, but flawed in subtle, dangerous, or downright weird ways. Hallucinations, bias, deepfakes, privacy violations, and job displacement are real concerns. Let's talk about what can go wrong and how to be a smart, responsible AI user.

Hallucinations — When AI Lies with Confidence

What are hallucinations? AI generating information that sounds correct but is completely made up — fake citations, invented statistics, nonexistent laws, fabricated people. The output reads with total confidence, making it especially dangerous.

Why Does This Happen?

LLMs are pattern-completion engines, not fact databases. They generate the most probable next words based on training data, not verified truth. They have no mechanism to "know" whether something is true — only whether it sounds plausible.

Real Examples

• Lawyers citing fake court cases generated by ChatGPT in a 2023 legal filing
• AI generating fake research papers with fabricated authors and journals
• Chatbots providing incorrect medical and legal information presented as fact

How to Protect Yourself

• Always verify facts, statistics, and citations independently
• Use AI for drafting, not as a definitive truth source
• Cross-check with tools like Perplexity that cite their sources
• Be especially cautious with names, dates, legal references, and academic citations

Bias — When AI Inherits Our Worst Patterns

AI systems can reflect and amplify biases present in their training data, leading to unfair or discriminatory outputs at scale.

Why Does This Happen?

AI learns from human-generated data — and human data contains historical biases. If the training data reflects decades of inequality, the model will reproduce those patterns as if they were neutral facts.

Examples of AI Bias

• Hiring algorithms scoring women lower for technical roles (trained on male-dominated historical hiring data)
• Facial recognition performing significantly worse on darker skin tones
• Language models associating certain professions with specific genders (doctors = male, nurses = female)
• Loan approval models penalizing certain ZIP codes as a proxy for race

Why This Is Especially Dangerous

• It automates and scales discrimination to millions of decisions
• It feels "objective" because it's algorithmic — harder to challenge than a human decision
• It can create feedback loops where biased outputs generate more biased training data

What You Can Do

• Question AI decisions that affect people's lives and opportunities
• Ask: "What data was this trained on? Who isn't represented?"
• Support transparency, auditing, and accountability for AI systems

Privacy Concerns — What Happens to Your Data?

Training Data Issues

AI models are trained on massive internet scrapes that often include personal information, copyrighted content, and private conversations — frequently without explicit consent from the people involved.

Conversation Data

What you type into AI tools may be used for model training. Always check the privacy policy of any AI service you use. Some tools retain conversations by default; others offer opt-out options.

Never Share with AI Tools

• Passwords or authentication credentials
• Financial details (account numbers, SSNs)
• Proprietary business data or trade secrets
• Personally identifiable information (PII) unless you trust the provider's policy

Privacy Tips

• Review privacy settings on every AI tool you use
• Use "opt out of training" features where available
• Be especially careful with company/client data in consumer AI tools
• Consider enterprise versions with stronger data protections for professional use

Deepfakes and Misinformation

AI can now generate convincing fake images, audio, and video of real people. The quality is increasing rapidly, making detection harder with each generation of models.

Impact

• Political misinformation: fake speeches, fabricated endorsements, manipulated events
• Celebrity fraud: fake endorsements and scam advertisements
• Revenge content: non-consensual synthetic media
• Financial scams: cloned voices used to impersonate executives and authorize transfers

How to Protect Yourself

• Be skeptical of sensational content, especially during elections or crises
• Check source reputation before sharing or believing
• Use reverse image search to verify visual content
• If something seems too outrageous to be true, it might be AI-generated

Copyright and Intellectual Property

One of the most contentious debates in AI today: models are trained on content created by human artists, writers, and musicians — often without their permission or compensation.

The Legal Landscape

• Legal battles are ongoing worldwide — the EU AI Act, US court cases, and international copyright disputes
• No clear consensus yet on whether training on copyrighted material constitutes fair use
• Artists and creators are pushing back through lawsuits, opt-out registries, and legislation

Practical Guidelines

• Don't assume AI-generated content is automatically copyright-free
• Check the terms of service for any AI tool you use commercially
• Give credit where appropriate
• Be transparent about AI involvement in your work

The Responsibility Framework

AI is a tool — and like any tool, it can be used to build or to break. A hammer builds a house or smashes a window. The tool isn't moral; the user is.

Human in the loop matters. AI should assist decision-making, not make autonomous high-stakes decisions without oversight.

Transparency: Be honest when content is AI-generated or AI-assisted. Trust depends on honesty.

Verification: Always check AI outputs in high-stakes contexts — medical, legal, financial, and journalistic work demands human verification.

Real-Life Examples

• Healthcare: An AI diagnostic tool trained primarily on one demographic may misdiagnose patients from other backgrounds, leading to delayed or incorrect treatment.
• Creative industry: Artists discovering their unique styles were used to train image generators without consent or compensation.
• Education: Students submitting AI-generated work as their own, undermining their own learning and academic integrity.
• Financial fraud: Scammers using AI-cloned voices to impersonate executives, authorizing fraudulent wire transfers worth millions.
• Positive developments: Anthropic, Google, and OpenAI investing heavily in safety research, red-teaming, and responsible deployment practices.