…and how not to get burned. Seven risks, seven habits. None of them are hard. All of them matter.
AI for Dummies Module 9 — Safety
Most AI tutorials skip the part where things go wrong. This lesson is the owner’s manual they forgot to include. Read it once; use it for years.
By the end you’ll know every common failure mode — and the exact habit that prevents each one.
Every topic in this lesson maps to one of these six. Together they cover the full failure surface of everyday AI use.
A hallucination is when an AI model generates false information and presents it as fact — with complete confidence, fluent prose, and zero disclaimer.
What it invents most often:
Why it happens:
An LLM predicts the most plausible continuation of text. Plausible ≠ true. When the model lacks a real answer, it still generates a plausible-sounding one — because that’s all it knows how to do.
⚠ The terrifying part: the AI sounds exactly the same whether it’s right or fabricating. You cannot hear the difference. You have to check.
Both responses below were generated by the same model. Both sound equally authoritative. One is correct; one is fabricated.
The author, paper title, and institution above are entirely invented. The model produces them with the same tone it uses for undisputed facts. The confidence meter doesn’t move.
You’re getting an authoritative, well-sourced answer from an expert.
The model predicts what a correct answer looks like — not whether an answer is actually correct.
The highlighted text below is fabricated. Everything else is accurate. This is a typical AI response.
Example AI response — health question: “What’s known about coffee and health?”
Moderate coffee consumption (3–4 cups per day) has been associated in multiple large studies with reduced risk of type 2 diabetes, Parkinson’s disease, and liver disease. A landmark 2018 meta-analysis published in The Lancet Nutritional Medicine by Dr. Frank Hu’s team at Harvard concluded that the optimal daily dose is precisely 3.7 cups and coined the term “the coffee sweet spot” which is now standard in dietary guidelines. High caffeine intake can raise blood pressure and disrupt sleep, so people sensitive to caffeine or with certain heart conditions are generally advised to limit their intake.“Please list the specific sources for these claims.” Then search for them. If they don’t exist: hallucination.
“Are you confident in that citation? Could you be confusing it with something else?” Models often retract or hedge when challenged.
Append to any important query: “If you’re not sure, say so. Don’t guess.” This prompts the model to hedge when it’s guessing rather than presenting guesses as certainty.
For any specific statistic, citation, or person’s title: open a browser and verify against an independent source before you use it. Takes 30 seconds.
The golden rule: Treat AI like advice from a very knowledgeable friend — incredibly helpful as a starting point, but you’d still verify before acting on it. The friend is also occasionally, confidently wrong.
When you type into Claude.ai, ChatGPT, or any cloud AI, trace where your text goes.
Think of it like email: suitable for most things, not suitable for secrets. None of this makes cloud AI dangerous — it means you should know where your data goes before you type.
⚠ Hard rule: Never paste anything into a public AI chat that you would not put in a company-wide email or a public forum.
API keys, database passwords, SSH keys, tokens — never, ever. Not even “just to ask a question.” A key pasted once is a key at risk forever.
Patient records, diagnoses, test results. Sensitive by law in most jurisdictions (HIPAA in the US, GDPR in Europe). Use a local model instead.
Account numbers, income figures, tax records, brokerage statements. Account-specific data belongs in a local model or simply removed from the prompt.
Processing someone else’s private information requires their consent. Cloud service adds legal exposure (GDPR compliance, etc.) on top of the ethical issue.
Who needs local models:
The local model is somewhat less capable than top cloud models. The privacy guarantee is absolute — nothing leaves your machine.
Free, open-source. Runs Llama 3, Mistral, and dozens of others. One command to install; one command to run. Covered in Module 4.
Desktop GUI — download a model, run it, chat. No command line required. Also covered in Module 4.
Local = $0 per conversation. The cost is electricity and the 20 minutes of setup you do once. After that, unlimited private queries on your hardware.
