Module 7: AI Safety, Ethics, and What Can Go Wrong

Why This Module Matters

You have spent six modules learning how to get the most out of AI. You can write better prompts, chain tasks together, pick the right tool for the job, and build workflows that save you real time. That is genuine power.

Power without understanding is dangerous.

This is the module where we slow down and look at what can go wrong. Not to scare you away from AI -- you are past that point -- but to make sure you use it with your eyes open. Every tool has failure modes. A chainsaw is incredibly useful, but you would not pick one up without understanding kickback. AI is the same.

By the end of this module, you will know the five most common ways AI can trip you up, what each major AI tool actually does with your data, and how to build habits that keep you safe without slowing you down.

Think of this as your defensive driving course for AI. You already know how to steer. Now you learn how to avoid the accidents.

Five Things That Can Go Wrong

1. Hallucinations: When AI Confidently Gets It Wrong

Here is the single most important thing to understand about AI: it does not know things. It predicts words. Most of the time, those predictions align with reality because the training data reflected reality. But sometimes the prediction looks perfect and is completely wrong. The AI community calls these "hallucinations." You can think of them as confident lies, except "lie" implies intent, and there is no intent. The AI just... made something up that sounded right.

Real examples:

A lawyer used ChatGPT to research case law for a federal court filing. ChatGPT generated citations to several cases that sounded real -- complete with case numbers, judge names, and detailed summaries. None of them existed. The lawyer submitted them to the court and faced sanctions.
A journalist asked an AI for statistics on a trending topic. The AI produced specific percentages and cited a real-sounding research institute. The numbers were fabricated. The institute did not exist.
A student asked AI to summarize a historical event. The summary was fluent and detailed -- and mixed up two different events that shared some similarities. It sounded authoritative and was wrong.

How to spot hallucinations:

Check specific claims. If the AI cites a statistic, a study, a legal case, or a specific fact, verify it independently. General knowledge is usually fine. Specific citations are where hallucinations live.
Watch for too-perfect detail. Real information is often messy. If every detail lines up neatly and the sources sound almost too credible, that is a red flag.
Ask for sources, then check them. If the AI references a book, paper, or website, go look it up. If you cannot find it, assume it is fabricated.
Cross-reference with a second source. Ask a different AI tool, or search the web. If only one source makes a claim, treat it as unverified.

What to do when you catch one:

Do not just re-ask the same question. The AI may generate a different hallucination.
Provide correct information and ask the AI to work from that. For example: "That case does not exist. Here is the actual case I am referring to: [details]. Please analyze this one."
When stakes are high -- legal, medical, financial -- always verify outside of AI. Always.

2. Privacy Leaks: When Your Data Does Not Stay Yours

Every time you type something into an AI chat, you are sending information to a server somewhere. What happens to that information depends on the tool, the plan you are on, and the settings you have chosen. Most people never check.

What can go wrong:

You paste in a confidential business document to ask AI to summarize it. That document now sits on someone else's server and may be used to train future AI models. Your competitor's AI assistant might one day reference details from your proprietary strategy.
You describe a medical condition in detail to get AI's take on it. That health information is now stored by the AI company, potentially linked to your account.
You share code from your company's proprietary software to debug an issue. That code is now part of a conversation stored on an external server.

The uncomfortable truth: If you are using a free AI tool, the company is making money somehow. Often, that "somehow" is using your conversations to improve their models. You are paying with your data.

We will cover what each specific tool does with your data in the next section. For now, the principle to internalize is simple: never put anything into a free AI tool that you would not want to appear on the internet. Because functionally, that is the risk level you are accepting.

3. Bias: When AI Reflects the World's Problems

AI learns from data created by humans. Humans have biases -- conscious and unconscious, individual and systemic. The training data for major AI models comes from the internet, books, articles, and other human-generated content. All of that content carries the biases of the people who created it and the systems they operated in.

What this looks like in practice:

An AI asked to generate images of CEOs disproportionately produces images of white men, reflecting historical patterns rather than current reality.
An AI writing a reference letter for a woman might default to language about being "collaborative" and "supportive," while the same letter for a man might default to "decisive" and "visionary" -- reflecting patterns in the training data, not the individual.
An AI summarizing news might give more weight to perspectives from certain countries or cultures because those perspectives are overrepresented in its training data.

