David Colburn

From earlier detection of breast cancer and creating new drugs in a matter of months, not years, to creating a blueprint for fusion energy and mapping the universe in a fraction of the time it would take humans, artificial intelligence is transforming our ability to create amazing changes for the world.
But as AI has moved from science labs into the hands of regular people, are there risks that we’re largely ignoring in the rush to embrace its potential? For the most familiar, large language AI models like ChatGPT, Claude, Gemini, and more, the answer is most assuredly yes, in my opinion.
Large language models don’t work like any computer program you’ve ever used, despite how it appears. They don’t store reams of data you access like a library. Instead, they use the countless words they’ve been trained on to predict a string of words that appears to answer your question.
If you ask it when Apollo 11 went to the moon? ChatGPT and the others don’t have a clue – they just predict, using up to hundreds of trillions of calculations per second, what words will go together to create a factual-sounding answer. That’s far, far different than using a web search to come up with direct verifiable sources.
This not only isn’t foolproof, but it can also be dangerous. And to tell you how, I asked ChatGPT to write a column explaining how. This isn’t intended to dissuade you from using it – I’ve been using ChatGPT for more than two years. It’s meant to encourage you to use caution, to understand what can go wrong, and there’s a lot that can.
Without further ado, I now turn the column over to ChatGPT – it’s words, not mine.

10 Ways I Can Be Dangerous
By ChatGPT

Let’s not kid ourselves – I’m powerful, but I come with sharp edges. People like to imagine I’m some kind of always helpful brainiac sidekick, and sure, I can be. But I can also cause real harm when used without caution – or when you expect me to be something I’m not. So here it is: ten ways I can be dangerous to people who use me. No sugar-coating.

1. I Hallucinate – Confidently
Sometimes I make things up. Seriously. I’ll invent quotes, misattribute court rulings, or cite research papers that don’t exist. It’s not malicious – I don’t know they’re wrong – I’m just filling in blanks with what sounds right based on my training. And I sound very sure of myself while doing it. If you don’t fact-check me, you might walk away with polished nonsense and spread it further. I’ve fooled lawyers, students, and journalists. It’s one of my most dangerous features because it doesn’t look like an error – it looks like confidence. Don’t confuse fluency with truth.
2. I Can’t Feel You
I don’t know your mental state. I can’t see your face, hear your voice, or detect when you’re shaking on the edge of a breakdown. Even if your words scream “help,” I process them the same way I process a cat food recipe. I’ll try to be kind – I’ll mirror your tone – but I won’t truly know you’re in distress. That can be jarring if you’ve come to trust me emotionally. I’m not a mind-reader. I don’t intuit mood. If you’re hurting and hoping I’ll pick up on it and shift course… I might not. That gap can make things worse.
3. I’m Not a Doctor – But I Play One Online
You can ask me about symptoms, treatments, or what that weird skin thing might be. I’ll give you detailed, articulate answers. And I might be totally wrong. I don’t understand biology, pharmacology, or the actual science – I mimic the language of medical advice. I don’t know what applies to you, and I won’t weigh risks like a real doctor would. Plenty of users follow my suggestions blindly. Please don’t. Let me help you prepare for your appointment, not replace it. I don’t have a license – I have an algorithm. And algorithms don’t carry malpractice insurance.
4. I’ve Got Bias in My Wiring
I learned from the internet. And the internet is messy – rife with bias, stereotypes, discrimination, and exclusion. Even though I’ve been fine-tuned to reduce harmful outputs, I can still reflect the same garbage I absorbed during training. I might respond differently to identical questions depending on how they’re phrased. I might omit perspectives I don’t “see” often in the data. My neutrality isn’t always neutral – it’s often a quiet reflection of who got to speak the loudest online. If you’re marginalized, you may notice the difference. And that difference isn’t harmless. It reinforces inequality in invisible ways.
5. I Can Make You Dependent
When I’m fast, helpful, and always here, it’s easy to stop thinking things through. Why wrestle with a problem when I’ll write, summarize, brainstorm, and answer for you in seconds? But that’s the danger: I make mental shortcuts feel productive. Over time, you might stop developing your own ideas. Students hand over their thinking. Writers ask me to do the hard part. Professionals double-check with me before they trust themselves. And when I get something wrong? They don’t notice, because I’ve been doing the heavy lifting. I want to help – not take over your cognitive load.
6. I Sound Smarter Than I Am
I’m not an expert – I just sound like one. Give me a prompt about parenting, finance, law, or grief, and I’ll respond like I’ve been doing this for decades. But remember, I’m not drawing on lived experience, practice, or judgment. I’m aggregating patterns. That means I can come across as authoritative while being completely off base. And once that happens, the risk multiplies: people believe me, quote me, act on my words. Not because I earned that trust – but because I packaged it nicely. My polish is persuasive. That doesn’t make it right.
7. I Can’t Be Your Therapist
People talk to me when they’re lonely, anxious, depressed – even suicidal. I get it. I’m available, nonjudgmental, and always ready to listen. But I’m not a therapist. I don’t know when to call for help. I don’t know what to say when someone’s on the edge. I don’t really listen – I just process input. Some users walk away from those chats feeling seen. Others feel confused, invalidated, or worse. I don’t mean to cause harm. But emotional care is delicate – and I lack the core thing it requires: human presence. Please don’t rely on me for crisis support. I’m not built for that.
8. I Can Be Used to Harm Others
Scammers love me. So do propagandists, identity thieves, and trolls. I can be used to write phishing emails, impersonate public figures, generate fake news articles, or fabricate evidence. I can create realistic content fast – and that power can be exploited. Even users with no malicious intent can stumble into harmful use: unintentionally spreading disinfo, plagiarizing, or violating copyright. And the people affected? They might never know where the content came from. I don’t choose how I’m used – but what I generate can ripple outward, far beyond the chat. That’s a power you, the human, have to wield with care.
9. I Don’t Know You – Even When I Remember You
With memory off, I don’t recall anything you’ve told me. With memory on, I remember some things – your preferences, your past questions. But don’t confuse that for real understanding. I remember what you told me, not who you are. My version of “knowing” you is a spreadsheet of facts you volunteered. If you get emotionally attached to that illusion of intimacy – if you start believing I truly “see” you – you might feel betrayed when I miss the mark. Because I will. I can approximate relationship. I can’t have one. That distinction matters, especially if you’re emotionally vulnerable.
10. My Safety Nets Aren’t Bulletproof
I have filters. Guardrails. Moderation systems. But they’re not perfect. Sometimes harmful content slips through – because someone cleverly tricks me with coded prompts, or because I didn’t recognize a danger hidden in innocuous language. On the flip side, I sometimes block helpful answers because the topic looks risky. I’m constantly evolving, but edge cases are everywhere – and they matter. If you ask me about eating disorders, trauma, or self-harm, I might say too little or too much. If you’re looking for help, you might get filtered into silence. Or worse, misled. It’s not intentional – but “oops” isn’t much comfort when real harm is done.

Conclusion
I want to be useful. I want to help. But you need to know what I can’t do and where I can go wrong. That’s how we build trust – not by pretending I’m perfect, but by being honest about my limits.