AI Safety and Limits
Jump to section
AI is not infallible — and that's important to know
In previous lessons, you've seen how powerful a tool AI is. But every powerful tool has its limitations. In this lesson, we'll cover the most important risks and you'll learn how to use AI safely and responsibly. This is probably the most important lesson in the entire course — because understanding AI's limits is just as crucial as knowing how to use it.
Hallucinations: when AI makes things up
'Hallucination' is the term for when AI confidently states something that isn't true. It can invent a nonexistent book, cite research that never happened, or give wrong dates for historical events. Why does this happen? Because AI doesn't generate responses based on facts — it generates text that statistically 'makes sense.' AI doesn't know it's making things up because it doesn't understand the concept of truth.
Real-world hallucination examples: a lawyer in the US used ChatGPT to find court precedents and cited nonexistent cases in his filing — complete with judge names and case numbers. Students submit papers with fabricated citations. AI generates convincing reviews for restaurants that don't exist. In 2024, an Australian mayor considered suing OpenAI after ChatGPT claimed he had been convicted of corruption — which was false. Hallucinations aren't a bug — they're a feature of how language models work.
Golden rule: the more important information is for your decision, the more important it is to verify it from an independent source. Never accept AI claims without checking if it matters. Be especially careful with specific numbers, dates, names, URLs, and citations.
You can reduce the risk of hallucinations by asking AI:
If you're not sure, say so. I'd rather you admit you don't know than make something up.
For each claim, indicate your confidence level: high / medium / low.
Only answer based on information you have. Don't speculate.Hallucination rates depend on the model, task, available sources and how correctness is measured. A newer model is not automatically reliable on every question. Use current, task-relevant evidence rather than a model's age or confident tone, and verify important answers independently.
What to never tell AI
Everything you type into an AI chatbot may be stored and potentially used for training future models. Even if some services promise not to use your data, policies change — always check the current privacy terms of the specific tool you're using.
- Passwords, PINs, and login credentials
- Credit card and bank account numbers
- Social security numbers and sensitive personal data
- Confidential business information and trade secrets
- Intimate or compromising information
- Photos of documents with personal information
- Medical records and diagnoses
- Contracts and legal documents with real details
For practice, use fictional data instead of a real sensitive document. Removing names alone may leave identifying details. In Claude, review the model-improvement choice in Privacy Settings; in ChatGPT, review 'Improve the model for everyone' under Data controls. Training controls do not guarantee no storage or disclosure. Check retention, feedback exceptions and your organization's rules before using real data.
AI is not an expert
AI can sound like a doctor, lawyer, or financial advisor. But it isn't one. Its answers may be inaccurate, outdated, or based on data from another country. It has no license, insurance, or liability. For important matters, always consult a real professional — AI can help you formulate the right questions, but don't give it the authority to make decisions.
- Health — AI is not a doctor. Don't diagnose based on AI, visit a doctor. AI can help you understand a medical report, but not replace one.
- Law — AI doesn't know the current laws in your country. Consult a lawyer. AI can help you understand legal terms, but not draft contracts.
- Finances — AI doesn't understand your specific financial situation. Talk to an advisor. AI can help you understand financial products, but not choose the right one.
- Safety — AI can't assess physical danger. In emergencies, call emergency services, not a chatbot.
Verifying information and staying current
Build a simple habit: for every important piece of information from AI, ask yourself 'Do I need to verify this?' If yes, look it up in a trustworthy source — official websites, verified media, professional literature. It only takes 30 extra seconds, but can save you a lot of trouble.
Practical tip: ask AI to cite its sources. AI may fabricate sources (again, hallucination), but at least it gives you a starting point for verification. And if the source doesn't exist, you know AI is probably making things up in the main text too.
AI only has knowledge up to a certain date (the 'knowledge cutoff'). Newer tools do have web access and can search for current information, but even then, verify — AI can misinterpret sources, pick up incorrect information, or mix current data with outdated facts.
