Generative AI's Six Landmines: How Many Have You Stepped On?

A while back, a friend who runs a company told me a story.

He said his company had rolled out generative AI last month. The marketing team started using ChatGPT to write copy and product descriptions. Efficiency went through the roof — they cranked through a month's worth of work in a single week. He was all set to roll it out across the entire company.

Then one day, legal came knocking. They said one of the product descriptions contained a term that counted as ESG-violating language (ESG = Environmental, Social, and Governance) in certain countries, and it needed to be pulled. Another piece cited a set of data that, when checked, turned out to be completely fabricated by AI.

He asked me: Liu Run, is this thing a superpower or a ticking time bomb?

I said: Both.

Today I want to talk about what the landmines of generative AI really are — and how to defuse them.

Landmine #1: AI Confidently Makes Things Up

What exactly is an AI hallucination?

It's when you ask AI a question, and it gives you an answer with absolute confidence. The tone is definitive, the logic flows, and it all looks legit. But the content is fabricated.

In May 2023, a lawyer in the U.S. used ChatGPT to write a legal motion. After it was submitted, the judge discovered that several of the cited cases simply didn't exist. ChatGPT had made them up out of thin air. The lawyer was later sanctioned by the court.

You see — even a lawyer dared to use it, and even a lawyer got burned.

It gets scarier. Researchers have found that when AI provides medical information about eye conditions, the content it generates is incomplete, contains errors, and could even be harmful. If a doctor actually used that information to make clinical decisions, the consequences would be unthinkable.

The data is even more alarming. In 2023, an analyst crunched the numbers and found that chatbots hallucinate roughly 27% of the time, and across all generative AI outputs, 46% contained factual errors.

Nearly a three-in-ten chance of spinning fiction. Nearly half the content is wrong. That's not a tool — that's playing Russian roulette.

OpenAI themselves admit that GPT-4 is better than before. But "better than before" doesn't mean "reliable." Google's Bard fumbled during its live launch demo — it answered an astronomy question incorrectly, and the stock price dropped that same day.

How to Defuse Landmine #1

Put simply, it comes down to one thing: humans must verify.

No matter how polished the AI output looks, before it goes out the door, a real person has to go through it. Not as a rubber stamp — as a genuine fact-check. Verify the numbers. Check whether the citations actually exist. Make sure the claims have a basis.

What if manual checking is too slow? There are now tools specifically designed to detect AI-generated content. They can automatically flag the numbers, facts, and quotes in your copy and route them to a human for verification.

Another approach is to look at what training data your AI model was fed. If it was trained on data scraped from across the internet, there's no telling how much misinformation got mixed in. But if it was trained on carefully curated, vetted data, the hallucination rate drops significantly.

Landmine #2: Your Data Is Leaking Out

In May 2023, Samsung sent out an internal notice: employees were banned from using ChatGPT.

Why? Because employees had been pasting confidential company code and meeting notes directly into the ChatGPT chat box. That content goes to someone else's servers — and you have no idea what they'll do with it.

Think about it. Your company's core code, client lists, product roadmaps — just sent over like that. You think you're using AI, but AI is using you right back.

ChatGPT itself has had its own mishaps. In March 2023, a bug in one of its underlying open-source libraries exposed some users' chat history titles and payment information. OpenAI had to take the service offline for an emergency fix.

This raises a critical question: How does the AI platform you're using handle your data? How long does it store it? Who can see it? Does it use your data to train its models?

If they don't proactively tell you, and you don't ask, you're flying blind.

Landmine #3: The Murky Waters of Copyright

In 2023, seventeen authors, including Sarah Silverman, jointly sued OpenAI and Microsoft. Their claim: their copyrighted works had been used to train GPT models without permission.

That case still isn't resolved. But it already makes one thing clear: the copyright issues around AI training data are a murky, unsettled mess.

For businesses, there are two layers of risk here.

The first layer: Does the data you're using to train or fine-tune your model contain someone else's copyrighted content? If so, you could face an infringement lawsuit.

The second layer: Do you actually hold the copyright to the content AI generates for you? The current stance of the U.S. Copyright Office is: it depends on how much human creative contribution was involved. Content that is purely AI-generated is likely not protected by copyright at all. In other words, what you paid AI to produce might not actually belong to you.