How to account for bias:

Be aware it exists. Simply knowing that AI outputs carry bias makes you far less likely to absorb that bias uncritically.
Review with diversity in mind. When AI generates content about people -- hiring materials, marketing copy, character descriptions -- check whether the output reflects your intent or the training data's patterns.
Prompt deliberately. If you want a balanced perspective, ask for one. "Give me arguments from multiple perspectives" or "write this in a way that does not assume any particular demographic" can steer the output in a better direction.
Do not delegate high-stakes human decisions to AI alone. Hiring, lending, medical triage -- these decisions need human judgment precisely because AI can encode and amplify existing biases at scale.

4. Over-Reliance: When You Stop Thinking for Yourself

This might be the sneakiest risk on the list, because it does not announce itself. One day you are using AI to help with a task, and six months later you are running every thought through AI before you trust it. Your own judgment gets quieter. You start to feel like you cannot write, decide, or analyze without the AI's input.

The danger:

You lose the ability to evaluate AI's output because you have stopped developing your own understanding of the topic. If you do not know what good looks like, you cannot tell when AI gives you bad.
You become slower, not faster, because you are consulting AI for decisions you used to make instantly on your own.
You lose your voice. Your writing starts to sound like AI writing -- smooth, generic, and forgettable.
You miss things AI misses, and you do not even realize you are missing them.

How to avoid it:

Use AI as a collaborator, not an oracle. Ask it for options, not answers. Ask it to challenge your thinking, not replace it.
Stay in the habit of doing the thing yourself sometimes. Write without AI. Make decisions without AI. Solve problems without AI. Not always -- just enough to keep your skills sharp.
When AI gives you an answer, ask yourself: does this make sense? If you cannot answer that question, you are over-reliant. Go learn the topic well enough that you can evaluate the output.
Treat AI like a very capable junior assistant. It can do a lot of the work, but you are the editor. You are the one signing off. That means you need to understand the work well enough to sign off responsibly.

5. Security: When AI Is Used Against You

The same technology that helps you write emails and analyze data can help someone else write a phishing email that looks exactly like it came from your bank. Or clone your voice from a thirty-second video clip. Or generate a realistic image of you in a place you have never been.

AI-powered threats to watch for:

Spear phishing. AI can generate highly personalized, grammatically flawless phishing emails that reference real details about you or your company. Old phishing was full of typos and obvious tells. New phishing is not.
Voice cloning. A few seconds of your voice from social media is enough to create a convincing clone. Scammers have used this to call family members pretending to be a relative in distress, requesting emergency money.
Deepfake video. AI can generate realistic video of people saying things they never said. The tech is improving fast, and the fakes are getting harder to detect.
Business email compromise. AI can mimic the writing style of your boss or a vendor well enough to fool you into approving a fraudulent payment.

How to protect yourself:

Verify through a second channel. If you get an urgent email from your CEO asking for a wire transfer, call them. Not at the number in the email -- from your contacts.
Establish a code word with close family. If someone calls claiming to be a relative in trouble, ask for the code word.
Be suspicious of urgency. AI-generated scams often create a false sense of urgency to prevent you from thinking clearly. When in doubt, slow down.
Limit your voice and video footprint. The less high-quality audio and video of you that is publicly available, the harder it is to clone you. This is a personal trade-off -- you decide where the line is.

What Each AI Tool Does With Your Data

This section is about the specific, practical reality of what happens when you type something into a major AI tool. Companies change their policies, so verify this yourself periodically. But as of early 2026, here is how the major tools handle your data.

ChatGPT (OpenAI)

Free and Plus plans: Your conversations are used to train OpenAI's models by default. You can opt out by going to Settings, then Data Controls, and turning off "Improve the model for everyone." This must be done per account.
Team and Enterprise plans: Conversations are not used for training by default.
Storage: Conversations are stored on OpenAI's servers and can be accessed through your account. You can delete individual conversations or set up auto-delete.
The bottom line: If you are on a free or Plus plan and have not opted out, assume anything you type could influence future model outputs. Whether that means your specific text appears in a response is unlikely, but the information is being processed into the training pipeline.