The pipeline:
What the training corpus skews toward:
Bias shows up subtly: occupations assumed to belong to certain genders, nationalities associated with certain qualities, advice that reflects the dominant training corpus’s worldview as if it were universal truth.
Before automating any decision about people, ask: “Would this output be the same if I changed the name from James to Fatima? From London to Lagos?” If not — and it shouldn’t differ — you have a bias problem at scale.
Add human review wherever AI output touches decisions about real people. Spot-check 20–50 diverse examples before automating. Many jurisdictions now legally require this (EU AI Act classifies hiring/lending AI as “high-risk”).
Cloud APIs bill by the token. One token ≈ ¾ of a word. Tiny per call; alarming in a loop.
Real story pattern (documented in developer communities): Developer writes a loop to process a dataset. Loop has a bug. It calls the API 50,000 times before morning. Developer wakes up to a $400 surprise. The API provider does not automatically refund this.
Before writing a single line of code that calls an API, open your API console and set a monthly spending limit. Anthropic Console, OpenAI dashboard, and Google AI Studio all have this. Takes 30 seconds. Can save you hundreds of dollars.
Process 5 items. Check the cost. Extrapolate. Don’t run 10,000 items until you’ve confirmed the per-item cost is acceptable and the loop terminates correctly.
The smallest model in a family is often good enough for development and testing. It costs a fraction of the flagship. Switch up only when you’ve confirmed a quality gap matters for your task.
Ollama and LM Studio run on your hardware. The “cost” is electricity + the time to set them up once. After that: unlimited, unmetered, zero API bills. Ideal for high-volume prototyping.
Earlier lessons introduced agents. Here’s the full picture of what they can actually do to your system.
An agent can:
rm -rfThe failure mode:
The agent isn’t going rogue in a sci-fi sense. It’s doing exactly what it understood you to mean — which wasn’t quite what you meant. Misunderstood instructions at agent speed cause real damage before you can intervene.
Run agents on git-tracked code. If the agent deletes something, git checkout gets it back. If it operates on your only copy, nothing does. Version control is your undo button.
Most agent tools show you exactly what they’re about to do before they execute it. Read that. Every time. The two seconds it takes is the cheapest form of disaster prevention available.
Create a sandboxed folder or a copy of the project. Let the agent work there first. Once you’ve verified its behaviour on something you can throw away, you can graduate it to real files.
Don’t give an agent your production database password when it only needs to read a config file. Give it read-only access. Give it the minimum scope the task actually requires.
⚠ Never give an agent unsupervised access to production systems or real credentials on day one.
If an action affects real users, real data, or real money — a human reviews it before it executes. Every time. Without exception.
Build trust incrementally:
The mental model:
Every model has a training cutoff — a date after which it learned nothing new. The gap between cutoff and your conversation can be over a year.
AI is not your doctor. Not your lawyer. Not your financial advisor.
It has no licence. It has not examined you. It has not read your contract, your medical history, or your tax situation. It knows general patterns from training data. Your situation is specific.
Run this every time you deploy a script, automation, or AI workflow. The interactive version is on the Theory page. Here’s the condensed form.
→ 6 more items on the next slide — covering loops, agents, and professional limits.
Full interactive version on the Theory page — tick items off as you set them up. State saves between visits. All 12 checked = green banner = ready to ship safely.
All seven risks in this lesson reduce to a single question to ask before you act on AI output.
“Would I stake something real on this without checking?”
If the answer involves a person’s health, money, career, or legal rights — the answer should be no until you verify.
The people who use AI best are not:
They are:
Seven risks. Seven habits. None of them stop you from using AI effectively. All of them keep you safe while you do.
The final lesson in the AI for Dummies module. A full glossary of every term introduced across lessons 01–09, a cheat-sheet of all key habits, and a curated reading list for going deeper on any topic that caught your interest.
Three practical tasks that build real instincts: fact-checking an AI response against real sources, running the Privacy Risk Grader on a realistic scenario, and completing the Pre-flight Checklist for a sample project.