The European Union is pushing legislation that would require AI companies to disclose the sources of their training data. In the U.S., lawmakers have introduced the Generative AI Copyright Disclosure Act of 2024. The direction is clear: transparency is becoming a hard requirement.

Landmine #4: Bias Hiding Inside the Algorithm

In 2024, the University of Washington ran an experiment. They had three mainstream AI models screen resumes.

The results were striking. These models exhibited clear racial and gender preferences. Resumes with white-sounding names ranked ahead 85% of the time. Male names were preferred 52% of the time. And Black male names never beat out white male names — not once.

Think about it. If your hiring system is hooked up to this kind of AI, you think you're doing efficient screening — but you're actually systematically discriminating against a segment of people.

Loan approvals, insurance pricing, customer segmentation — any scenario that involves making judgments about people, as long as AI runs with bias baked in, you're planting landmines for yourself. It's only a matter of time before regulators come knocking, and only a matter of time before you end up in court.

Landmine #5: Getting Locked In by Your Vendor

There's a category of risk that doesn't come from AI itself — it comes from your relationship with your AI vendor.

Imagine you've built your entire workflow on a single platform. Your data is in there. Your models are in there. Your team only knows how to use this one system. Then one day, they raise the price. Or a key feature gets cut. What do you do?

Switch to another vendor? Your data needs to be migrated, your models need to be retrained, and your team needs to learn everything from scratch. That cost could end up being higher than what you paid to get on the system in the first place.

That's vendor lock-in. You thought you bought a tool, but you actually locked yourself in a cage.

Landmine #6: People Get Lazy

The last landmine is the most hidden — and the easiest to overlook.

When AI can write your copy, run your analysis, and draft your proposals, your team will gradually start to depend on it. First they stop doing their own research. Then they stop thinking independently. Eventually, even their judgment starts to atrophy.

If a marketing team relies entirely on AI to produce content, efficiency goes up, sure. But when you read what comes out, you can tell — it doesn't sound like your company anymore. That signature voice is gone.

AI is an amplifier, not a replacement. Give it good input, and it gives you good output. But if your own judgment is weak, it'll just amplify your mediocrity tenfold.

So What Do You Do? Three Lines of Defense

Having talked about all these landmines, I'm not saying don't use AI. You absolutely should. But you need defenses.

First line: Choose the right platform.

Enterprise-grade AI platforms and consumer-grade AI tools are two different animals. You need to check whether they have data security certifications — things like SOC 2, GDPR, CCPA, HIPAA. You need to check whether they train on your data. You need to check whether they can customize output to your brand guidelines and terminology standards.

A platform that won't even tell you how long it retains your data has no place in an enterprise setting.

Second line: Build your own capabilities.

Your data is yours. Your use cases are yours. Your team is yours. You can't hand all of that over to a vendor.

You need your own training data, your own business logic, and people who understand how to use and fine-tune the tools. That way, even if your vendor runs into trouble, your core assets stay in your hands.

Third line: Humans stay in the driver's seat.

No matter how impressive AI gets, the final call must be made by a human. Before anything goes out, a person reviews it. When it comes to major decisions, a human is the safety net.

Teams need training — not on how to use the tools, but on how to judge whether what AI gives them is correct. Using it well and using it correctly are two different things.

A Real-World Example

Commvault is a company that does data protection. When they were going through a brand transformation, they used Writer's knowledge graph to build an internal app called Ask Commvault Cloud, which lets the sales team quickly pull up brand messaging and product information.

Their Chief Marketing Officer, Anna Griffin, said something notable: what used to take 8 to 12 hours a day now produces a first draft in 20 minutes, ready for human review.

8 to 12 hours, down to 20 minutes.

But notice — she said "ready for review." Not "AI sends it out directly." AI does the work, and humans call the shots. That's the human element done right.

A Final Word

Generative AI is a very sharp knife. It slices through prep work lightning fast — but it's just as easy to cut yourself.

Should you stop using it because you might cut yourself? Of course not. But you need to learn how to hold it, how to set it down, and when to pull back.

Pick the right platform, set up the right processes, and keep human judgment sharp — and this knife can do a lot of work for you.

Rush in with no guardrails, on the other hand, and those six landmines are waiting.

Technology won't do the thinking for you. The one who can defuse these landmines will always be you.