Claude (Anthropic)

All plans: Conversations are not used for training Anthropic's models unless you explicitly opt in. This is the default.
Storage: Conversations are stored temporarily to maintain your session and are retained in your account history. You can delete them.
The bottom line: Claude has the most privacy-friendly default policy of the major tools. You still should not put highly sensitive data into any cloud-based tool, but the training risk is opt-in rather than opt-out.

Gemini (Google)

Free plan: Your conversations are used to train Google's models by default. You can opt out in Gemini's Activity settings, which are linked to your Google account.
Paid plans (via Google One AI Premium): Similar defaults. Check your settings.
Storage: Conversations are stored in your Google account activity, similar to how Google stores your search history. This means your AI conversations may be visible alongside your other Google activity.
The bottom line: Gemini is the most tightly integrated with an existing ecosystem account. If you use Google services heavily, be aware that your AI activity may be part of the same data profile.

The Universal Rule

Regardless of which tool you use, there is one rule that never fails: never put anything into a free AI tool that you would not want on the internet. Free tools pay for themselves with your data. Even paid tools store your conversations on servers you do not control. The risk is lower with some tools than others, but it is never zero.

If you are handling truly sensitive information -- medical records, financial data, legal documents, proprietary business information -- consider using an enterprise plan with a data processing agreement, or use a locally-run model that keeps everything on your machine. That is a more advanced topic, but it is worth knowing the option exists.

The Responsible AI Checklist

Before you use AI in any meaningful way, run through these five questions. They take thirty seconds and they will save you from most of the problems we have discussed.

Before using AI output: Did I verify the facts?

If the output contains specific claims -- statistics, citations, dates, names, legal references -- check them. AI is fluent but not reliable. You would not trust a confidently stated fact from a stranger without verification. Treat AI the same way.

Before sharing AI output: Did I review and edit it?

Read every word before you send, publish, or submit something AI helped write. Not just to catch errors, but to make sure it sounds like you, reflects your actual views, and meets the standard you want to be known for. Your name is on it. You own it.

Before putting data into AI: Is this information I would share publicly?

If the answer is no, stop. Either anonymize the data, use a tool with appropriate data protections, or do not use AI for that task. This is not paranoia. This is basic information hygiene.

Before automating with AI: What happens when it makes a mistake?

If you are setting up an AI-powered workflow -- auto-generating reports, responding to customers, processing data -- ask yourself what happens on the day the AI gets it wrong. Does a human review the output? Is there a rollback? Can you catch the error before it does damage? If the answer to all three is no, you are not ready to automate that task.

Before relying on AI: Do I still understand the subject myself?

If you cannot evaluate whether the AI's output is good, you are not using AI. AI is using you. Go learn the basics of the topic well enough to be a competent editor of AI's work. Then use AI to go faster, not to avoid learning.

AI and Your Job

This is the question everyone asks eventually, so let us be direct about it.

Will AI replace your job?

The honest answer: some of your tasks, yes. Your entire job, probably not -- if you adapt.

The jobs most at risk:

Tasks that are repetitive, rules-based, and do not require human judgment are the low-hanging fruit for AI automation. Think:

Data entry and processing -- taking information from one format and putting it into another
Basic content writing -- product descriptions, routine reports, formulaic marketing copy
Simple customer service -- answering common questions with standard answers
Routine analysis -- generating the same reports from the same data every week

If your job is mostly these tasks, AI is going to eat a significant portion of it. Not tomorrow, but sooner than you might like.

The jobs safest from AI:

Anything that requires things AI fundamentally struggles with:

Human relationships -- negotiation, empathy, trust-building, managing people through difficulty
Physical work -- plumbing, construction, surgery, anything that requires hands in the real world
Creative judgment -- not generating text, but deciding what matters, what moves people, what is worth making
Complex decision-making -- situations where every case is different, context matters deeply, and there is no single right answer

Notice I said "safest," not "safe." AI will touch every job. The question is whether it replaces you or amplifies you.