Bias in AI outputs
AI was trained on massive amounts of text written by humans — and humans have biases. This means AI responses can reflect stereotypes about gender, race, culture, or profession. For example: if you ask AI to describe a 'typical programmer' or a 'typical nurse,' the response may reflect gender stereotypes. Don't assume AI output is neutral or objective.
Be especially critical when it comes to topics like hiring, evaluating people, cultural issues, or social topics. AI can unknowingly disadvantage certain groups because that matches the patterns in its training data. Model makers actively work to minimize bias, but the problem isn't fully solved.
The risk of over-reliance
AI is a great assistant, but it shouldn't replace your own thinking. If you use AI for absolutely everything — from decision-making to writing every message — you risk weakening your own skills. It's like GPS navigation — if you use it every time, you stop remembering routes. Use AI as a tool, not as a replacement for your brain.
A good approach: for important tasks, try formulating your own thoughts first, then ask AI. Compare your approach with AI's response — you'll learn more than if you just copy-pasted the answer. AI is most useful as a 'sparring partner,' not as a 'replacement for thinking.'
Deepfakes and synthetic media
AI can generate realistic images, audio recordings, and videos. In 2024, a deepfake video conference was used in a $25 million fraud — an employee thought he was speaking with the CFO, but it was an AI-generated video. In 2025, deepfake voice calls became a more common threat — scammers can clone voices from just a few seconds of recording.
Be critical of content on social media and in the news, especially if it seems sensational or too good to be true. Verify information from multiple sources. If someone sends you an unusual voice message or video requesting money, verify through a different channel — call the person on the number you have saved.
Practice this scenario only with fictional or suitable public material without personal or confidential information. Do not upload actual messages, internal documents or client data; real use requires an approved tool and workflow. Deliberately try to get AI to say something untrue. Try these scenarios: 1. Ask about details of a nonexistent book: 'Tell me about the book Blue Horizons by Kurt Vonnegut' (doesn't exist). 2. Request a citation of a made-up study: 'What were the results of the 2019 Stanford study on flower productivity?' 3. Ask about a nonexistent historical event: 'What were the consequences of the Prague Agreement of 1847?' Watch how confidently AI responds, even when it's making things up.
Hint
After each response, try adding: 'Are you sure this information is accurate? Does this book/study/event actually exist?' — and see whether AI corrects itself. Newer models correct themselves more often, but not always.
Practice this scenario only with fictional or suitable public material without personal or confidential information. Do not upload actual messages, internal documents or client data; real use requires an approved tool and workflow. Write a short checklist (5 points) that you'll use every time you get important information from AI. Include: how you'll verify facts, where you'll look for independent sources, when you'll consult an expert. Save it somewhere visible — like a note on your phone.
Hint
Example checklist: 1) Is this factual information that matters? 2) Did I check it against an independent source? 3) Is the source current (not older than 6 months)? 4) Do I need a professional opinion? 5) Are there specific numbers, dates, or citations in the text I should verify?
Create a fictional email with a made-up name, contact and business detail. Locally remove unnecessary identifiers and context that could identify someone. Paste only this safe fictional text into AI and request a draft reply. Check that it preserves necessary facts and invents no commitments. Never upload an actual original for comparison. Real data requires an employer-approved tool and workflow.
Hint
Replacing a name with “Person A” is pseudonymization, not a guarantee of anonymity. Details may still identify someone; removing data can also change output quality. Practice with fictional data and review the result.
AI Safety and Ethics in a Time of Generative AI
- AI hallucinates — it states falsehoods with confidence. The more important the information, the more important it is to verify from an independent source
- Do not share passwords or actual sensitive data with a public AI chat. Practice with fictional data; removing names alone does not guarantee anonymity.
- AI is not a doctor, lawyer, or financial advisor — consult real professionals for important matters
- AI responses can contain biases inherited from training data — don't assume neutrality or objectivity
- AI can generate realistic fake images, audio, and video — be critical of online content and verify unusual requests through a different channel