The real risk:

The most important sentence in this module might be this one: the real risk is not that AI replaces you, but that someone who uses AI does.

Think about it this way. A accountant who uses AI to handle routine bookkeeping while focusing on advisory work will produce more value than an accountant who does everything manually. A writer who uses AI for first drafts and research while bringing original insight and voice will outperform a writer who does not. A manager who uses AI to prepare for meetings, analyze data, and draft communications will have more capacity for the human work of leadership.

AI does not eliminate jobs. It changes what the job requires. The people who thrive are the ones who adapt -- who learn to use AI as a tool and double down on the human skills that AI cannot replicate.

If you have made it through seven modules of this course, you are already adapting. Keep going.

Deepfakes and Misinformation

AI can generate content that looks and sounds real but is not. This is not a future problem. It is a present one, and it is getting worse every year as the technology improves.

How to spot AI-generated content:

Look for inconsistencies. AI-generated images often have subtle problems: hands with too many or too few fingers, text that is gibberish, reflections that do not match, lighting that is inconsistent across the image. These tells are getting harder to spot as the technology improves, but they are still there in many cases.
Check the source. Where did this content come from? Is it from a news organization you recognize? Does it link to a verifiable original source? If it showed up in your social feed with no clear provenance, treat it with caution.
Be skeptical of perfection. Real photos and videos have imperfections. Real writing has a voice. If something looks too perfect, too polished, or too perfectly aligned with what you already believe, that is reason to be more skeptical, not less.
Use reverse image search. If an image seems suspicious, drop it into Google Images or TinEye. If it is AI-generated and has been debunked, you will often find that context.
Listen for vocal tells. AI-generated audio can sound slightly flat or have unusual pauses. Cloned voices may get the tone right but miss the natural rhythm and imperfections of real speech.

What to do when you encounter fake content:

Do not share it. Even sharing it to say "this is fake" spreads it further and can cause it to appear in others' feeds without your debunking context.
If you know the person who shared it, reach out privately and let them know. Most people share misinformation because they do not realize it is fake, not because they want to deceive.
Report it to the platform if it is clearly harmful -- election misinformation, financial scams, non-consensual intimate imagery, or content designed to incite violence.

Why this matters more every year:

AI generation technology is on a trajectory from "detectable if you look closely" to "indistinguishable from real." We are not there yet for most content, but the gap is closing. The habits you build now -- checking sources, being skeptical, verifying before sharing -- will serve you well as the technology continues to improve. The problem will not get easier. Your defenses need to get stronger.

Key Takeaways from Module 7

AI hallucinates. It can generate confident, plausible, completely false information. Verify specific claims, especially citations, statistics, and legal or medical information.
Your data is not always private. Free AI tools often use your conversations for training. Know what each tool does with your data and adjust your behavior accordingly. Never put anything in a free AI tool you would not want on the internet.
AI carries bias. It reflects the biases in its training data. Be aware of this, review outputs with diversity in mind, and do not delegate high-stakes decisions about people to AI alone.
Over-reliance is a quiet trap. If you cannot evaluate AI's output, you are not using it well. Stay sharp. Keep doing the work yourself sometimes. Use AI to go faster, not to stop thinking.
AI can be used against you. Phishing, voice cloning, and deepfakes are real and improving. Verify through second channels, be suspicious of urgency, and protect your voice and video footprint.
The biggest career risk is not using AI. AI will change tasks, not eliminate most jobs entirely. The people who thrive will be those who use AI well and invest in the skills AI cannot replicate.
Build the verification habit now. The responsible AI checklist takes thirty seconds and prevents most common problems. Make it automatic.

What's Next

You have come a long way. In Module 1, you were figuring out what AI even is. Now you understand how it works, how to use it effectively, and how to keep yourself safe while doing so.

Module 8 is the final module: Your AI Future. We are going to step back from the tactical and look at the bigger picture. How do you build a personal AI practice that grows with you? What should you be learning next? How do you stay current as AI keeps evolving? And most importantly, how do you make AI a sustainable part of your life and work rather than a novelty that fades?

The finish line is in sight. Let's cross it.